Intel's 2026 roadmap is unlike any the company has published in recent years, because its manufacturing ambitions and its product launches have to succeed simultaneously.

Panther Lake, the Core Ultra Series 3 laptop processor unveiled at CES in January, is the first consumer chip built on Intel 18A — the company's new process node combining RibbonFET GAA transistors with PowerVia backside power delivery. Clearwater Forest, the next-generation Xeon E-core server CPU formally introduced March 3 at MWC 2026, is the server counterpart to it, and both are proof points for a foundry business that Intel has publicly stated could not justify proceeding to its next node, 14A, without first securing a major external customer.

Meanwhile, Intel is currently shipping the AI data center chip Gaudi 3, which has been available through cloud partners since late 2024. The chip was supposed to be followed by Falcon Shores, but Intel cancelled it for commercial release and confirmed it would deploy the chip internally instead, redirecting its GPU roadmap toward inference workloads. That produced Crescent Island, an inference-focused data center GPU which is expected to enter customer testing in the second half of 2026, with a potential successor in ‘Jaguar Shores’, due 2027.

Meteor Lake to Nova Lake

Since 2023, Intel's consumer CPU roadmap has focused on architectural consolidation, including the abandonment of the monolithic die. Meteor Lake, which launched in December 2023 as the first Core Ultra series processor, moved Intel's consumer laptop chips onto Intel 4 with Foveros 3D packaging, splitting compute, graphics, SoC, and I/O functions across separate tiles connected via hybrid bonding. That was an inflection point, with every subsequent generation iterating on that foundation rather than departing from it.

Then came Lunar Lake, the Core Ultra 200V series that launched in September 2024, which Intel hailed as its most power-efficient x86 platform, targeting the Copilot+ PC category with a fourth-generation NPU and the debut of the Xe2 graphics architecture. Arrow Lake followed in October 2024 as the desktop counterpart under the Core Ultra 200S branding.

While both share the multi-tile approach, they diverge at the process level. Arrow Lake consumer parts don’t use Intel 20A; Intel publicly confirmed the decision to use external nodes instead — almost certainly from TSMC — for the consumer desktop line. Intel originally said that 20A would be the node that would introduce RibbonFET and PowerVia, but the company moved those technologies to 18A instead and treated 20A as a stepping stone it bypassed for production.

Panther Lake, announced at CES in January 2026 as Core Ultra Series 3, is the first client platform built on Intel 18A. Intel cited over 200 system designs in development across laptop partners, alongside a claimed 60% better multi-threaded performance versus Lunar Lake at similar power, and up to 180 total platform TOPS — 120 of which come from the Xe3 integrated GPU and 50 from the NPU 5 architecture. Those figures are Intel estimates tied to specific workloads and comparison generations; the NPU alone meets Microsoft's 40 TOPS threshold for Copilot+ PC certification, but the 180 TOPS figure reflects all three compute engines combined.

Nova Lake is next, with Intel's Q4 2025 earnings guidance initially targeting an end-of-2026 launch. This, as we understand, is likely to be delayed to 2027; process node and die configuration details remain unconfirmed, and it’s far too early to speculate given that the upcoming Arrow Lake refresh (Core Ultra 200K Plus) is still to come.

Xeon and data center CPUs

Xeon 6 formalized a split Intel had been building toward for several years: P-core variants targeted at compute-intensive and AI inference workloads, and E-core variants aimed at density, throughput-per-watt, and scale-out workloads like containerized cloud infrastructure.

Sierra Forest launched in June 2024 as the first Intel 3 server product. Its E-core design packs a high thread count into a constrained thermal envelope, making it well-suited for high-density rack deployments. Granite Rapids, the P-core counterpart, followed in September 2024, targeting scientific computing, high-performance databases, and AI inference on large models. Both families share a common platform foundation — a unified I/O die connected via EMIB packaging — which reduces platform churn for OEMs and provides a validation reuse advantage across derivative SKUs.

Meanwhile, Clearwater Forest, introduced March 3 at MWC 2026, is Intel's first 18A server CPU. Expected to be released later this year, the chip packs 288 Darkmont E-cores across 12 compute chiplets in its maximum configuration, each with 24 cores all built on 18A. Those compute tiles are stacked on three active base dies fabricated on Intel 3 using Foveros Direct 3D, while two I/O tiles on Intel 7 handle connectivity, and lateral integration across the package is handled by EMIB.

EMIB 3.5D then extends this further by combining those Foveros-stacked modules with Intel's second-generation EMIB bridges — scaled from 55-micron to 45-micron bump pitch — to link heterogeneous tiles laterally across the package, whether those are identical compute modules or disparate I/O and memory dies. The result is a package whose total silicon area far exceeds what a conventional silicon interposer could accommodate. A clean Clearwater Forest launch would therefore validate both Intel 18A and its advanced packaging simultaneously.

Finally, Diamond Rapids will arrive as an exclusively 16-channel platform after Intel cancelled the 8-channel SKUs that were originally planned for the Xeon 7 lineup. The remaining parts are expected to pack up to 192 P-cores across four compute tiles in an LGA9324 package, with 2nd-generation MRDIMM support pushing memory bandwidth to roughly 1.6 TB/s — nearly double Granite Rapids' ~844 GB/s. Intel has indicated a 2H 2026 launch window, but has said nothing more solid at this stage.

AI accelerators

Intel’s AI accelerator portfolio hasn’t followed as clean a generational progression as its CPUs have. Gaudi 3, as previously mentioned, is the current shipping product and has been available through cloud partners and direct customers since late 2024, with Intel expanding availability throughout 2025.

Intel has marketed Gaudi 3 around openness and software portability, with the argument being that customers locked into Nvidia’s CUDA ecosystem face procurement and pricing constraints that a chip running on open frameworks like PyTorch and oneAPI can avoid. While this has let the chip find some traction, Gaudi 3 hasn’t achieved a meaningful share in large-scale training clusters where Nvidia’s accelerators still dominate by a huge margin.

The most concrete successor to Gaudi 3 in the near-term is Crescent Island, which Intel announced as an inference-focused data center GPU in October 2025 at the OCP Global Summit, with customer sampling due to begin in the second half of 2026. The card is built on the Xe3P architecture, a performance-enhanced version of the Xe3 GPU used in Panther Lake, and carries 160 GB of LPDDR5X memory.

That memory choice is a deliberate departure from the HBM stacks used by Nvidia and AMD in their high-end accelerators: Intel is positioning Crescent Island as a power- and cost-optimized part for air-cooled enterprise servers, with Intel CTO Sachin Katti citing "tokens-as-a-service" providers as the primary target. No performance figures have been disclosed.

When and if it does sample later this year, it will be going up against AMD's Instinct MI450 and Nvidia's Vera Rubin architecture, both of which use HBM4 and target a broader range of workloads. Crescent Island's narrower inference focus could make it competitive on cost-per-token, but the 160GB LPDDR5X configuration offers substantially less memory bandwidth than HBM-based competitors, which remains the main bottleneck for large model inference.

Jaguar Shores, meanwhile, has been confirmed by Intel as a product, though technical details about it remain sparse. Intel products chief Michelle Johnston Holthaus stated during the company's Q1 2025 earnings call that Jaguar Shores remains on the AI roadmap despite the cancellation of its predecessor, Falcon Shores, and described it as a rack-scale design incorporating silicon photonics interconnects. Intel has also confirmed, via a slide shown at its AI Summit, that Jaguar Shores will carry the Gaudi brand and use HBM4 memory from SK hynix.

Should it launch, Jaguar Shores would be Intel’s first return to HBM-based AI acceleration since Ponte Vecchio, but specifications remain unconfirmed, and we’re very unlikely to see a release until 2027 at the earliest. That would put it up against Nvidia’s Vera Rubin successors and AMD’s Instinct MI500 series — and whether it can be competitive by then depends heavily on software maturity, an area where Intel’s track record in AI acceleration has been consistently weak.

Process nodes and packaging

Intel 4, which debuted with Meteor Lake, was Intel's first EUV-enabled manufacturing node, claiming 21.5% higher frequencies at the same power as Intel 7, or 40% lower power consumption at the same frequency, alongside a 2x transistor density improvement for high-performance libraries. Intel 4 also introduced second-generation Contact-over-Active-Gate, enhanced copper interconnects with cobalt cladding for better performance and electromigration resistance, and doubled MIM capacitance density to reduce voltage droop.

Production ran at Intel's D1 facility in Hillsboro, Oregon, with Fab 34 in Ireland coming online for Intel 4 volume production in late 2023. Notably, only Meteor Lake's compute tile used Intel 4; the graphics, SoC, and I/O tiles were sourced from TSMC and older Intel nodes, reflecting the limited scope of Intel 4 as a chiplet-specific node.

Intel 3 followed as an 18% performance-per-watt improvement over Intel 4, with broader EUV usage, improved transistor cells, and both I/O and high-density cell libraries suited for server workloads. Sierra Forest, which launched in June 2024 as the first E-core Xeon 6, was its first flagship product, followed by Granite Rapids with P-cores in September 2024. Unlike Intel 4, Intel 3 was designed as a more general-purpose node from the start, underpinning Intel's server ramp and serving as the base die for Clearwater Forest's heterogeneous packaging.

Intel 20A, meanwhile, was the planned introduction point for RibbonFET and PowerVia in production, and Intel confirmed it entered production readiness in 2024. But Intel also confirmed the decision to shift Arrow Lake consumer parts away from Intel 20A to external nodes. The only logical explanation for this is that Intel concentrated its 20A engineering on proving the key technologies it needed for 18A rather than committing a high-volume product line to an intermediate node.

Every product on Intel's 2026-2028 roadmap runs on Intel 18A, the company's first node to combine RibbonFET gate-all-around transistors with PowerVia backside power delivery. RibbonFET wraps the gate entirely around the channel on all four sides, improving electrostatic control and reducing leakage compared to the FinFET structures Intel used through its 10th Gen era. PowerVia routes power through the back of the silicon wafer, freeing front-side routing resources for signal interconnects. 18A entered high-volume manufacturing in October, but yields remain below profitable levels and, per CFO David Zinsner, will not reach desired cost thresholds until the end of 2026 at the earliest.

Intel 14A, which uses High-NA EUV — which Intel is the first to deploy — remains contingent on securing a major external foundry customer. The good news is that Intel has said it has two prospective customers in the works following early PDK access, and CEO Lip-Bu Tan reckons that firm supplier decisions will be made in the “second half of this year… extending into the first half of 2027.” A lot is riding on these prospective customers, with Intel having publicly discussed the possibility of slowing or cancelling 14A and subsequent nodes if external foundry revenue does not materialize at scale. Without it, the capital expenditure required to develop and ramp leading-edge nodes past 18A will become extremely difficult to justify.

The future of Intel

Whether Clearwater Forest's 2026 launch materializes will be a solid indication of whether 18A performs at the scale Intel has projected, while Panther Lake's rollout through laptop OEMs will test whether 18A volume manufacturing is genuinely ramping up or still constrained to early production quantities.

Meanwhile, any announcement from Intel Foundry on an external customer committing to 18A or beginning 14A engagement could substantially change the economics of Intel’s roadmap.

During the 10nm era, Intel's manufacturing problems were visible and protracted over several years. Today's timeline is more compressed, and Intel’s public milestones — Panther Lake and Clearwater Forest shipping on 18A in close succession — are specific enough to hold the company to account.

Intel this week reiterated its plans to address emerging and niche AI inference use cases with products that offer the best performance efficiency and value. Among the products Intel mentioned is a hybrid solution that combines multiple IPs in a bid to address specific use cases that require an x86 CPU, a fixed-function AI accelerator, and programmable logic.

"Over the last several quarters, we have been developing a broader AI and accelerator strategy that we plan to refine in the coming months," said Lip-Bu Tan, chief executive of Intel, during an earnings call with investors and analysts. "This will include innovative options to integrate our x86 CPUs with fixed-function and programmable accelerator IP."

A mysterious hybrid design

Based on the description provided by the head of the company, the hybrid AI solution is set to pack Intel's x86 cores, Intel's fixed-function accelerator IP (which may derive from its Xe GPUs, or from something more compute-oriented), and programmable logic to accelerate emerging workloads (perhaps licensing FPGA IP from Altera, or from QuickLogic, which seems to have hard eFPGA IP implemented on Intel 18A (note that this is speculation, not confirmation). The goal is to produce a hybrid AI accelerator capable of targeting both existing and emerging inference workloads.

"Our focus is on the emerging wave of AI workloads — reasoning models, agentic and physical AI, and inference at scale — where we believe Intel can truly disrupt and differentiate," Tan said.

Building a hybrid processor to address AI workloads is not entirely new. When AMD and Intel developed their product strategies for artificial intelligence and supercomputing early this decade, they envisioned processors combining x86 cores and GPU-based accelerators, perhaps envisioning bursty workloads where control flows will dominate execution.

Over time, it appeared that the compute performance appetites of AI workloads are so high that developers of frontier AI models prefer to use servers featuring one multi-core CPU and up to eight purebred compute GPUs, rather than hybrid processors featuring both x86 and GPU IP (i.e., execution dominating the control flow). As a result, both companies quietly shut down their hybrid projects and focused on AI accelerators based on architectures derived from GPUs.

GPUs are still on the table

When it comes to Intel, it first removed x86 IP from its codenamed Falcon Shores product and converted it into a pureblood AI GPU, then decided against releasing the product commercially, instead using it as a development vehicle for the software stack and rack-scale AI platform. As a result, Intel appears to have two AI accelerators within its AI roadmap (if nothing has changed): Crescent Island, due this year, and Jaguar Shores due in 2027.

Crescent Island is an inference-optimized Data Center GPU that carries one or two high-performance processors based on the Xe3P architecture (performance-enhanced 3^rd Generation Xe GPU architecture) along with 160 GB of LPDDR5X memory. The product is designed to handle a 'broad range of data types' relevant for inference workloads, cloud providers, and on-prem corporate deployments, which is why it is 'power and cost optimized for air-cooled enterprise servers.' Crescent Island will be a workhorse for accelerating mainstream AI models optimized for GPUs.

As for Jaguar Shores, this one is expected to be based on an architecture optimized for compute as well as for AI training and inference, so it will be a full-fat purebred data center GPU with plenty of HBM4 memory onboard. Meanwhile, rack-scale solutions based on Jaguar Shores are projected to feature silicon photonics interconnects to maximize the performance of massive clusters. Unless Intel repositions Jaguar Shores, this one could indeed offer formidable performance.

Where does the new mysterious design fit in?

Given that Intel has the Crescent Island and Jaguar Shores GPUs (and likely, their successors) serving very different market segments — from edge AI inference workloads to on-prem deployments to large-scale inference and training — in the roadmap, at first glance, it does not seem that Intel may need anything else, other than positioning the pair properly. However, these observations are based solely on first impressions, so it's worth digging deeper.

First up, the new product is not meant to compete against heavy-duty Rubin or Feynman GPUs or systems on their base (potentially Jaguar Shores). It is also not meant to compete with Crescent Island. But perhaps it can apply to heterogeneous, latency-sensitive, retrieval-augmented (RAG) AI workloads on-prem? On-prem data center deployments is a market that Intel has served well historically, so this is almost certainly the market it plans to address.

This market now potentially includes a stable of AI workloads, spanning recommendation systems and fraud detection, all the way to physical AI. While clearly distinct, these are all bursty, lightweight models and rule engines under strict latency agreements (SLAs), which is exactly where a tightly integrated hardware succeeds.

Agentic AI systems also fall into this category: they spend much of their time planning, branching, calling tools or databases, and then make small inference steps (albeit with limited batching), which keeps GPUs underutilized, thus demonstrating CPU-GPU synchronization overheads.

In all of the aforementioned cases, workloads are bursty, batch sizes are small, and control flow essentially dominates execution, exactly where tightly integrated heterogeneous processors — with CPUs, fixed-function acceleration, and low-latency interconnects — are a better architectural match than discrete GPU-only platforms.

An elephant in the room

Despite all of this, there is one elephant in the room for Intel to address. The programmable logic mentioned by the Intel chief exec. Given how fast AI models, and therefore workloads, are evolving, fixed-function silicon would limit flexibility, falling back to CPU and GPU hardware, and to a large degree destroying synchronization. This is where programmable logic comes into play, as it allows frequently used but evolving parts of the workload to be accelerated. However, this part requires very close integration of hardware and software. Whether or not Intel's oneAPI can deliver this remains to be seen, and might even demand its own analysis.

The hybrid future

Intel has quietly brought a hybrid AI processors back onto its roadmap: a mysterious design that combines an x86 CPU, fixed-function AI acceleration (with no details about the architecture) to its roadmap. Coupled with Crescent Island and Jaguar Shores products, this one may offer something that AMD and Nvidia do not address at this time.

If executed well, the hybrid approach could let Intel differentiate through tight integration and programmability, though its success will depend on software maturity, particularly around oneAPI.

The talk of the town in the world of AI right now is almost exclusively about the amounts of money shifting from one company to the next, but the latest event might have farther-reaching effects than most business deals. Eric Demers, who designed ATI's best GPUs and spearheaded almost all of Qualcomm's Adreno designs, has now joined Intel's GPU team "with a focus on AI."

The blue team's GPU efforts are all but guaranteed to be significantly bolstered by Demers, a particularly welcome development in these troubled times for the company. According to Moor Insights and Strategy, this move is "bigger than people realize", as "[Demers] is an executive, but also he is a GPU architect, of which there are not that many that are at the level that he is at because he can basically build a GPU architecture from the ground up.”

Although so far Intel has been quiet on the exact wording of his new position, the reports so far predictably indicate that Demers will be in charge of designing AI accelerator GPUs, much to the chagrin of hopeful gamers who would like to see Intel's Arc series get an influx of brainpower.

Nvidia and AMD's accelerators are the first ports of call for datacenter-grade AI chips, and Intel wants in on that action, having produced three generations of Gaudi accelerators. The last one, Gaudi 3, is from 2024 and was presented as a more affordable alternative to Nvidia's now-aging H100. Gaudi is set to be superseded in the coming years by Falcon Shores and Jaguar Shores chips. The Shores silicon will exist alongside Crescent Island, a bespoke design for inference tasks.

The chip architect knows GPUs from the first transistor to the video outputs, having spent most of career designing them for AMD (formerly ATI), and Qualcomm in the past 14 years. His designs live in millions of smartphones right now as part of Snapdragon chips, and he was the lead architect for ATI's R300 and R600 series. For those who remember the names, he was at Silicon Graphics and even Matrox during his early years.

The R300 is fondly remembered by any techie in the early 2000s in the form of the Radeon 9700 and 9500 series that delivered a one-two punch to Nvidia's offerings of the time, namely the much-maligned FX 5800, known still today as the Dustbuster. When AMD absorbed ATI, Demers became the company's graphics Chief Technical Officer, a position he held until 2012 when he joined Qualcomm, and now Intel in 2026. Reports of loud expletives heard from Nvidia's offices are as of yet unconfirmed.

The artificial intelligence industry is developing so rapidly that the leading suppliers of AI accelerators — AMD and Nvidia — have moved to a yearly product release cadence. Furthermore, it appears that hyperscalers who can afford to develop their own silicon followed suit, so Amazon Web Services, Google, and Meta are also going to release new AI accelerators every year through to the late 2020s.

But which processors are on the horizon? We drew the big picture of the AI and HPC accelerator industry over the next several years. Here's what that looks like.

AMD

Traditionally, high-performance AI and HPC accelerators from the early 2020s — such as AMD's Instinct MI100 and MI200-series, and Nvidia's A100 and H100-series — were essentially the same product. This is perhaps why many hyperscalers have decided to build their own custom accelerators, dedicated specifically to AI workloads, to optimize costs, performance, and power consumption.

While Nvidia's Blackwell and Blackwell Ultra GPUs are tailored primarily for AI and FP32 and FP64 performance, they're not competitive in HPC because they lack sufficient FPUs with appropriate capabilities. AMD's latest Instinct MI350-series is still aimed at both AI and HPC workloads. Luckily, the new lineup of AMD's compute GPUs supports FP4 and FP6 data formats for AI inference, but since the GPU also supports FP64, the company had to sacrifice some performance in lower-precision workloads.

However, things are going to change for AMD with the Instinct MI400-series, set to land sometime in the second half of 2026. The upcoming MI450X will focus on AI workloads, while the MI430X will target traditional supercomputing applications. Both processors are expected to be made using TSMC's N2 (2nm-class) fabrication process, packaged using CoWoS-L technology, and equipped with HBM4 memory.

Each Instinct MI400-series processor will be built on different subsets of AMD’s CDNA Next architecture, according to reports. The MI450X will focus on low-precision formats such as FP4, FP8, and BF16, while the MI430X will support high-precision formats like FP32 and FP64. This separation is expected to help AMD eliminate unnecessary compute blocks from each chip, thus ensuring more efficient use of silicon and better tuning for specific workloads.

Both accelerators will include Infinity Fabric and UALink connectivity. While this makes them among the first GPUs to integrate UALink, adoption is expected to be limited at launch because external partners like Astera Labs, Auradine, Enfabrica, and XConn are not expected to have switching hardware ready by the second half of 2026.

Without these switches, large-scale deployments using UALink will not be possible in 2026, restricting systems to small-scale topologies like mesh and torus. Nonetheless, AMD will still offer its Helios rack-scale solution with 72 GPUs, which will scale out using Ultra Ethernet technology already supported by existing network cards, including AMD's own Pensando Pollara 400 and the upcoming Pensando Vulcano cards.

AMD's Instinct MI400-series will be followed by the Instinct MI500-series GPUs, which are expected to hit the market in 2027. Therefore, expect MI500 processors to be made on TSMC's N2 production node and packaged using CoWoS-L, though it remains to be seen whether AMD adopts HBM4E for these units. The Instinct MI500-series processors will power AMD's next-generation AI rack-scale solution, which will carry 256 GPUs..

Amazon

Amazon exclusively uses its AI accelerators at its own data centers, so the company does not disclose too many details about its two chips. Amazon uses its Trainium chips for both training and inference, its Inferentia chips solely for inference workloads.

Amazon has successfully deployed at least two generations of Trainium and Inferentia processors, including the latest Trainium2 Ultra (667 BF16 TFLOPS, 1300 FP8 TFLOPS, 96 GB of HBM3E, CoWoS-R). Interestingly, it appears that Amazon does not have plans to build more Inferentia processors in the coming years, as it intends to focus on Trainium. However, this change has not been officially confirmed by the company, and it remains a rumor.

Amazon's plans for late 2025 – early 2026 include Trainium3, which is expected to offer higher performance, support for new data formats, and up to 128 GB of HBM3E memory onboard. The chip is expected to be produced by TSMC on one of its 3nm-class nodes, most likely N3P. The Trainium3 chip will be followed by Trainium3 Ultra, featuring 128 GB of HBM4 in 2026 - 2027, which will further increase performance. After that, expect Amazon to release a 2nm-based Trainium4 chip in 2027 – 2028, though there are currently no details on estimated performance or features.

Google

Google has been developing its AI accelerators since 2015, meticulously increasing performance and adding features to its performance (e.g., TPU v5p) and efficient (e.g., TPU v5e) tensor processing units (TPUs).

Last year, the company introduced its Trillium (TPU v6e) chip, equipped with 32GB of HBM3, aimed primarily at low-power inference workloads. This year, the company rolled out its Ironwood accelerator, its 7th Generation TPU, built for large-scale training and inference workloads. Each chip is made on TSMC's 3nm-class process technology, which features 4,614 FP8 TFLOPS performance (10 times more than its predecessor, TPU v5p, but only slightly higher than Nvidia's H100) and comes with 192 GB of HBM3E.

The unit delivers 7.37 TB/s of HBM bandwidth, around 2.64 times that of v5p, and 1.2 TB/s bidirectional Inter-Chip Interconnect (ICI) bandwidth. Ironwood pods scale up to 9,216 chips, delivering a total of 42.5 FP8 ExaFLOPS, which makes it one of the most powerful systems built to date. In terms of efficiency, Ironwood offers around two times better performance per watt compared to Trillium, and is nearly 30 times more power efficient than Google’s first Cloud TPU.

After Ironwood (TPU v7p), Google is expected to release its 8th Generation TPUs — v8p and v8e — which are rumored to be made on TSMC's 3nm-class process technology and feature up to 288 GB of HBM3E memory. So, do not expect a major performance increase from these parts. Google's v8p and v8e accelerators are slated for 2026.

Google's TPUs will likely see a major performance increase in 2027 or 2028, when the company rolls out its 9th Generation TPU based on an all-new architecture with HBM4 memory. These parts are projected to be created using TSMC's N2 fabrication process and use CoWoS-L packaging, which suggests very high internal bandwidth for system-in-packages. Some believe that TSMC and Google might even adopt hybrid bonding, but that remains to be seen.

Intel

Although Intel's CPUs are widely used in AI servers, it does not look like its Gaudi 3 accelerators have gained any traction so far. While multiple loyal partners offer Gaudi 3-based servers, with Dell offering a workstation with a Gaudi 3 card, the company's share of the AI accelerator market is negligible.

One reason large companies might hesitate to develop or deploy AI models for Gaudi is that Intel has already announced Gaudi will be discontinued when it launches its AI GPUs in 2026 – 2027. Investing tens of millions of dollars in a platform set to go extinct in a couple of years is not something big players do, so Gaudi is unlikely to take off. In fact, it remains to be seen whether it ever releases a refresh for its Gaudi 3 processor.

As for Intel's GPU plans, the firm will use its codenamed Falcon Shores GPU for internal development purposes, and may offer access to select AI companies. Intel's first compute GPU for AI workloads, codenamed Jaguar Shores, will be available to a wide range of clients and is set to be released in 2027.

Meta

Although Meta is among the leaders when it comes to hardware investment in AI, the company's in-house AI silicon efforts are behind those of its rivals. Meta's own AI chips are called Meta Training and Inference Accelerators (MTIA). They are developed in collaboration with Andes, which provides RISC-V-based processing elements (PEs), and Broadcom, which designs processors that use a systolic array architecture.

Meta's 1st Generation MTIA v1 was introduced in 2023: it used a chip built by TSMC on its 7nm-class process technology, and was equipped with 64GB of LPDDR5 memory. Meta itself stated that MTIA v1 was deployed in its data centers and was used to serve recommendation and ranking models in production.

However, that deployment seems limited to internal workloads and was not necessarily at the scale one would expect for a full infrastructure shift. The company's more recent MTIA chips are made on TSMC's 5nm-class fabrication process and double onboard memory to 128 GB (MTIA 2) and 256 GB (MTIA 2.5). However, the company will get more aggressive with subsequent generations of MTIA.

Meta's MTIA v3 — due in 2026 — is projected to be a considerably higher-performance solution, as it's expected to use a compute chiplet made on TSMC's N3 fabrication process, and is expected to use HBM3E memory. The company is also expected to release MTIA v4 in 2027. This accelerator will likely use two or more chiplets fabbed on TSMC's 2nm fabrication process and equipped with HBM4 memory.

Microsoft

Microsoft was a bit late to the custom AI silicon party with its its Maia 100-series launching in late 2023, years after Amazon, Google, and Meta. Maia 100 is built on a 5nm-class production node by TSMC, contains 105 billion transistors, and is equipped with 64 GB of HBM2E memory.

According to Microsoft's own blog, Maia 100 has been deployed in Azure to support large‑scale AI training and inference workloads. However, external reporting suggests that Maia's deployment has not been widespread across Microsoft's entire service infrastructure, and the vast majority of Azure AI services still rely on partner hardware, such as GPUs from AMD and Nvidia.

The company is currently working on its next-generation Maia processors: the codenamed Braga (Maia 200?) chip will use TSMC's 3nm node and HBM4 memory. Braga is allegedly due in 2026, with its successor, Clea (Maia 300?), due at a later date. However, considering the limited adoption of Maia, Microsoft might want to recalibrate the positioning of its own AI accelerators to reduce complexity and risk.

Nvidia

While many companies like Broadcom and Marvell develop XPUs for AI workloads, Nvidia continues to lead the market for both training and inference workloads. The company's latest Blackwell Ultra architecture sacrifices INT8, FP32, and FP64 performance in favor of NVFP4 performance, mainly used for inference. So, Nvidia is certainly doing everything to keep its crown as the dominant force in the AI industry.

Nvidia's upcoming data center GPU carries the codename Rubin, and is expected in late 2026. The initial model, tentatively referred to as R100/R200, will consist of two reticle-sized GPU dies and two dedicated I/O chiplets, all built using TSMC's 3nm-class node, likely N3P or something customized for Nvidia's needs. Memory-wise, it will integrate 288 GB of HBM4 across eight stacks, each running at 6.4 GT/s, yielding an impressive ~13 TB/s of total memory bandwidth.

The R100/R200 is designed primarily for AI acceleration; it is expected to hit 50 PFLOPS of NVFP4 performance for inference, and roughly 17 FP8 TFLOPS for training workloads. Performance data for other formats has yet to be shared, but incremental improvements across the board are anticipated, relative to the Blackwell generation. This performance uplift will come at a cost: each VR200 unit is projected to draw 1,800 watts, posing new power and cooling demands on data center infrastructure.

But Nvidia's Rubin GPUs for AI will not come alone. Nvidia is also set to offer its Rubin CPX, a newly introduced GPU, which is designed to handle the context phase of long-context inference. This workload is becoming increasingly common in next-generation AI models, which must process up to 1 million tokens before generating output.

Rather than relying solely on high-power, high-bandwidth GPUs like the Rubin R100/R200 (which uses HBM4), CPX offloads this specific task onto a compute-dense (but bandwidth-light) CPX GPU that features 128 GB of GDDR7, offering a cheaper, cooler, and simpler memory alternative to HBM. With 30 petaFLOPS of NVFP4 compute performance, hardware acceleration for attention mechanisms, and even support for video encoding/decoding, CPX is designed specifically for rapid input processing at a relatively low cost and power consumption. Nvidia's Rubin CPX GPU will work alongside Rubin and Vera CPUs in the NVL144 CPX system, which will deliver 8 exaFLOPS of NVFP4 compute and 100 TB of memory per rack.

In 2027, Nvidia plans to introduce a substantially upgraded variant: Rubin Ultra (VR300). This will double the compute complex to four reticle-sized GPU tiles, alongside two I/O dies, and support 1 TB of HBM4E memory made up of 16 stacks, offering a blistering 32 TB/s bandwidth. Targeted FP4 performance is 100 PFLOPS, making it twice as fast as VR200 for inference.

However, the VR300’s scale demands enormous power — 3,600 watts per package — which makes it suitable only for highly specialized, high-density deployments with advanced liquid cooling systems. Nvidia will continue using CoWoS-L packaging for both VR200 and VR300, but the Rubin Ultra variant's footprint requires either TSMC's forthcoming 9.5-reticle CoWoS-L interposer (sized 120 × 150 mm) or an arrangement of stitched smaller interposers, since no vertical die stacking is visible in official slides.

Looking beyond Rubin, Nvidia's 2028 plans include a follow-up GPU family codenamed Feynman. While specific architectural details are unavailable, it is noteworthy that these processors will likely be made on TSMC's A16 process technology with a backside power delivery (which will provide an ultimate uplift both for transistor density and for performance) and are said to adopt next-generation HBM. Whether this refers to a specialized form of HBM4E or early HBM5 is still unclear — though the latter seems premature for that timeframe.

OpenAI

OpenAI is perhaps the most known AI company due to popularity of its ChatGPT service, but for AI hardware, it is a new kid on the block. The company has reportedly been working on its own AI accelerator since at least late 2023 presumably with Broadcom.

Recently, Broadcom confirmed that an undisclosed client intends to procure $10 billion worth of custom AI processors, which are set to be delivered in the third quarter of 2026. While the industry believes that the product in question is OpenAI's first custom AI processor, this has never been formally confirmed.

Although OpenAI will allegedly spend $10 billion procuring custom processors made according to its needs, which may point to 1 – 2 million of XPUs, depending on unit prices, its recent agreement with Nvidia indicates that it will gain access to $100 billion worth of Nvidia GPU hardware, presumably over several generations of GPUs. This suggests that the lion's share of OpenAI's workloads will still rely on Nvidia and its models will be optimized for the CUDA platform.

This article covers our picks for the best graphics cards for gaming in 2026. Amid the AI gold rush and consequent supply crunch for consumer silicon, no truly new gaming GPUs have been introduced in almost a year. If you haven't already upgraded your graphics card after the GeForce RTX 50-series and Radeon RX 9000-series launches in 2025, well, you're still looking at the exact same products now.

AMD did make its formerly China-only Radeon RX 9070 GRE available globally after Computex 2026, but in our review, we found that $549 product to be too expensive given the level of performance it delivers and the compromises made to hit its price point, so it isn't joining the list here. Check out that coverage for all the details.

Most of the products we recommend remain at elevated prices compared to their MSRPs, but this is just life in mid-2026.

It's admittedly cold comfort, but unless you're shopping for an RTX 5090, graphics card prices haven't risen much more than they already did earlier this year. Compared to the doubling or tripling of prices we've seen for RAM kits and SSDs in 2026 versus last year, a GPU upgrade remains a relatively affordable (and self-contained) option, either as a boost for an existing PC or part of an all-new parts list.

Even if you can’t build an all-new system, you can just put a new graphics card in an older PC and still enjoy boosts to gaming performance, image quality, or both—especially if you can upgrade your monitor at the same time.

As we discuss in further depth below, the arrival of DLSS 4.5 upscaling (for RTX 40-series and 50-series cards, at least) and expanded multipliers for Multi-Frame Generation, which now can boost frame rates by up to 5x or 6x, means that driving a high-resolution, high-refresh-rate monitor is now easier than ever if you're considering a GeForce RTX 50-series graphics card.

Read on to see our picks in today's gaming graphics card market.

Best graphics cards for gaming, at a glance

The above list shows all the latest-gen graphics cards we feel stand out in their segments. If you want to see how all of the current and prior generation GPUs stack up, check our GPU benchmarks hierarchy. We also have performance benchmarks further below.

When accounting for pricing, we perform our own research to find the midpoint of current prices for a given graphics card, rather than taking a vendor's MSRP at face value. We feel this method tends to be most representative of the price you're likely to see for products in stock.

If you can find a card for less than this midpoint, it's likely closer to (or even less than) a vendor's MSRP and a better value. Conversely, if you find one for more than this midpoint, it could be a worse value (or too close in price to a more powerful card that's a step up). Tread carefully.

The overall performance ranking incorporates 19 games from our 2026 test suite, which takes the geometric mean (i.e., equal weighting) for both rasterization and ray tracing games. Note that we are not including any upscaling or frame generation results in the table.

Raw performance may be the most important consideration for most gamers, but it's not the only metric that matters. Our subjective rankings below factor in price, power usage, and power efficiency, and features colored by our own years of experience. Others may offer a slightly different take, but all of the cards on this list are worthy of your consideration.

Upscaling and frame generation mean GPUs are more than just a chip

GPU performance goes beyond the hardware these days. Choosing a particular GPU vendor means you're buying into a complex software stack that includes upscaling, frame generation, and (more rarely) AI-powered RT denoising technologies.

In Nvidia's corner, the DLSS 4.5 upscaling model and its second-generation transformer architecture offer superior image quality to other upscaling tech (and with lower input resolutions, meaning higher potential performance), but it's more computationally expensive than past DLSS models and works best on RTX 50-series and 40-series cards.

The DLSS 4 model and its first-gen transformer architecture still work with cards going all the way back to the RTX 20-series family. Not all games implement DLSS 4 natively, but Nvidia allows you to force the usage of that model in many older titles through the Nvidia App utility, so you can practically always get the latest and greatest.

Between native support and driver overrides, DLSS is available in virtually any modern game you might want to play. Nvidia recently marked DLSS feature availability in over 1000 titles.

RTX 50-series GPUs are Nvidia's first with support for multi-frame generation (MFG), which allows Blackwell GPUs to insert anywhere from one to five AI-generated intermediate frames between each native one (for a 2x, 3x, 4x, or even 5x or 6x frame rate boost). RTX 40-series GPUs also support framegen, but only with a 2x boost.

Meanwhile, AMD's FSR 4 offers AI-enhanced upscaling with superior image quality to other FSR versions, but official support for it is limited to RX 9000-series Radeons for now. AMD will bring FSR 4 upscaling to RX 7000-series cards in July 2026 and RX 6000-series cards in early 2027.

In the meantime, AMD's FSR 3.1 and earlier upscalers still work on any GPU, but the image quality tends to be noticeably lower than both DLSS and FSR 4.

AI-enhanced FSR framegen (aka ML Frame Generation) arrived on AMD cards as part of the FSR Redstone update late last year. Like FSR 4 upscaling, ML Frame Generation is limited to Radeon RX 9000 cards, and it can be enabled in compatible games using a control panel override for titles that don't natively have it.

Legacy FSR frame gen remains available, too. Its framerate-doubling boost remains cross-compatible with GPUs from all vendors, but its image quality can't keep up with the AI-powered frame gen tech of the latest AMD and Nvidia models.

Intel XeSS upscaling can be superior to FSR 3.x, but it isn't available in as many games as FSR or DLSS. It works best on Arc GPUs, but like FSR, it's cross-compatible with a wide range of graphics cards from AMD and Nvidia if you need it.

XeSS 2 with AI-enhanced frame generation is now available in 95 games as of this writing and requires an Arc GPU. XeSS 3 brings multi frame generation to the party through both native support and a driver override in compatible titles.

All that said, we don't think you should go out of your way to buy an Intel Arc card for gaming in 2026 for reasons we'll get into later.

1. Best high-end graphics card: GeForce RTX 5070 Ti, $1099.99

If you want the best blend of high performance and cutting-edge graphics tech out there for 1440p or 4K gaming, the GeForce RTX 5070 Ti is it. This card comes with full support for Nvidia’s latest DLSS 4.5 upscaling and Multi Frame Generation tech, and its 16GB of VRAM gives you full freedom to enable every DLSS 4 feature.

AMD’s closest competitor, the Radeon RX 9070 XT, is way cheaper than the RTX 5070 Ti right now, but the AMD card obviously doesn’t support DLSS 4 or MFG. For the privilege of those capabilities, you'll generally need to spend a whopping 45% more cash right now for just 5% more baseline performance than AMD’s best before you start enabling all the DLSS 4 features Blackwell supports.

Is that worth it? Yes, if you can swing it. Here's why: getting the best gaming performance on modern graphics cards is as much a software problem as a hardware one. Tuning your gaming experience to taste requires access to high-quality upscaling, frame generation, and (more infrequently) an AI-powered RT denoiser like DLSS Ray Reconstruction.

Even amid its shift to AI and data center products, Nvidia ensures that its full suite of DLSS tech is adopted in virtually every new game, whereas AMD's support of FSR 4 adoption has become rather hit-or-miss.

The RTX 5070 Ti also offers superior RT performance versus the RX 9070 XT across our 2026 test suite. Beyond that baseline, Nvidia is working with developers to enable impressive path-traced effects in many of the latest AAA releases.

In our recent experience, path-traced games play best with DLSS 4.5 upscaling and MFG at your disposal, and being able to consistently rely on the availability of those features makes the extra cash for the 5070 Ti worth it.

Our recent GPU Hierarchy retesting has shown that high-end graphics cards are becoming five- to eight-year investments, and Nvidia's ongoing commitment to developer relations and new software features means that you'll enjoy a first-class gaming experience throughout the life of your 5070 Ti no matter what games you want to play on it.

Spread out over that time span, the extra cost of the RTX 5070 Ti versus the RX 9070 XT is worth it for the better experience.

What about the RTX 5080? Nvidia's second-fastest Blackwell card is anywhere from 8% to 16% faster than the 5070 Ti, with the biggest gap at 4K. Prices for the 5080 in June 2026 remain insane, however, and at the midpoint of current prices, the 5080 is 33% more expensive than the 5070 Ti.

There's no way the RTX 5080 offers anywhere close to enough value for the money to justify the step up right now unless you're looking for the fastest thing this side of a 5090 for 4K gaming.

Read: Nvidia GeForce RTX 5070 Ti review

2. Best enthusiast value graphics card: Radeon RX 9070 XT, $759.99

The Radeon RX 9070 XT is AMD's most well-rounded graphics card in years. It delivers raw gaming performance within spitting distance of the GeForce RTX 5070 Ti for far less money, making it a strong value at first glance. But that appealing price tag comes with a number of asterisks in mid-2026 that mean it's no longer our first pick for high-end PC gaming.

AMD shored up two of its greatest weaknesses against Nvidia with the RX 9070 XT's RDNA 4 architecture: RT performance and AI acceleration, both of which are closer to Nvidia's latest and greatest. And AMD did all that while keeping power efficiency right there with Nvidia, too.

The FSR 4 upscaler is a big jump in image quality over FSR 3, and FSR ML Frame Generation now offers higher-quality framegen on the RX 9070 XT than FSR 3's approach, although it's still limited to a simple doubling of frame rates versus DLSS Multi Frame Generation's versatility.

The problem for the RX 9070 XT in mid-2026 is that FSR 4.x upscaling still trails Nvidia's flagship DLSS 4.5 in image quality, and AMD isn't driving the adoption of FSR 4 features nearly as aggressively as Nvidia is for DLSS. Driver-level overrides for those features can't entirely close the gap.

Worse, you might find the RX 9070 XT entirely shut out of features that you might want to enable in certain games. For just a couple of examples, Radeon gamers can't enable path-traced effects at all in recent titles like Pragmata and Resident Evil Requiem, and 007 First Light sticks you with FSR 3 upscaling that can't be overriden through a driver toggle due to the way it's implemented.

And in our latest GPU Hierarchy retesting with the RX 9070 XT, we saw major performance issues in Grand Theft Auto V Enhanced and minor visual corruptions in Stalker 2 that weren't present on GeForces or on RX 7000- or RX 6000-series cards. We don't think these issues should have slipped past any QA program, especially for such popular and high-profile games, but they sting especially hard on a current-gen product.

All that means the overall ownership experience of an RTX 5070 Ti and an RX 9070 XT is significantly different in mid-2026. We think that gamers shopping in this price class should be able to expect a consistently high level of software feature support and quality across all the games they might want to play, and Nvidia provides that assurance better than AMD does right now.

If you're willing to gamble with the availability of FSR 4 features, for RT or path-traced effects, and don't care to tune the smoothness of your gaming experience with frame generation, the RX 9070 XT's shortcomings versus the RTX 5070 Ti may be easier to overlook given the large amount of cash that will remain in your pocket.

But we also think that you should look closely at what you're giving up before reflexively choosing an RX 9070 XT over an RTX 5070 Ti, despite its strong value at a glance.

Read: AMD Radeon RX 9070 XT review

3. Best midrange graphics card: GeForce RTX 5070, $659.99

Until 2026 rolled around, the GeForce RTX 5060 Ti 16GB had been our entry-to-midrange Blackwell gaming favorite. But the $579 midpoint of current pricing puts the 5060 Ti 16GB's on-shelf price above that of the RTX 5070's $549 MSRP, and the 5070 is one of the least marked-up graphics cards out there at the moment.

As a result, it's possible to find RTX 5070s for about $670, and that makes it an easy call to step up for less than $100 more than 5060 Ti 16GBs.

The RTX 5070 is about 30% faster than the RTX 5060 Ti 16GB across our 2026 test suite, and that's a huge leap that you'll easily see on the right monitor for just 15% more money.

The advent of DLSS 4.5 upscaling, which makes it possible to achieve superior delivered image quality at lower input resolutions than older DLSS versions, also takes some VRAM pressure off the RTX 5070's 12GB of GDDR7, making the deployment of RT and DLSS MFG more practical on this card than it has been in the past.

And as with the RTX 5070 Ti, the universal availability of DLSS 4.5 (both natively and through app overrides) plus MFG makes this card a fast and flexible performer across all of the games you might want to play in 2026.

Given the image quality of DLSS 4.5 and the smoothness boost of MFG, along with the higher baseline RT performance of this card versus the 9070 in our 2026 testing, we think the 5070 should be your first pick for a midrange gaming card right now.

In an ideal world, the RTX 5070 would have more VRAM to allow for unhindered exploration of everything DLSS 4 and MFG have to offer, especially at a native 4K resolution. If you're pushing those limits, we'd still recommend the Radeon RX 9070 thanks to its 16GB of VRAM.

But if you're on a 1440p monitor where VRAM is less of an issue and want DLSS 4.5 over FSR 4, as most gamers do, the RTX 5070 is still a strong performer, and you're less likely to run into its limits.

Read: Nvidia GeForce RTX 5070 review

4. An AMD midrange alternative: Radeon RX 9070, $629.99

If you primarily play raster titles at native resolution and aren't on board with upscaling or framegen, the Radeon RX 9070 remains a strong midrange alternative to the RTX 5070. It's one of the least marked-up 16GB graphics cards available even with today's AI headwinds, and in a world where MSRPs have largely been forgotten, that makes the RX 9070 a strong value.

The GeForce RTX 5070 and RX 9070 go neck-and-neck in our test suite, but the RX 9070 has 16GB of VRAM and the RTX 5070 has just 12GB. Especially if you're trying to push 4K games at native resolution, that extra VRAM matters.

But the advent of DLSS 4.5 upscaling, which provides image quality that's practically indistinguishable from native rendering even at relatively low input resolutions, means that the RTX 5070 is a more potent midrange graphics card in 2026 than it was at launch.

On top of their inherent technical superiority, you can find DLSS 4 and MFG in most every game released today, which can't be said for FSR 4.x upscaling or ML framegen. AMD's driver overrides make up some of the gap, to be sure, but not all of it.

And as with the RX 9070 XT, AMD gamers may find themselves locked out of certain features like path tracing (in major releases like Pragmata and Resident Evil Requiem) or FSR overrides (in 007 First Light) entirely.

The RX 9070 is subject to the same minor image quality issues and performance hitches we saw with the RX 9070 XT, and those issues could certainly be overcome with future software updates. But we think that if you're looking for the best midrange graphics card, it should be free of those issues entirely, and so the RTX 5070 is our first pick for this price point right now.

Read: AMD Radeon RX 9070 review

5. Best enthusiast value graphics card: Radeon RX 9060 XT 16GB, $469.99

AMD's Radeon RX 9060 XT 16GB can handle basically anything the mainstream gamer can throw at it at 1920x1080 and 2560x1440, all at a price that comes in way under the sky-high markups on the RTX 5060 Ti 16GB as of this writing.

At a midpoint of $459.99 in mid-2026, prices for the RX 9060 XT 16GB are the highest we've ever seen, dulling the 9060 XT 16GB's reputation as the value-minded builder's GPU of choice. But with no cheaper Radeons worth recommending in the lineup, what can you do?

In any case, the RX 9060 XT enjoys the much-improved ray-tracing and AI performance of the RDNA 4 architecture, both of which bring Radeons a lot closer to the latest Nvidia competition. And its 16GB of VRAM gives mainstream gamers the assurance they'll basically never find VRAM a bottleneck in modern games at 1080p and 1440p resolutions.

Like the RX 9070 XT, the 9060 XT 16GB gives you access to AMD's much-improved FSR 4 upscaling tech, allowing you to boost performance with a small hit to image quality in the small but growing list of titles that support it.

Even with its new ML-powered model, FSR Frame Generation remains limited to a doubling of output frame rate at best, so it’s not a direct competitor to Nvidia’s DLSS 4 with MFG.

If you want more frames, AMD just launched the RX 9070 GRE globally for $549, and it provides a decent step up in performance for less than $100 more, especially if you're only gaming at 1080p or 1440p. But the more powerful RX 9070 can be found for just $50 more than the GRE, and then you're contemplating the even more powerful and versatile RTX 5070, too.

The RTX 5060 Ti 8GB is the RX 9060 XT 16GB's closest Nvidia competition, dollar for dollar, but we can’t recommend it at all. If you're spending over $350 on a GPU, we don't think you should have to fine-tune every setting to avoid running out of VRAM. The RX 9060 XT is easy to live with for a wide range of gamers in a wide range of titles, and that’s why it won our Editor’s Choice award.

Read: AMD Radeon RX 9060 XT 16GB review

6. The best graphics card for 1080p gaming: Nvidia GeForce RTX 5060, $369.99

If you want to get your game on at 1080p, we think the RTX 5060 is still your best bet in mid-2026. The midpoint of RTX 5060 prices is around $370 right now thanks to the AI crunch, but you can still find them for as little as $350 if you're willing to shop around.

The RTX 5060 has impressive baseline performance for 1080p gaming in wildly popular titles like Fortnite, Counter-Strike 2, Marvel Rivals, and Apex Legends that aren't hungry for giant pools of VRAM. And if you are trying to push higher output resolutions in demanding AAA games, the universal availability of DLSS 4.5 upscaling means that it's easy to achieve near-native image quality at lower input resolutions than before, making the RTX 5060 a more flexible performer than ever.

If you can tune your settings right, enabling DLSS 4 Multi-Frame Generation could make for an even smoother ride on this card, but we find that 8GB of VRAM isn't enough to consistently enable framegen in the titles where you'd really want it. The feature often doesn't work if you're already at the limits of the RTX 5060's memory pool (or that of any 8GB Blackwell card), since the MFG AI model needs some VRAM of its own to run.

AMD's toughest competition for the RTX 5060 is the RX 9060 XT 8GB, which also lists for $299 but is now selling for about the same $350 as you'll see RTX 5060s going for. Supply of those cards has largely dried up in mid-2026, however, and you're likely to see only a couple options for them from any e-tailer.

Despite its much-maligned 8GB of VRAM, the 9060 XT 8GB put in a strong showing in our 2026 GPU Hierarchy testing, but not consistently enough to beat out the RTX 5060 and take home our general recommendation.

When the RX 9060 XT can bring its full compute horsepower to bear in certain games, it can handily outpace the RTX 5060, so it's worth checking out results like those from our RX 9070 GRE review and seeing whether a game you love benefits from the Radeon's raw muscle.

But if you want a more consistently solid gaming experience, we'd still recommend the RTX 5060.

Read: Nvidia GeForce RTX 5060 Ti review

7. The best graphics card, period: GeForce RTX 5090, $4299

There's nothing else like the GeForce RTX 5090. If you want to turn on every bell and whistle in modern games at 4K (or beyond), the RTX 5090's sheer shader and Tensor Core horsepower, along with support for Nvidia's DLSS 4 upscaling and multi-frame generation, lets you tune your gaming experience to perfection even on high-refresh-rate 4K displays.

If you're a hardcore PC gamer who demands only the best, the hair will stand up on the back of your neck when you watch the RTX 5090 breeze through workloads that other graphics cards leak out all their thermal gel about.

Prices for the RTX 5090 have always been elevated, but they're stratospheric in early 2026. Major e-tailers only have a few different models listed, and prices start at $3500 or so and only go up from there. Nvidia's $1999 MSRP is pure imagination in current market conditions.

At those prices, an RTX 5090 is an indulgence of the highest order, but then again, it always has been. Without a compelling AMD alternative even on the horizon, considerations of value don't really apply here. If you truly need (or want) this class of gaming or AI performance, you're going to have to pay up.

This card needs a system with a massive power supply, one of our best gaming CPUs, and a top-shelf monitor to take full advantage of its astounding capabilities, and all those spendy components add up quick. But if you have a big enough bankroll to consider shopping for a graphics card of this caliber, you probably don't need us to tell you all that.

If Nvidia and its industry partners fixed the meltdown-prone ATX12V-2x6 connector, the RTX 5090 would be as close to gaming perfection as any graphics card that's ever been made. Guess that's something to improve on the RTX 6090, if it ever arrives.

Read: Nvidia GeForce RTX 5090 Founders Edition review

8. The cheapest graphics card worth buying: GeForce RTX 5050

If you have to save every penny on a new graphics card in 2026, you're hard up for good budget options. We previously recommended Intel's Arc B570 here, but after completing our 2026 GPU Hierarchy retesting, we're bumping it in favor of the RTX 5050, which is currently selling for about $300, or about $50 more than the B570.

Here's why: we think if you're spending any amount of money on a graphics card, it should just work. You should expect consistent feature support over time in games, universal support for upscaling and (optionally) frame generation when you need them, and consistently high performance in games.

The RTX 5050 unreservedly checks all those boxes, while we couldn't even complete our testing of the Arc B570 (or B580) for our 2026 GPU Hierarchy until the literal day before this guide update goes live, due to a months-long settings lockout with UE5's Nanite and Lumen in a little title you may have heard of called Fortnite.

We can't say when a similarly major issue might occur again with the Arc B570 in any game, and so we're no longer recommending it. Unless you're willing to gamble and need to save every possible dollar on a graphics card, we think you should just save up a bit more cash and buy an RTX 5050.

The RTX 5050 isn't the fastest GPU around, to be sure, and its 8GB of VRAM is a constraint for anything beyond 1080p gaming in mid-2026. But it delivers solid enough native raster performance at 1080p, and it beats out the Arc B570 even before you enable DLSS 4.5 upscaling. And if you do want the performance boost of DLSS, you're getting access to the best and most widely adopted upscaler on the market.

On top of that, the extra $50 over the Arc B570 means that you have the full strength of Nvidia's developer relations team and software support behind you when you go to play the latest games, and we think that reliable software support makes all the difference between a GPU that's fun and affordable and one that's merely cheap.

Read: GeForce RTX 5050 review

How we test the best graphics cards

Determining pure graphics card performance is best done by eliminating all other bottlenecks — as much as possible, at least. To that end, we've selected components for our test rig , most notably AMD's Ryzen 7 9800X3D CPU, one of the best CPUs for gaming.

We test across the three most common gaming resolutions, 1080p, 1440p, and 4K, using a mix of high and ultra settings, depending on the title. Where possible, we use 'reference' cards for all of these tests, like Nvidia's Founders Edition models and AMD's reference designs. Most midrange and lower GPUs don't get reference models, however, and in some cases we only have factory-overclocked cards for testing. We do our best to select cards that are close to the reference specs in such cases.

For each graphics card, we follow the same testing procedure. We run one pass of each benchmark to "warm up" the GPU after launching the game, then perform our actual test runs across each resolution.

We carefully review our test data and check for anomalies. For example, we always expect the RTX 5080 to be faster than the RTX 5070 Ti. If it's not, and we're not in a CPU limited situation, we'll recheck both cards to ensure that our standings our accurate. We also check and retest in cases of subtler issues, as when a transient hitch or frame-time spike causes a large dip in 1% low FPS.

Due to the length of time required for testing each GPU, updated drivers and game patches inevitably come out that can impact performance. We periodically retest a few sample cards to verify our results are still valid, and if not, we go through and retest the affected game(s) and GPU(s). We may also add games to our test suite over time, if one comes out that is popular and conducive to testing. See what makes a good game benchmark for our selection criteria.

Best graphics cards performance results

Our updated test suite of games consists of 19 games at present, eight of which have ray tracing enabled (or require RT to run at all).

We test without any upscaling or frame generation technologies enabled. We expect that most gamers will want to enable these features, but they complicate apples-to-apples comparisons between GPU vendors due to inherent differences in output image quality. To keep it simple, we present native resolution performance as a baseline.

The data in the following charts is from testing conducted during the past several months. We've tested all of the latest GPUs at every resolution and setting, even where it generally doesn't make sense (e.g. 4K with ray tracing at single digit framerates).

For each resolution and setting, the first chart shows the geometric mean (i.e. equal weighting) for all tested games. The second chart shows performance in the 11 pure raster games, and the third chart focuses in on ray tracing performance in eight games.

The charts below contain all the current Nvidia RTX 50-series and AMD RX 9000-series graphics cards. We're leaving Intel Arc cards out of the standings for now due to software compatibility issues with our test suite, and we'll include those results when those issues are corrected and we have the opportunity to retest them.

Our GPU benchmarks hierarchy contains additional data for every GPU spanning multiple generations of hardware. The charts are color coded with AMD in red, Nvidia in blue, and Intel in gray to make it easier to see what's going on.

The following charts are up to date as of June 2026.

Best Graphics Cards — 1080p

Best Graphics Cards — 1440p

Best Graphics Cards — 4K

Additional Shopping Tips

When buying a graphics card, consider the following:

• Monitor Resolution: The more pixels you're pushing, the more performance you need. You don't need a top-of-the-line GPU to game at 1080p, but you will certainly want more power at 1440p or 4K.
• PSU: Make sure that your power supply has enough juice and the right 6-, 8- and/or 16-pin connector(s). Nvidia, AMD, and Intel board partners will all make PSU recommendations alongside their products that you can use as a baseline, so if you're unsure whether your PC can provide enough power, be sure to check those spec sheets first. If you have an older PSU, be mindful that power supplies do lose capacity with time, so if you're contemplating a high-end GPU, it might be time to upgrade your GPU, too.
• Video Memory: In 2026, 8GB of VRAM is the bare minimum you'll want to play the latest games at 1080p, and it's the smallest amount of memory you'll find on a new card. Midrange cards tend to feature 12GB of VRAM, which is generally enough for raster gaming all the way out to 4K but may present limitations for RT even at 1440p. If you're planning to push a 4K display without upscaling or want to explore RT gaming without restriction, we recommend a 16GB card.
• FreeSync or G-Sync? Either variable refresh rate (VRR) technology will synchronize your GPU's frame delivery with your screen's refresh rate. Nvidia supports G-Sync and G-Sync Compatible displays (for recommendations, see our Best Gaming Monitors list). And most every G-Sync Compatible display also supports AMD FreeSync these days, so this vendor war is largely over.
• Upscaling and Frame Generation technologies: Nvidia's DLSS is in practically every game, and the latest DLSS 4.5 tech provides high-quality upscaling and frame generation (on RTX 40-series to boost performance to taste with practically no loss of image quality. AMD FSR 4 provides AI-enhanced upscaling on RX 9000-series cards, and a version compatible with RX 7000-series cards arrives in July 2026. Intel XeSS can deliver better image quality than older versions of FSR, but the core upscaler hasn't been updated in some time, and it's not as widely adopted as either DLSS or FSR, so it shouldn't influence your buying decision either way.

Finding Discounts on the Best Graphics Cards

While deep discounts are rare on graphics cards in 2026, you might find some particularly tasty deals on occasion. Check out the latest Newegg promo codes, Best Buy promo codes and Micro Center coupon codes for potential savings.

Want to comment on our best graphics picks for gaming? Let us know what you think in the Tom's Hardware Forums.

MORE: HDMI vs. DisplayPort: Which Is Better For Gaming?

MORE: GPU Benchmarks and Hierarchy

In a surprising twist, Intel announced on Thursday that its Falcon Shores GPU for AI and HPC applications will not be released to the market but will remain an internal test processor to develop the hardware and software foundations for its successor, codenamed Jaguar Shores. This decision makes Intel's struggling Gaudi 3 processor, which has suffered from limited uptake amid the company's disclosed software issues, the company's only viable solution for AI applications for the next two years. Meanwhile, the company will work on developing rack-scale solutions, the only true way to compete with AI behemoth Nvidia.

"Many of you heard me temper expectations on Falcon Shores last month," interim co-CEO Michelle Johnston Holthaus said during the company's earnings call on Thursday. "Based on industry feedback, we have decided to leverage Falcon Shores as an internal test chip. Without bringing it to market, we will support our efforts to develop a system-level solution at rack scale with Jaguar Shores to address the AI data center more broadly." 

The company originally expected Falcon Shores to serve the AI and HPC markets currently covered by Gaudi processors. However, a quarter ago, Intel's interim co-CEO told analysts and investors to temper their expectations, as Falcon Shores would mostly serve as a vehicle to develop a hardware and software ecosystem around its hybrid processors. 

Apparently, after careful consideration, Intel decided not to launch Falcon Shores commercially at all but to use it solely for internal development to perfect hardware and software rather than launching a half-baked product commercially that could potentially damage the company's reputation. 

Intel described Falcon Shores as its first multi-chiplet design featuring Xe-HPC (or Xe3-HPC) GPU chiplets for highly parallel AI and HPC workloads. The product was meant to greatly increase performance and performance-per-watt efficiency compared to Intel's AI and HPC processors, though the company refrained from giving actual numbers. 

Oddly, Intel stated that Falcon Shores would be an "internal test chip." When developing an ecosystem, some processors that are not launched commercially are still shipped to external partners, including independent hardware vendors (IHVs) and independent software vendors (ISVs). For example, Intel's first-generation Xeon Phi processor (derived from the company's Larrabee GPU project), codenamed Knights Ferry, was not offered as a mainstream, fully supported commercial product. It was provided in limited quantities to select developers and research partners to begin porting and optimizing code for Intel's Many Integrated Core (MIC) architecture, which later became the commercial Xeon Phi product family. 

Since Falcon Shores and its successor, Jaguar Shores, have entirely new GPU microarchitectures, Intel would typically send samples of these processors to ISVs to ensure their software can work efficiently with the hardware. Also, since Intel is focusing on rack-scale solutions, it would stand to reason that it would also supply samples of Falcon Shores to its IHV partners. However, with Falcon's new designation as an internal test chip only, that strategy doesn't appear to be planned. 

Intel dumped its Xe-HPC GPU accelerators in favor of Gaudi but is still missing its sales goals with its Gaudi 3 AI accelerator. Thus, Intel is pinning its AI hopes on its next-gen Falcon Shores GPU platform. However, interim co-CEO Michelle Johnston Holthaus said this week that rather than Falcon Shores being a first-gen breakthrough, it would require fast iteration to become a competitive platform.

"We really need to think about how we go from Gaudi to our first-generation Falcon shores, which is a GPU," said at the Barclays 22nd Annual Global Technology Conference. "I will tell you right now, is it going to be wonderful? No, but it is a good first step in getting the platform done, learning from it, understanding how all that software is going to work, how the ecosystem is going to respond, so then we can very quickly iterate after that."

Intel admitted that its Gaudi 3 platform would miss its 2024 sales targets primarily due to imperfect software. At the Barclays Conference, the company shed additional light on the situation, saying that the Gaudi platform is complex to deploy, particularly in large clusters used for training. This is why it is primarily used for inference on the edge.

"Gaudi does not allow me to get to the masses; it is not a GPU that is easily deployed in systems around the globe," said the interim co-CEO. "When you think about those that deploy Gaudi, it is from the largest hyperscalers to smaller customers that are deploying at the edge."

There are good things about Intel's Gaudi platform, too, as it enables the company to learn more about the platform and software design. While the learning from the hardware side of the business can be used for next-generation AI platforms, it remains to be seen how the lessons learned from Gaudi can be applied to entirely different next-gen platforms.

Intel's Falcon Shores is thought to be a multi-chiplet design with Xe-HPC (or at least Xe-HPC-like) and x86 chiplets with unified HBM memory. Integrating x86 CPUs and Xe-HPC GPUs into a single module with unified memory architecture will enable Intel to achieve over 5x higher compute density, memory capacity, bandwidth, and performance per watt compared to February 2022 platforms, the company said in 2022. 

Keeping in mind that Falcon shores will adopt both refined architectures and process technologies, it is reasonable to expect this unit to be dramatically faster compared to the company's 2022 products, which are Xeon Scalable 'Ice Lake' processors and the first-gen Gaudi accelerator.

Logically, Falcon Shores will be a learning vehicle for Intel and its ISV partners. The company's Data Center GPU Max 'Ponte Vecchio' has not gained significant traction in the AI realm, so independent software vendors will have to learn how to use Intel's Xe-HPC (or rather Xe-AI?) architecture on Falcon Shores. Since developing AI software takes quite some time, Intel's next-next-gen Jaguar Shores would likely be the first platform with a shot at mass adoption. 

Pat Gelsinger, the CEO of embattled tech company Intel, praised the xAI team after visiting Elon Musk’s Memphis Supercluster. He posted on X (formerly Twitter) that xAI uses Intel Xeon processors for its AI head node — the dedicated server that manages the entire 100,000-GPU-strong cluster — and that it’s “incredible what’s been built in such a short amount of time!”

After losing over $1.6 billion in its data center and foundry businesses, Intel is struggling. It has also missed the AI bandwagon, especially as its Gaudi 3 AI accelerator still suffers from issues.

On the flip side, Elon Musk has so far spent around $10 billion on AI training hardware this year, allowing his team to set up 100,000 Nvidia H200 GPUs in just 19 days — a process that usually takes four years, according to Nvidia CEO Jensen Huang. Although Nvidia powers Elon’s AI processors, he still needs a CPU to direct the AI cluster’s massive processing power. Pat Gelsinger’s post confirms that the xAI team chose Intel Xeon processors to drive their AI head node. Team Blue launched its latest 128-core flagship CPU called Granite Rapids in September 2024, but Gelsinger did not confirm which model the team uses.

The Intel CEO also praised Michael Dell, the founder and CEO of Dell Technologies and the current provider of xAI’s head node servers — at least those pictured here. Musk previously purchased systems from Supermicro, but it's unclear if the Dell servers are used in place of other Supermicro offerings — it is entirely possible that the company chose to use Dell head nodes while continuing to outfit the rest of the data center with Supermicro gear. The use of Dell head nodes will further rumors that Supermicro's legal issues have pushed xAI to switch suppliers, but that might not be the case.

Aside from exploring xAI’s massive AI cluster, we don’t know if Pat Gelsinger had other reasons for visiting the Memphis data center. Musk plans to double the GPUs on the site to 200,000 shortly, and he even mentioned plans to go as high as 300,000, although it seems that it will be a later phase of expansion. All these additional GPUs will still reportedly be Nvidia AI accelerators and could be Blackwell GPUs, so it’s unlikely that Pat is selling Elon some of Intel’s Gaudi 3 chips.

However, such a massive GPU purchase means that xAI would also need many CPUs, so he might try staying in Musk’s good graces to sell more Xeon chips. After all, even though xAI isn’t buying Intel’s AI chips yet, it still would do the company well if it could move its data center CPUs, helping it recover from its financial troubles revealed in August this year.

While the industry anticipates the launch of its next-gen dedicated Gaudi 3 ASICs followed by Falcon Shores, both pushed into 2025; Intel has already revealed Jaguar Shores as the next successor in this lineup, per HPCwire. However, we aren't sure if Jaguar Shores will combine the CPU and GPU into a single platform, as was initially planned with Falcon Shores.

, The report says that Intel's Habana Labs division - probably unintentionally - revealed this codename during a technical workshop at the SC2024 conference. When I inquired for more details, Intel refused to comment, but that's expected since most specifications are likely not finalized by now.

For the uninitiated, Intel's current lineup includes dedicated ASICs termed "Gaudi," whose third iteration - Gaudi 3 - has been pushed back to 2025 and competes against Nvidia's last-gen Hopper-based H100. Intel's HPC and AI GPUs are side by side, and Ponte Vecchio currently sits in the world's third-fastest supercomputer, Aurora.

Intel merged these two families into a disaggregated XPU (CPU+GPU) codenamed "Falcon Shores" to further streamline its AI offerings. However, those plans have been axed, and Falcon Shores will arrive in a GPU-only configuration by late 2025. With that context in mind, we cannot say much about the design choices behind the upcoming Jaguar Shores.

Jaguar Shores is likely a GPGPU (General-Purpose GPU)—akin to Nvidia's B100, B200, and B300 chips—fabbed using an Angstrom-grade node (Intel 18A/14A). HPCwire suggests that Intel teased a Falcon Shores successor last year, slated for a 2026 launch. But will it be able to compete against Nvidia's Rubin chips by then?

Admittedly, Intel's AI accelerators pale in contrast to AMD and Nvidia, and the future of AI in the company still looks bleak. The data center is the bread and butter of Intel's financials. This likely influenced its decision to reserve an X3D-esque server cache even though Intel Foundry is more than capable of manufacturing it for the mainstream market. But not all is doom and gloom since the latest Granite Rapids CPUs with upwards of 128-core signal a potential resurgence for Intel, at least in servers and workstations.

Our HDD benchmarks hierarchy shows all of the high-capacity hard drives that we've tested over the years, ranked in order of overall sequential throughput. It's a companion to our guide of the best hard drives, as well as the best SSDs and the SSD benchmarks hierarchy — we strongly recommend any modern PC use an SSD for the OS and boot drive, while HDDs are best as secondary storage.

We regularly review hard drives and update the HDD hierarchy with new models. For now, we've included 20 HDDs that have been tested and reviewed in the past couple of years. Our test PC consists of a Core i9-12900K running Windows 11 22H2, with 32GB of DDR4 memory. If you represent an HDD manufacturer and don't see a particular model in our list, drop me an email, and we can see about testing it.

We have a variety of sizes that we've tested over time, ranging from 6TB up to 22TB, with spindle speeds of 5400 RPM to 7200 RPM. Everything is grouped into a single large table for now, but as we add more models over time, we'll see about separating things according to capacity.

HDD Hierarchy

The highest-performing HDDs tend to be larger models, though that alone isn't always the determining factor. In some cases, lower-capacity drives can outperform higher-capacity drives — like the Toshiba X300 Pro 12TB and 20TB models. But while we've ranked everything by performance, pricing is arguably the more important element, and we have the best current price we're tracking listed in the table with the price per TB.

If you're after the best price per TB, the top two options are the Seagate BarraCuda 8TB and the Seagate Exos X20 20TB. The 8TB has a slightly better value at the time of writing, but the 20TB model more than doubles the raw capacity, plus it has a higher 7200 RPM rotational speed, and that's hard to pass up if you're planning to store lots of files. The worst value, looking purely at the cost per GB (TB), is the Seagate FireCuda 8TB at over $30 per TB — that's almost getting into the range of budget SSDs! Almost.

Performance can still matter, though if you're mostly interested in performance you should go with an SSD first. The Seagate Barracuda 8TB, which has the lowest price per TB, was also the slowest drive we tested. That's due in part to the 5400 RPM spindle speed, but also the use of SMR technology — shingled magnetic recording. SMR can help pack more bits per mm^2, increasing areal density, but writes in particular can be much slower as the drives fill up. Still, we'll see more SMR-type technologies in the future as we move into the 30TB and higher range.

The fastest drive overall is the Seagate Exos X20, thanks to its combination of high sequential throughput and improved random IO. Again, considering it's the second-best value overall (not factoring in performance), it's a great option. A few other drives might be slightly faster in the copy test or sequential throughput, but no other drive fully eclipses the Exos X20 — not even the newer Skyhawk AI.

Prices do change on a regular basis, of course, so it's a good idea to look around at similar capacity HDDs before making any purchase decision. A simple price cut could easily turn one of the lower ranking values into the new top option.

American lawmakers have quietly been discussing options on how they could help Intel get back on its feet should its financial situation deteriorate. According to a Semafor report, sources say this possible rescue package will be above and beyond the scope of the CHIPS Act, which would award the company at least $8.5 billion before the end of 2024.

It should be noted, though, that these are just precautionary discussions of backup plans in case the company folds. After all, the company had reported a strong outlook on its third quarterly earnings call for 2024.

These discussions show how much Washington, D.C. values Intel, primarily as the U.S. competes with China for global dominance in advanced technologies. Although both AMD and Nvidia, semiconductor giants in their rights, are also American companies, Intel is the only one that both designs and manufactures chips. “Intel is the only American company that designs and manufactures leading-edge chips and is playing a critical role in enabling a globally competitive semiconductor ecosystem in the U.S.,” said an Intel spokesperson to Semafor.

If Intel were to fail, the U.S. would have to rely on TSMC and Samsung to make its most advanced chips. Even though both of these companies already have fabs in the U.S., they only provide a fraction of their total output. Furthermore, although both companies have headquarters based in allied countries, South Korea and Taiwan are at risk because of their proximity to China.

Another reason why the Capitol and the White House wouldn’t want to see Intel fail is because it’s one of America’s top exporters, with its export revenue in 2023 exceeding $40 billion. Aside from that, the company is also working with the Pentagon’s Secure Enclave program to build leading-edge chips for the military, making it crucial for the country’s economy and security. It’s also a major employer — even though it’s in the process of laying off over 16,000 people, it still has over 120,000 employees on its payroll.

However, sources say policymakers are shying away from lump sum bailout payments as the White House did for Chrysler and General Motors in 2008. Instead, one of the solutions they’re considering is a government-encouraged private-sector merger, potentially with other Intel rivals like AMD or Marvell. After all, there have already been some rumors that Arm and Qualcomm are allegedly interested in acquiring a part or all of Intel—but some experts say that breaking up Intel will not do anyone good.

Whatever the case, let’s hope that Intel will not get into a position where it would require federal assistance to ensure its survival. Intel’s 18A chip is already showing promise, with the company saying that Amazon and two other companies are committing to using its technology for their next-generation semiconductors. If everything goes according to plan, Intel should be able to pick itself up and compete again. Still, it wouldn’t hurt the American government to have a backup plan ready if something else goes wrong.

Intel says it will now be unable to meet its goal of $500 million in Gaudi 3 sales due to software issues. Meanwhile, AMD plans to rake in $3 billion from its AI GPUs, and while Nvidia doesn't specifically state the amount it makes from AI GPUs for the data center, it is expected to be well north of $80 to $90 billion.  

Intel claims its Gaudi 3 accelerator for AI offers tangible performance improvements compared to its predecessors, and given its claimed advantages amid relatively low prices, Intel expected sales of these products to exceed half a billion dollars this year. However, the new unit was formally launched in late September, and Intel now says the software was not fully baked. Still, some Gaudi 3 accelerators will be available at IBM Cloud.

"While the Gaudi3 benchmarks have been impressive, and we are pleased by our recent collaboration IBM to deploy Gaudi 3 as a service on IBM Cloud, the overall uptake of Gaudi has been slower than we anticipated, as adoption rates were impacted by the product transition from Gaudi 2 to Gaudi 3, and software ease of use," said Pat Gelsinger, chief executive of Intel, at the company's earnings call with analysts and investors. "As a result, we will not achieve our target of $500 million in revenue for Gaudi in 2024."

Intel's Gaudi 3 relies on two interconnected chiplets housing 64 tensor processing cores, designed with a 256x256 matrix structure that uses FP32 accumulators and eight matrix engines using 256-bit wide vector capabilities. It also includes 96MB of internal SRAM cache, offering data transfer rates up to 19.2 TB/s. 

Additionally, Gaudi 3 has 24 networking interfaces running at 200 GbE and 14 media processors capable of handling video and image formats like H.265, H.264, JPEG, and VP9 for visual data processing. The chip has 128GB of HBM2E memory across eight stacks, delivering a high bandwidth of 3.67 TB/s. Compared to its predecessor, Gaudi 3 marks a substantial leap forward as Gaudi 2 contained only 24 tensor cores, two matrix engines, and 96GB of HBM2E memory.

Intel says the new Gaudi 3 accelerator offers tangible performance advantages over Gaudi 2 and can even challenge Nvidia's H100 (at least when this GPU does not use sparsity) in some cases. It is just as important that Gaudi 3 is significantly cheaper than the H100. Earlier this year, Intel disclosed that a kit featuring eight Gaudi 3 chips on a baseboard would be priced at $125,000, roughly $15,625 per chip. In comparison, a single Nvidia H100 card is currently priced at $30,678, around two times higher.

However, despite all the advantages that Gaudi 3 has, it looks like Intel's software was not exactly ready for prime time, which slowed down hardware purchases. Now, Intel expects Gaudi 3 sales to ramp up in 2025.

In 2005, Intel CEO Paul Otellini surprised the company's board. According to a report from the New York Times, he suggested that Intel buy Nvidia for "as much as" $20 billion.

According to the Times's sources ("two people familiar with the boardroom discussion"), even some Intel executives thought that Nvidia's designs could eventually play an important role in data centers. While that idea would come to fruition with the modern AI boom, the board pushed back against it. It would have been Intel's most expensive acquisition, and there were worries about integrating the company. Otellini backed off, and that was that.

Instead, Intel's board backed an internal graphics project, Larabee, which now-CEO Pat Gelsinger helmed. It used Intel's x86 technologies, and the GPU was a sort of hybrid of a CPU and GPU. Intel ultimately pulled the plug on the project, though it would later return to graphics with its Xe and Arc projects.

On the AI side, Intel has made a handful of purchases, including Nervana Systems and Movidius in 2016 and Habana Labs in 2019. But none held a candle to where Nvidia is today—a juggernaut with a market cap of over $3 trillion. Intel's Gaudi 3 AI chip is positioned as a cheaper alternative to Nvidia's offerings, but the company is primarily thought to have missed the boat on AI.

With its other struggles in manufacturing and building customers for its foundry business, Intel is now a much smaller company than Nvidia at under $100 billion. Intel is working on NPUs for consumer technology, including laptops and now its desktop CPUs.

This isn't the only time Intel gave up on getting into AI early. In 2017 and 2018, Intel had the opportunity to buy a stake in OpenAI when it was still a tiny non-profit research firm. But then-CEO Bob Swan put the kibosh on that deal, assuming that AI models were far from reaching a broad market.

Intel officially released its Lunar Lake mobile processors today, and we have a review of the Asus Zenbook S14 with the Core Ultra 7 258V. This is supposed to be the new king of laptop processors, with improved battery life being a key element. But it's also the showcase for Intel's latest GPU architecture, and while no integrated graphics solution will rank among the best graphics cards and these aren't positioned as gaming laptops, we want to see how Intel's latest stacks up against the competition.

The Core Ultra 9 288V comes with Intel Arc Graphics 140V, with the V indicating the power level of the chip. It's a nominal 30W part with a maximum turbo power of 37W. We're using the Core Ultra 9 in order to potentially remove power restrictions that might impact the Core Ultra 7 258V, with the hope of seeing higher, more consistent performance. This is also the debut of Intel's Battlemage GPU architecture, sort of — power limits and shared memory will be major limiting factors for any GPU workload.

To see what Lunar Lake and Battlemage bring to the table, we'll compare the new processor with the previous generation Meteor Lake chip along with AMD's Ryzen AI chip. Here's the quick rundown of core specs for the various chips we're looking at.

[Note: We have retested the games that used FSR3 upscaling and framegen to eliminate those technologies from the equation, as it seems framegen may have failed to work in at least one instance. We also tested with XeSS alongside FSR3 upscaling on both AMD and Intel GPUs, without framegen.]

What should be clear from the above table is that the three laptops we have for testing are not equivalent in a variety of ways. There are slightly different power targets, different CPU core counts, different tiers (Core Ultra 7/9 and Ryzen 9), different memory, and other places where they don't match up. There are also wildly different NPUs in play for AI workloads and even wildly different AI capabilities for the GPUs.

But I'm mostly interested in the graphics performance, so that will be the focus. And even here, we have some interesting changes relative to the previous generation Meteor Lake GPU. Both have eight Xe-cores (roughly similar to an AMD Compute Unit, aka CU), but where Meteor Lake's Arc Xe GPU omitted the XMX functionality, Lunar Lake's Arc Xe2 includes that. It should allow for better XeSS upscaling quality and performance; otherwise, we don't expect that to make much of a difference in graphics performance.

The Lunar Lake GPU also has a peak clock speed of just 2.05 GHz, down 200 MHz from Meteor Lake's 2.25 GHz and way, way below AMD's potential 2.9 GHz. AMD also has twice as many graphics clusters (CUs) as Intel, but half as many shaders per CU, giving peak theoretical compute of 5.94 TFLOPS. That's 41% more than the 4.20 TFLOPS on Lunar Lake and 29% more than Meteor Lake's 4.61 TFLOPS. Theoretical TFLOPS rarely tell the whole story, so it will be interesting to see how the new Arc Graphics 140V stacks up against the previous Arc Graphics; Intel says the new chips should be universally faster.

All three laptops support DDR5 (or LPDDR5) memory, but the actual memory used is often up to the laptop vendors — though not on Lunar Lake. We're using Asus Zenbook models in all three cases, and both the Meteor Lake and Strix Point models have DDR5 running at 7500 MT/s (the MTL says 7467 MT/s, but close enough). With dual-channel memory that works out to 120 GB/s of raw memory bandwidth — a far cry from what you get with even a modest dedicated GPU, but it should be reasonably sufficient for a laptop with integrated graphics. The new Lunar Lake laptop has 8533 MT/s LPDDR5X memory integrated onto the package, yielding a moderately higher 136.5 GB/s of bandwidth.

As I said, we're looking at apples and oranges in some respects, as there are many differences in the core specifications. Still, it's what we have available, so let's go ahead and run some gaming and graphics benchmarks to see where things land.

We have three laptops, all from Asus, and the MyAsus app also allows you to set a fan speed profile. The default is "Standard," but we've found that selecting "Performance" seems to increase the TDP limit and can help quite a bit with gaming performance. We initially tested each laptop with both the Performance and Standard fan modes (flip through the galleries to see the framegen results), but we'll stick with just Performance mode for the additional non-framegen testing. Windows 11's power profile was set to "Balanced" mode for all the testing shown here.

First up is 3DMark Time Spy. I don't like using synthetic benchmarks for normal GPU testing, but in this case it's a reasonable way to try and remove driver issues from the equation. Every GPU manufacturer should be intimately familiar with 3DMark, and ensuring the drivers work optimally is standard procedure. If the drivers are fully tuned for any games that we test, we'd expect the results to at least echo what Time Spy shows.

And what it shows is that, despite having theoretically less GPU compute available, Lunar Lake's Battlemage GPU comes out on top. It's not massively faster than the other chips, but it shows 13% higher performance than the Radeon 890M when both are using the MyAsus Performance mode fan profile. Compared to the previous generation MTL GPU, it's also 14% faster. Putting all three laptops into the Standard mode fan profile does change things a bit. All three laptops run slower, and the Lunar Lake GPU is now only 5% faster than the 890M but still 12% ahead of the Meteor Lake GPU.

One critical item to point out is that memory bandwidth on Lunar Lake is 13–14 percent higher than on the other two laptops, so instead of showing true gaming potential, 3DMark Time Spy might actually end up scaling more with bandwidth. Some games do behave that way, but plenty of others depend more on the GPU computational abilities rather than bandwidth. So let's check out some games...

Black Myth Wukong is a recent release, which makes it a good showcase for any potential driver concerns. We're using Intel's latest preview drivers for the Lunar Lake system, AMD's 24.8.1 drivers on the Ryzen AI / Strix Point laptop, and the latest publicly available Intel 6077 drivers for Meteor Lake. (6078/5736 drivers just came out today for the retail Lunar Lake launch, but we used 6077 from late last week.)

Starting with the native 720p results, the standings don't look the same as 3DMark. The new Lunar Lake GPU still comes out 11% ahead of the Meteor Lake GPU, but now AMD's 890M ends up 10% ahead of Lunar Lake. This is the issue with synthetic graphics tests like 3DMark versus real-world gaming performance. We'll want to look into testing more games to come to any firm conclusion.

Bumping the resolution to 1080p and enabling quality mode upscaling in both XeSS and FSR3, AMD still comes out on top, with slightly higher XeSS performance compared to FSR3 performance. However, XeSS 1.3 uses larger upscaling factors so that's likely the cause. The two Intel GPUs also show better performance with XeSS than FSR3, though the biggest gap remains less than 10%.

Last in the gallery are the FSR3 results with frame generation enabled, which we retested as there were some anomalies. Things take a turn for the worse on Intel's new GPU here, as framegen does something for the Arc 140V, but it's definitely not working that well. Lunar Lake sees a 17% increase in fps — and half of the framegen frames have no user input so it will feel noticeably worse. That's in stark contrast to Meteor Lake, which improved by 50%, and AMD's 890M that achieved a 63% higher result than the non-framegen 32 fps.

Cyberpunk 2077 has been around for a while, but it just (finally) received the promised FSR3 support, so we'll see how that goes. But let's start with the native 720p result, which removes any upscaling or framegen shenanigans from the equation.

Of our tests, this represents the best results from Intel's new Arc Graphics 140V. It's 32% faster than the previous Arc Graphics solution found in Meteor Lake, and since both are running in basically the same Asus 14-inch chassis, these are directly comparable. AMD's chip still beats Lunar Lake by 18%, but that's at least partially influenced by the larger 16-inch chassis that may have a higher TDP with our test settings.

Bumping to 1080p and enabling upscaling, again noting that XeSS 1.3 uses a larger scaling factor than FSR3 (and DLSS), the Arc 140V takes the top spot with 51 fps, narrowly edging past the GPU chip. AMD's performance is basically the same with XeSS as with FSR3, while Intel (not surprisingly) sees a bigger boost with XeSS. That applies to both Meteor Lake and Lunar Lake, with LNL getting a 6% increase while MTL gets a 9% boost.

Turning on FSR3 framegen completely changes the picture, however, with the Radeon 890M dominating the competition. It's 47% faster than MTL in performance mode and 27% faster in standard mode — so in this case, engaging the performance mode fan profile in the MyAsus app boosts performance by over 40%. Again, chassis size looks to be a factor, with the MTL laptop showing a 24% boost in performance while the newer LNL laptop gets a 14% boost.

As noted with Black Myth Wukong, Lunar Lake doesn't seem to appreciate FSR3 framegen much right now. It saw basically no benefit at all. AMD's GPU meanwhile got a 60% boost to fps in performance mode, and Meteor Lake saw a 54% increase. It's odd that framegen, which just uses GPU shader computation as far as we're aware, doesn't seem to want to run on Lunar Lake right now.

Last, we have Shadow of the Tomb Raider, an older game that should be good and fully optimized for any reasonable drivers. And, interestingly, we get a chart that looks more like 3DMark. There's no framegen this time, and only XeSS (or DLSS if you have an Nvidia GPU), so we opted to include native 1080p performance.

At 720p, the new Lunar Lake GPU claims the top spot with 74 fps, leading AMD's 890M by 6% and beating the Meteor Lake GPU by 25%. That's a good result for what is still likely a power constrained GPU — and having fewer CPU cores probably helps Lunar Lake's GPU make the most of the horsepower it has on tap.

Our second chart is with XeSS Quality mode enabled. Now, I think this is XeSS 1.2 or 1.1, not the latest 1.3, so the scaling factor is 67% — 2.25X upscaling. And the DP4a mode used on the Meteor Lake and AMD GPUs doesn't tend to look as good as the XMX mode running on the Arc 140V. AMD and Meteor Lake end up basically tied, while the new Arc 140V holds a 25% lead.

Finally, we have the native 1080p medium results. Overall, the Core Ultra 9 288V takes top honors with 51 fps in performance mode and 40 fps in standard mode. That's also the biggest improvement from engaging the higher speed fan mode, a 24% increase. The Core Ultra 7 155H only increased by 14%, while the Ryzen AI 9 HX 370 saw a 15% increase. The net result is that Lunar Lake comes out 13% ahead of Strix Point with both performance modes but only 5% ahead in standard mode. It's also 21% faster than Meteor Lake in performance mode and 11% faster in standard mode.

First Impressions of Lunar Lake's Arc Xe2 Graphics

With updated testing results that don't use FSR3 frame generation in Black Myth Wukong and Cyberpunk 2077, our overall impression of Lunar Lake's GPU improves quite a bit. It's not a massive generational performance improvement compared to Meteor Lake, but the 28–30 watt TDP of the two chips we tested probably limits how far Intel can push performance regardless — and memory bandwidth also makes a difference.

There are, however, still some oddities present, things which will probably end up being fixed in Intel's drivers. Why didn't framegen work at all on Lunar Lake, or work very poorly, when it seems to do just fine on Meteor Lake? It doesn't make sense for Intel to intentionally break AMD's tech, but it's also weird that it doesn't seem to work in general. We've tested FSR3 framegen on old Nvidia GTX cards and it tends to work fine, so why would Lunar Lake be any different?

Not shown in the charts, we also ran the same graphics tests on the 'lesser' Core Ultra 7 258V. The main differences between that chip and the Core Ultra 9 288V are the base power (17W vs 30W) and maximum clocks (100 MHz higher for the GPU clock on 288V). In our testing, at least in the standard fan mode, the two chips are within spitting distance of each other on our graphics tests. So, you can safely opt for the slower and less expensive Lunar Lake chip and only give up a very small amount of performance — which makes sense. Lunar Lake doesn't have lots of CPU cores even on the Core Ultra 9 288V.

Overall, the graphics performance from Lunar Lake and the Arc Graphics 140V, aka Arc Xe2 Battlemage, doesn't look particularly amazing, but we didn't expect to see much since we're dealing with integrated graphics. Power and memory bandwidth constraints often eclipse other factors. We've seen claims of double the performance (or at least performance per watt) on various integrated GPU solutions in the past, and those rarely pan out in the real world. With its current focus on ultraportable, low voltage, long battery life laptops, this is hardly the best showcase for Intel's next-generation Battlemage GPU architecture.

We hope that the results for the future desktop cards end up being far more impressive once Battlemage GPUs begin shipping. We also hope that driver issues and oddities continue to get resolved. Intel has improved in that area since the initial Arc launch, but AMD and Nvidia remain far more consistent in both game support and performance. There are times when Lunar Lake might come out ahead of AMD's Radeon 890M, but we suspect that over a larger test suite, Intel will encounter far more frequent anomalies in terms of game support and optimizations.

Update 10/1/2024: Added more information on Intel's Guadi 3 and corrected supported data formats.

Intel formally introduced its Gaudi 3 accelerator for AI workloads today. The new processors are slower than Nvidia's popular H100 and H200 GPUs for AI and HPC, so Intel is betting the success of its Gaudi 3 on its lower price and lower total cost of ownership (TCO).

Intel's Gaudi 3 processor uses two chiplets that pack 64 tensor processor cores (TPCs, 256x256 MAC structure with FP32 accumulators), eight matrix multiplication engines (MMEs, 256-bit wide vector processor), and 96MB of on-die SRAM cache with a 19.2 TB/s bandwidth. Also, Gaudi 3 integrates 24 200 GbE networking interfaces and 14 media engines — with the latter capable of handling H.265, H.264, JPEG, and VP9 to support vision processing. The processor is accompanied by 128GB of HBM2E memory in eight memory stacks, offering a massive bandwidth of 3.67 TB/s.

Intel's Gaudi 3 represents a massive improvement when compared to Gaudi 2, which has 24 TPCs and two MMEs and carries 96GB of HBM2E memory. Intel did not simplify its TPCs and MMEs as the Gaudi 3 processor supports FP8, BF16, FP16, TF32, and FP32 matrix operations, as well as FP8, BF16, FP16, and FP32 vector operations.

*with sparsity. 

When it comes to performance, Intel says that Gaudi 3 can offer up to 1,835 BF16/FP8 matrix TFLOPS as well as up to 28.7 BF16 vector TFLOPS at around 600W TDP. Compared to Nvidia's H100, at least on paper, Gaudi 3 offers higher BF16 matrix performance or just slightly lower when Nvidia's hardware uses sparsity feature (1,835 vs 1,979 TFLOPS), slightly lower FP8 matrix performance or two times lower FP8 matrix performance when H100 uses sparsity (1,835 vs 3,958 TFLOPS), and significantly lower BF16 vector performance (28.7 vs 133.8 TFLOPS). 

More important than the raw specs will be the actual real-world performance of Gaudi 3. It needs to compete with AMD's Instinct MI300-series and Nvidia's H100 and B100/B200 processors. This remains to be seen, as a lot depends on software and other factors. For now, Intel showed some slides claiming that Gaudi 3 can offer a significant price-performance advantage compared to Nvidia's H100.

Earlier this year, Intel indicated that an accelerator kit based on eight Gaudi 3 processors on a baseboard will cost $125,000, which means that each one will cost around $15,625. By contrast, an Nvidia H100 card is currently available for $30,678, so Intel plans to have a big price advantage over its competitor. Yet, with the potentially massive performance advantages offered by Blackwell-based B100/B200 GPUs, it remains to be seen whether the blue company will be able to maintain its advantage over its rival.  

"Demand for AI is leading to a massive transformation in the datacenter, and the industry is asking for choice in hardware, software, and developer tools," said Justin Hotard, Intel executive vice president and general manager of the Data Center and Artificial Intelligence Group. "With our launch of Xeon 6 with P-cores and Gaudi 3 AI accelerators, Intel is enabling an open ecosystem that allows our customers to implement all of their workloads with greater performance, efficiency, and security."

Intel's Gaudi 3 AI accelerators will be available from IBM Cloud and Intel Tiber Developer Cloud. Also, systems based on Intel's Xeon 6 and Gaudi 3 will be generally available from Dell, HPE, and Supermicro in the fourth quarter, with systems from Dell and Supermicro shipping in October and machines from Supermicro shipping in December.

Unlike rivals like Nvidia and AMD, Intel hasn't made a massive splash in artificial intelligence products aside from some laptop chips with NPUs. According to a report from Reuters, things could have gone differently. In 2017 and 2018, Intel had an opportunity to buy a stake in OpenAI, which was then a tiny non-profit research firm.

The report, citing four people with knowledge of the situation, suggests that Intel and OpenAI discussed several different investment options, including purchasing a 15% stake in the company for $1 billion or taking a stake in OpenAI while selling hardware to the AI company "at cost price."

Former Intel CEO Bob Swan reportedly scuttled the deal because had didn't see AI models hitting the broad market anytime soon, which would mean that Intel wouldn't be repaid on its investment. OpenAI launched ChatGPT to massive acclaim in 2022. Additionally, the sources said that Intel's data center unit "did not want to make products at cost.

On OpenAI's side, teaming up with Intel would mean it wouldn't have been as dependent on Nvidia. Nvidia has since become the biggest name in AI hardware with its GPUs.

Microsoft poured billions into OpenAI in 2019, 2021, and 2023, turning the company into a big player on the AI stage. AMD, too, has become an AI company, though perhaps not on Nvidia's scale. AMD's data center business has boomed with the AMD Instinct MI300.

In its recent earnings, Intel's data center and AI Group (DCAI) reported reduced revenue and operating income. Still, its Xeon 6 'Sierra Forest' processor is in volume production, and Intel is set to start shipments for Xeon 6 'Granite Rapids' CPU. The company's Gaudi 3 AI accelerators are projected to ramp up in the second half of the year. Intel has suggested it has "over 20 customers" set to use Gaudi 3.

Neither Intel nor OpenAI responded to requests for comment before publication.

As CEO of Intel from 2018 (interim CEO) to 2021 (when he was replaced in the full-time gig by Pat Gelsinger), Swan often faced criticism due to a lack of technical acumen. He has previously served as the company's CFO, focusing more on the books than the tech. In retrospect, this missed opportunity to get in on AI earlier may further cement Swan's legacy.

Swan's predecessor, Brian Krzanich, bought Nervana Systems 2016 to compete with Google's tensor processing units. IN 2019, Intel purchased another AI startup, Habana Labs, and prioritized it over Nervana, shuttering the latter. Intel also owns Movidius, which it bought in 2016 for its vision processing units.

Etched, a startup that builds transformer-focused chips, just announced Sohu, an application-specific integrated circuit (ASIC) that claims to beat Nvidia’s H100 in terms of AI LLM inference. A single 8xSohu server is said to equal the performance of 160 H100 GPUs, meaning data processing centers can save both on initial and operational costs if the Sohu meets expectations.

According to the company, current AI accelerators, whether CPUs or GPUs, are designed to work with different AI architectures. These differing frameworks and designs mean hardware must be able to support various models, like convolution neural networks, long short-term memory networks, state space models, and so on. Because these models are tuned to different architectures, most current AI chips allocate a large portion of their computing power to programmability.

Most large language models (LLMs) use matrix multiplication for the majority of their compute tasks and Etched estimated that Nvidia’s H100 GPUs only use 3.3% percent of their transistors for this key task. This means that the remaining 96.7% silicon is used for other tasks, which are still essential for general-purpose AI chips.

However, the transformer AI architecture has become very popular as of late. For example, ChatGPT, arguably the most popular LLM today, is based on a transformer model. In fact, it’s in the name — Chat generative pre-trained transformer (GPT). Other competing models like Sora, Gemini, Stable Diffusion, and DALL-E are all also based on transformer models.

Etched made a huge bet on transformers a couple of years ago when it started the Sohu project. This chip bakes in the transformer architecture into the hardware, thus allowing it to allocate more transistors to AI compute. We can liken this with processors and graphics cards — let’s say current AI chips are CPUs, which can do many different things, and then the transformer model is like the graphics demands of a game title. Sure, the CPU can still process these graphics demands, but it won’t do it as fast or as efficiently as a GPU. A GPU that’s specialized in processing visuals will make graphics rendering faster and more efficient, that’s because its hardware is specifically designed for that.

This is what Etched did with Sohu. Instead of making a chip that can accommodate every single AI architecture, it built one that only works with transformer models. When it started the project in 2022, ChatGPT didn’t even exist. But then it exploded in popularity in 2023, and the company’s gamble now looks like it is about to pay off — big time.

Nvidia is currently one of the most valuable companies in the world, posting record revenues ever since the demand for AI GPUs surged. It even shipped 3.76M data center GPUs in 2023, and this is trending to grow more this year. But Sohu’s launch could threaten Nvidia’s leadership in the AI space, especially if companies that exclusively use transformer models move to Sohu. After all, efficiency is the key to winning the AI race, and anyone who can run these models on the fastest, most affordable hardware will take the lead.

Ever since AI data centers started popping up left and right, many experts have raised their concerns over the power consumption crisis this power-hungry infrastructure will lead us to. Meta founder Mark Zuckerberg says electricity supply will constrain AI growth, and even the U.S. government has stepped in to discuss AI power demands. All the GPUs sold last year consume more power than 1.3 million homes, but if Etched’s approach to AI computing with Sohu takes off, we can perhaps reduce AI power demands to more manageable levels, allowing the electricity grid to catch up as our computing needs grow more sustainably.

When Intel announced its Gaudi 2 and Gaudi 3 accelerators for AI, it said that the total cost of ownership would be a fraction of that of Nvidia's H100, but never revealed any actual numbers. At the Intel's keynote at Computex 2024 the company finally revealed pricing of its latest AI processors and it appears that eight Gaudi 2 processors on a baseboard will cost less than one Nvidia B200 is rumored to cost.

Intel charges $65,000 per eight Gaudi 2 accelerators on a baseboard and $125,000 per eight Gaudi 3 processors on a baseboard, as shown by our colleague Jim McGregor from Tirias Research. This means that a single Gaudi 2 is priced at around $8,125, whereas a Gaudi 3 costs around $15,650 — of course, when purchased in bulk and with a baseboard. Yet, for large customers buying in huge quantities Intel can certainly make further per-unit discounts.

By contrast, Nvidia's H100 80GB cards cost $30,000 — and more when purchased retail, though these cards offer lower performance than H100 80GB SXM modules. HSBC projects that Nvidia's 'entry-level' next-generation B100 GPU based on the Blackwell architecture will have an average selling price (ASP) ranging from $30,000 to $35,000, which is comparable to the price of Nvidia's H100. The more powerful GB200, which integrates a single Grace CPU with two B200 GPUs, is expected to be priced between $60,000 and $70,000. However, it's important to note that these are just analyst estimates, and that the actual cost could be significantly higher.

Selling Gaudi accelerators for AI at a discount is not unexpected. Intel is currently an underdog in a market dominated by Nvidia, which benefits from the fact that many AI workloads are tailored for its CUDA platform in general and the Hopper architecture in particular. Intel has to fight not only against Nvidia, but against custom in-house-designed AI processors used by cloud service providers like AWS and Google. These accelerators come without a markup and for Intel it is particularly hard to compete against them. In fact, competing against custom silicon could be even harder for Intel because it needs to earn selling its products, while companies like AWS do not.

Intel's codenamed Falcon Shores hybrid processor combines x86 and Xe GPU cores to offer formidable performance for AI and HPC workloads, but it will consume an extreme 1500W of power, reports ComputerBase.de, citing a comment by Intel. Such an extreme power consumption will require Intel to use advanced cooling methods. 

Intel's Falcon Shores will be a multi-tile processor featuring both x86 cores (tiles) for general-purpose processing and Xe cores (tiles) for highly parallel AI and HPC workloads. Intel itself once said that it would offer five times higher performance per watt and five times higher memory capacity and bandwidth compared to its 2022 products while also offering a 'simplified' programming model.  

The company still hasn't revealed detailed performance expectations for its Falcon Shores processors. To feed Falcon Shores, Intel will probably have to use proprietary modules (or promote a new OAM specification) as even the latest OAM specification (OAM 2.0), which features a new high-power connector, can only support power levels of around 1000W. Even the power consumption of Nvidia's B200 won't exceed 1,200W. 

Cooling a 1,500W processor is another matter. Some of Intel's partners may use liquid cooling, but others could probably opt for liquid immersion cooling, a technology that Intel has been promoting for several years now. 

Intel seems to be pinning a lot of hope on its Falcon Shores processor. Although Intel's Gaudi 3 is significantly more powerful than its predecessors, and Intel expects the industry to adopt this accelerator for various uses, the company seems somewhat cautious about the product's success. This is perhaps because its expectations for its Falcon Shores processor are considerably higher. 

Considering that by the time Falcon Shores arrives in 2025, there will be more software developers familiar with Xe architecture for supercomputers, the adoption of Falcon Shores will likely be relatively smooth for a new product. 

Intel has officially started the process of sunsetting its Ponte Vecchio HPC GPUs. According to ServeTheHome, Intel is starting to retire these GPUs just two years after initial release, as the company looks forward to launching Ponte Vecchio's successor Falcon Shores, as well as the Gaudi 2 and Gaudi 3 AI accelerators.

Intel isn't fully retiring Ponte Vecchio just yet, but it is no longer prioritizing Ponte Vecchio sales. To be more specific, Intel will still be producing Ponte Vecchio GPUs, but only for existing customers. This means that new customers will be out of luck if they want to buy Intel's latest (at the time of writing) GPU hardware and will be forced to look at alternatives from AMD and Nvidia.

Intel's decision to begin sunsetting Ponte Vecchio makes sense — the HPC and AI market is incredibly competitive right now, with the current AI boom. Intel Ponte Vecchio is now effectively a previous-generation architecture, thanks to AMD's introduction of the much newer MI300 and Nvidia's introduction of the B200 Blackwell. As a result, there's not much incentive for customers to buy Intel's inferior current GPU hardware when they can get newer, better hardware from competitors.

Ponte Vecchio originally launched in 2022 and it's the biggest GPU Intel has produced, so far. The architecture itself sports over 100 billion transistors, spread across 47 active tiles, and manufactured on five different process nodes. Depending on the configuration, one Ponte Vecchio GPU could have as many as eight compute tiles, or GPU dies, and four HBMe2 stacks based on Intel's Xe HPC architecture — which is even more potent compared to the vanilla Xe architecture Intel is using in its consumer Arc Alchemist GPUs.

It's also worth noting that the biggest showcase for Ponte Vecchio hasn't blown anyone away: The Aurora Supercomputer underperforms and uses more power than AMD's Frontier supercomputer. According to the latest Top500 numbers, Aurora takes the number two spot with 1,012 petaflops and 38,698 kW of power, compared to the older Frontier's 1,206 petaflops and 22,786 kW. There are workloads (mixed precision) where Aurora does better, but it was supposed to hit 2 exaflops (2,000 petaflops) — so it's short by half.

Despite its impressive size, Ponte Vecchio is now outdated compared to AMD's newer MI300 and Nvidia's new B200 APUs/GPUs. Intel, as usual, is already super late in delivering its MI300/B200 competitor in the form of Falcon Shores. This new architecture was supposed to debut in 2024 with a mixture of CPU and GPU cores to compete with AMD's MI300 APU (which also boasts CPU and GPU cores) and Nvidia's Grace Hopper Superchip, but Intel has been forced to delay Falcon Shores to 2025 and limit it to a GPU-only style chip.

Beginning the retirement process of Ponte Vecchio before Falcon Shores arrives will give Intel more resources to focus on Falcon Shores and accelerating its production after debut. This should help Intel catch up to AMD and Nvidia, but it will still be behind the development curve since AMD and Nvidia are also underway, building next-generation HPC hardware.

The August 1982 release of the Commodore 64 is historic, as Commodore's hit personal computer managed to be one of the best-selling PCs of all time — and, it turns out, this historic era of Commodore 64 computing has produced hardware that can do AI image generation — with caveats, of course.

Developer and hobbyist Nick Bild has successfully built and documented a Generative AI tool for Commodore 64 that can be used to create 8x8 sprites that are then displayed at 64x64 resolution. These are intended to help inspire game design concepts, but certainly aren't up to the level of generating entire sprite sheets off one prompt. There are much higher-end implementations of AI in existing games, as well.

In any case, it's fascinating that any kind of generative AI model can be run on hardware this old. It still takes twenty minutes to run 90 iterations for a final image, but that's not bad at all considering the age of the hardware. This also recalls a story from mid-April where the Commodore 64 managed to outperform a modern IBM QPU (Quantum Processing Unit) in a quantum utility experiment.

No reliance on something like OpenAI is needed, though the "probabilistic PCA algorithm" running on the Commodore 64 used for this project was actually trained on a modern computer. So while the model runs on Commodore 64 as advertised, a modern PC was still needed to get this up and running to begin with.

It seems that while the entry-level for "real" AI PCs is hotly up for debate by manufacturers, the ever-reliable Commodore 64 reminds us that the true entry-level starts wherever you, the end user, want it to. With skill, determination, and patience almost anything is possible, though of course, practicality is another question entirely. Even other Commodore 64 mods, like the Raspberry Pi C64 expansion cartridge playing Doom, may be more practical for end users than 8x8 pixel AI sprite generation running on 40-year-old PC hardware.

AMD this week increased its datacenter GPU sales guidance by $500 million to $4 billion as the number of customers interested in its latest Instinct MI300-series processors is increasing. $4 billion sales of GPUs for AI and HPC applications is a lot of money in general, but it is a tiny fraction of what Nvidia will earn on its H100, H200, and B200 products. 

"[Instinct] MI300 demand continues to strengthen," said Lisa Su, chief executive of AMD, at the company's earnings call with financial analysts and investors (via SeekingAlpha). "Based on our expanding customer engagements, we netback data center GPU revenue to exceed $4 billion in 2024, up from the $3.5 billion we guided in January."  

AMD began to ship its Instinct MI300-series products in Q4 2023 (the first customers reported that they had received the compute GPUs in mid-January 2024). Its sales have surpassed $1 billion, according to AMD, which is in line with the company's expectations that the Instinct MI300-series would be its fastest product to $1 billion revenue. It is highly likely that AMD begain to ship its Instinct MI300 products in very late Q4 2023, so the majority of shipments were made in the first quarter of this year. Such a high adoption rate points to significant interest in the new series and AMD's conservative guidance about Instinct MI300's sales in 2024.

Meanwhile, it should be noted that AMD is not supply constrained with the Instinct MI300 and demand for the product is relatively modest at this point. 

"Our $4 billion number is not supply capped, […] we do have supply capability above that," said Su. "It is more back half weighted. So, if you are looking at sort of the near-term, I would say, for example, in the second quarter, we do have more demand than we have supply right now, and we are continuing to work on pulling in some of that supply."

Last week Intel guided that sales of its Gaudi 3 products will achieve $500 million this year, which is not much compared to how much its Xeon processors for datacenters earn. Gaudi 3 has yet to ramp, so it is possible that the processor will not have enough time to sell in huge quantities in 2024. 

Expectations for sales of AI processors by AMD and Intel pale when compared to what Nvidia is expected to make in 2024. Analysts believe that the company could sell $40 billion worth of datacenter GPUs.

It is noteworthy that while AMD's datacenter GPU sales look impressive, they are still considerably lower compared to sales of AMD's gaming solutions, comprising of AMD's system-on-chips for Microsoft's Xbox and Sony's PlayStation consoles, the company's discrete Radeon GPUs for desktops and laptops, and not including built-in Radeon graphics. Sales of AMD's gaming products exceeded a billion dollars per quarter for years and only dropped to $922 million this quarter.

More than 500 AI models now run optimized on new Intel Core Ultra processors, the chipmaker announced on May 1, 2024. Intel’s promotion of artificial intelligence research and development across various aspects of the technology made the milestone possible.

Intel’s contributions to the AI development community have made the Core Ultra processor the fastest-growing AI processor to date, the company said. In October 2023, Intel launched an AI Developer Program.

Then, the company added even more programs in 2024" to promote building AI applications and hardware. Part of that initiative included providing hands-on lab time to developers with Intel’s new Core Ultra Meteor Lake NUC development kit.

The sheer number of AI models supported by Intel’s Core Ultra matters because they form "the backbone of AI-enhanced software features like object removal, image super resolution or text summarization," the chipmaker said in its press release. More optimized models means more AI features devs can bring to users.

Intel sold more AI PC processors in a single quarter than its competitors did throughout all of 2023, the company noted.

"Intel has a rich history of working with the ecosystem to bring AI applications to client devices, and today we celebrate another strong chapter in the heritage of client AI by surpassing 500 pre-trained AI models running optimized on Intel Core Ultra processors," said Robert Hallock, vice president of AI at Intel. "This unmatched selection reflects our commitment to building not only the PC industry’s most robust toolchain for AI developers, but a rock-solid foundation AI software users can implicitly trust."

Intel’s next-gen laptop chip, Lunar Lake, is expected to spur even more growth. The CPU provides 100+ tera-operations per second (TOPS) of AI workload performance, three times what the current generation can handle. TOPS is used specifically to measure how many computing operations an AI accelerator can complete in one second at full utilization.

During its AI Summit in Taipei, Intel set the bar for next-gen AI PCs at 45 TOPS of Neural Processing Unit (NPU) performance, which Lunar Lake provides. By way of comparison, Intel’s Meteor Lake chips provide more than 65 TOPS of performance, but only 10 TOPS of that performance come from the NPU. AMD, on the other hand, lags but says it will catch up. Its current-gen Ryzen Hawk Point platform has an NPU with 16 TOPS of performance. However, the company says its next generation will triple that performance. Qualcomm will launch its Snapdragon X lineup this year, which will offer 45 TOPS on the NPU.

The 500 AI models running on Intel Core Ultra processors come from more families than just the ChatGPT-like services most folks are familiar with. They include large language models (ChatGPT, Microsoft Copilot, Google Gemini, etc.) as well as computer vision, diffusion, image classification/segmentation, object detection, super resolution, and more. Some of the industry sources Intel pointed out as using its technology include OpenVINO Model Zoo, Hugging Face, ONNX Model Zoo, and PyTorch. These sources span more than 20 categories of AI, encompassing just about anything you might want to do with the technology.

Intel's Gaudi 3 processor for AI applications is significantly faster than its predecessor, which makes this offering significantly more competitive against rivals. However, the company gave a rather uninspiring guidance for Gaudi 3 sales this year: the company expects Gaudi 3 to generate $500 million in revenue in 2024. Compared to AI-related expectations of AMD and Nvidia, this looks like a negligible sum.

"Our Gaudi 3 launch gave us a strong offering to improve our position in accelerated computing for the data center and cloud," said Pat Gelsinger, Intel's chief executive, at the earnings call last week. "We now expect over $500 million in accelerated revenue in the second half of 2024 with increasing momentum into 2025 based on Gaudi 3's vastly superior TCO as well as our own expanding supply."

AMD expects its Instinct MI300-series offerings to generate about $3.5 billion in revenue this year, which is a modest expectation compared to Nvidia, whose data center GPU revenue — dominated by GPUs for AI and HPC applications — is expected to hit around $40 billion this year. In fact, keeping in mind that Nvidia is now focusing on shipping machines based on its Blackwell compute GPUs rather than individual processors, it is likely that Nvidia's data center revenue will be significantly more than $47 billion in 2023.

There could be several reasons why Intel has such humble expectations for Gaudi 3's business performance in 2024. Perhaps the company will start shipping Gaudi 3 in meaningful volumes only very late in the year and will not have enough time to ship tens of thousands of processors for revenue in 2024. It is also possible that Intel will be supply-constrained or that it expects sales of Gaudi 3 to take off slowly because customers yet have to tailor their AI software for Intel's hardware.

Based on Intel's claims, the interest in Gaudi 3 is significant, so the product should be quite popular, which means that over time, it will generate much more than $0.5 billion for the company.

"At our Vision event, we had over 20 customers publicly describing their embrace of Gaudi 2 and Gaudi 3," said David Zinsner, Intel's chief financial officer, at the conference call. "I was super pleased to see the breadth of those customers. It was CSPs like Naver and Ola and IBM Cloud. It was ISVs like Seekr, right, coming on board. But maybe most importantly, enterprise customers."

An engineer has shared his experience of designing a CPU from scratch, over two weeks, “with no prior experience.” During this brief period, Adam Majmudar claims to have learned the fundamentals of chip architecture, absorbed the finer points of chip fabrication, and prepared his first full chip layout using EDA tools. The next step in his “speed-running the chip stack” to-do list is designing a GPU from scratch. When finished this project is destined for production via Matthew Venn’s TinyTapeout 6.

We’ve reported on enthusiast DIY CPU designs previously, as well as DIY GPU projects. However, some of those feats have eaten through years of spare time for the people involved. Majmudar must be on vacation and spending all his surplus time on this “speed run” project to have gotten as far as he has “from scratch.”

The fledgling chip designer, who describes himself as one of the founding engineers at a web3 development company, outlines the steps he has made so far in his quest. You can click and read through all the steps leading up to the current GPU focus via the embedded Tweet, above. We’ve also bullet-pointed the speed run steps which have been completed to date, below.

• Learning the fundamentals of chip architecture – a strong understanding is a critical foundation
• Learning the fundamentals of chip fabrication – materials, wafer prep, patterning and packaging
• Starting electronic design automation by making a CMOS transistor, layer-by-layer
• Creating my first full circuit in Verilog – “my first experience with programming hardware using software.”
• Implementing simulation & formal verification for my circuit
• Designing my first full chip layout – designing and optimizing using OpenLane, which is an open-source EDA tool

As we mentioned in the intro, the significant step that Majmudar now faces is designing a GPU from scratch. He knows this will be a difficult task and admits that, after initial investigations, it is harder than expected. The fledgling chip designer explains that there simply aren’t the learning resources online for building a GPU. “Because GPU companies are all trying to keep their secrets from each other, most of the GPU architecture data is all proprietary and closed source,” the engineer finds.

Despite this hurdle, Majmudar says the big GPU-makers’ secrecy has made this part of the project “way more fun for me.” Interestingly, Anthropic’s Claude Opus AI tools have been useful during this GPU designing stage. “I've been proposing my ideas for how each unit must work to Claude, and then somehow it will guide me toward the right implementation approaches which I can then go and confirm with open-source repos,” explained the engineer. However, he observed that “if I search some of the things publicly, nothing shows up which is a testament to how well hidden the implementation details are.”

After taking just two weeks or so to get through three out of five legs of his speed run, the above concerns expressed regarding GPUs might make readers worry that Majmudar may have hit a speed bump, a snag, or even a brick wall. That doesn’t seem to be the case, as he optimistically predicts that his GPU design will be shipping “in the next few days,” and a cut-down version sent to be taped out.

It may be well worth keeping an eye open for this engineer’s next updates. However, we know it can take quite some time between submitting work to projects like TinyTapeout and the production run. The maker of the Rickroll ASIC, for example, said there were nine months between submitting his design and receiving the silicon. Please note that TT06 closes just eight days from now.

Intel announced China-specific versions of its Gaudi 3 AI accelerators along with the regular versions this week. By contrast, nothing is clear about AMD's Instinct MI300-series products for China or Nvidia Blackwell-based offerings for the People's Republic. Will they ever be released? Analysts from TrendForce have no idea.

Intel is prepping two Gaudi 3 products that are neutered to meet U.S. export control rules for China. One product is the Gaudi 3 HL-328 OAM module, and another is the Gaudi 3 HL-388 PCIe card. Intel doesn't disclose the performance levels of the two AI processors, but both are said to feature a thermal design power of 450W (down from 900W and 600W of fully-fledged versions), which points to radically lowered performance. Gaudi 3-series China-oriented processors have 128 GB of HBM2E memory (using eight stacks with a 1024-bit interface) and 3.7 TB/s of bandwidth. The scale-out capabilities of PRC-specific Gaudi 3 are similar to other Gaudi 3 accelerators. 

Neither AMD nor Nvidia has officially introduced their China-specific offerings based on their latest architectures, such as the CDNA 3 and Blackwell. In fact, we have heard rumors about AMD's Instinct MI309 product designed for export to China, which has failed to meet the U.S. government's export control rules. This indicates that the company wants to sell its CDNA-based products in the People's Republic. We have no idea about Nvidia's Blackwell-related plans for China. 

From a business standpoint, heavily reducing the performance of flagship AI and HPC products to sell them to China is a questionable tactic. On the one hand, companies like AMD, Intel, and Nvidia can sell cut-down processors with minor defects that otherwise would have to be scrapped. On the other hand, TSMC can yield even large chips pretty well, so some cut-down processors are made from fully functional chips. 

One would ask why companies like AMD and Nvidia tend to cut down their latest offerings rather than keep selling previous-generation offerings in China. While this tactic may work, it should be noted that the latest processors also support data formats that are more efficient for AI training and offer better performance-per-watt, so their customers are more inclined to install them rather than previous-generation products. To that end, it makes sense to expect Blackwell and CDNA 3-based products for China.

In an interview cited by The Wall Street Journal earlier this week, Rene Has, CEO of Arm, warned of AI's "insatiable" thirst for electricity, stating an increase to as much as 25% of the U.S.' current 4% power grid usage from AI datacenters is possible. 

Rene himself may have been citing an International Energy Agency report from January stating that ChatGPT consumes roughly 2.9 watt-hours of electricity per request, which is 10 times as much as a standard Google search. Thus, if Google made the full hardware and software switch with its search engine, Google would consume at least 11 terawatt-hours of electricity per year from its current 1 TWh. 

The original report says one example of a standard 2.9-watt-hour would be running a 60-watt-hour lightbulb for just under three minutes. Similar to the standard deviation of ChatGPT queries to standard search engines, industry-wide expectations for Artificial Intelligence power demands are expected to increase tenfold.

These statements were made ahead of an expected U.S. and Japanese partnership in AI and alongside recent developments like OpenAI's Sora, the current version of which Factorial Funds estimates to consume at least one Nvidia H100 GPU per hour to generate five minutes of video. Grok 3 has also been estimated to require 100,000 Nvidia H100s just for training. A single, 700-watt Nvidia H100 can consume roughly 3740 kilowatt-hours per year.

Without great improvements to efficiency and/or greatly increased government regulation, Rene declares the current trend is "hardly very sustainable," and he might be correct. 

The US Energy Information Administration (EIA) stated that the United States generated a total of 4.24 trillion kilowatt-hours, or 4240 terawatt-hours, in 2022, with only 22% of that coming from renewables. This is compared to a total consumption of 3.9 trillion kWh, or 3900 terawatt-hours of the available ~42. 

That's 11 of the 340 remaining terawatt-hours left at current levels that the AI industry seems to be aiming for in the next decade. Sustainability must also keep in mind the likely increasing demands of other industries and the scale of renewable to non-renewable resources. Given that the cost of power has nearly doubled since 1990 (per Statista), perhaps calls for more regulation are justified.

Of course, outlets like The New York Times are also outright suing OpenAI and Microsoft, so it's not like the current AI industry is without existing legal challenges. Rene Haas expressed hope that the international partnership between Japan and the U.S. may yet improve these dramatically high power estimations. However, corporate greed and compute demand are also international, so only time will tell.

According to Digitimes, Dell Taiwan General Manager Terence Liao reports that the delivery lead times of Nvidia H100 AI GPUs have been reduced over the past few months from 3-4 months, as we previously reported, down to just 2-3 months (8-12 weeks). Server ODMs revealed that supply is finally easing up compared to 2023 when it was virtually impossible to attain Nvidia's H100 GPUs.

Despite decreasing lead times, Liao states that demand for AI-capable hardware is still extremely high. Specifically, AI server purchases are supplanting general-purpose server purchases in businesses, even though AI servers are incredibly costly. However, he believes that procurement time is the only reason for this.

A 2-3 month delivery window is the shortest lead time seen for Nvidia's H100 GPUs. Just six months ago, lead times reached 11 months, meaning that most of Nvidia's customers had to wait a year to have their AI GPU orders fulfilled. 

Since the start of 2024, lead times have decreased significantly. First, we saw a huge reduction to 3-4 months earlier in the year. Now, lead times have shrunk by another month. At this rate, we could see lead times evaporate completely by the end of the year or sooner.

This behavior is potentially the result of knock-on effects from some companies having a surplus of H100 GPUs and re-selling some of their supply to offset high maintenance costs of unused inventory. Additionally, AWS has made it easier to rent Nvidia H100 GPUs through the cloud, which is also helping alleviate some of the H100 demand.

The only remaining Nvidia customers struggling with supply constraints are large companies like OpenAI, which are developing their own LLMs. These companies need tens and hundreds of thousands of GPUs to train their LLMs quickly and effectively. 

The good news is that this shouldn't be a problem for long. If lead times continue to get exponentially shorter, as they have in just the past four months, Nvidia's largest customers should be able to get their hands on all the GPUs they need, at least in theory.

Intel has announced that it will be the exclusive CPU launch partner for Star War Outlaws. Touted by Ubisoft and Massive Entertainment as the “first-ever open-world Star Wars game,” Intel partners have already started to bundle 14th Gen processor packing laptops with promotional codes for a free digital copy of the game. Moreover, PC desktop buyers and DIYers will be pleased that they haven’t been excluded from the galaxy’s most wanted bunde promo.

The Intel 14th Gen CPU Star Wars Outlaws bundle began on Tuesday and runs until July 31, 2024. Intel’s full T&Cs are here, but to sum up, the most important step is to make sure the purchasing store or reseller is participating in the Star Wars Outlaws promo. You can do this by checking the online product listings or even asking someone at the store. Make sure that your product is included in the offer, so it will be equipped with one of the qualifying laptop or desktop CPUs as listed below.

Massive Entertainment has used Ubisoft’s Snowdrop engine for the lavish graphics in Star War Outlaws. We note that Massive used this engine to great effect in the open-world The Division 2. In the new game, there are claimed to be “new Intel hybrid CPU optimizations for the Snowdrop game engine,” so things should be even more detailed and smooth. Specifically, Intel claims its engineers have worked with Ubisoft on “improving CPU scheduling to fully utilize the processing power of Intel’s hybrid architecture,” but we would expect the same optimizations to benefit 12th and 13th Gen Core hybrid CPU users. Moreover, in previous Snowdrop games, the GPU choice was probably more important to deliver a fast and smooth gaming experience - after rising beyond a certain minimum CPU spec.

Star War Outlaws takes place in the period between the stories told during The Empire Strikes Back and Return of the Jedi. According to the official promotional material and the story video trailer, embedded below, you will play as Kay Vess, “an emerging scoundrel seeking freedom and the means to start a new life.” During your open-world adventure, living the life of an outlaw, you will explore distinct planets throughout the Star Wars universe. In bustling cities or vast outdoor landscapes, you will face diverse challenges, with great rewards - if you are willing to take the risk. As well as your companion Nix, you will have the Trailblazer ship and planet Speeder to help you fight, steal, and outwit your way through the galaxy.

Those who qualify for and receive their free game code will still have at least a month to wait to start playing, as the game is scheduled for release on August 30, 2024. If you are interested in this game and pondering a new CPU or laptop, this offer could save you $69.99.

A few days ago, we reported that Samsung's investment in Taylor, Texas increased to around $44 billion. Thanks to an additional Reuters report, we now know that a large U.S. government subsidy also seems to be supporting that venture. In exchange for Samsung's more-than-doubling its U.S. investment to $44 million, the government will reward the company with a $6.6 billion subsidy. At least, according to Reuter's sources, who remain undisclosed.

When Reuters reached out to the Commerce Department and Samsung to confirm details of the story, both parties declined to comment. While this could mean the story is untrue, that seems unlikely considering the sheer scale of U.S. investment into Big Tech lately, especially when evaluating the budget behind the CHIPS Act.

In our previous reporting, details pointed to Samsung's Taylor, Texas operations as a focus for semiconductor production and HBM manufacturing. The Samsung campus in Taylor will play a crucial role in advancing the U.S.'s position in the semiconductor industry. While the details of the Taylor facility still seem to be true, Samsung may also be expanding elsewhere in Texas with the help of the subsidy they're receiving, according to the details of the Reuters report. The second location besides Taylor, Texas is currently unknown.

In any case, reports like these indicate that the Biden administration is taking the ongoing Chip War against China very seriously. With billions of dollars in agreements being thrown around seemingly every week, there's no doubt that the country is in the midst of a "fab renaissance." 

This approach may be necessary in today's era of cutting-edge technology used globally for economic, military, and political operations. In any case, the likes of Samsung, TSMC, Intel, and SK hynix all have a lot to gain in the coming year as the CHIPS Act and related initiatives continue funneling investment into these companies.

The Biden administration is expected to officially announce its Samsung subsidy at some point next week.

Elon Musk, CEO of Tesla and founder of xAI, made some bold predictions about the development of artificial general intelligence (AGI) and discussed the challenges facing the AI industry. He predicts that AGI could surpass human intelligence as soon as next year or by 2026, but that it will take an extreme number of processors to train, which in turn requires huge amounts of electricity, reports Reuters.

Musk's venture, xAI, is currently training the second version of its Grok large language mode and expects to complete its next training phase by May. The training of Grok's version 2 model required as many as 20,000 Nvidia H100 GPUs, and Musk anticipates that future iterations will demand even greater resources, with the Grok 3 model needing around 100,000 Nvidia H100 chips to train.

The advancement of AI technology, according to Musk, is currently hampered by two main factors: supply shortages on advanced processors — like Nvidia's H100, as it's not easy to get 100,000 of them quickly — and the availability of electricity.

Nvidia's H100 GPU consumes around 700W when fully utilized, and thus 100,000 GPUs for AI and HPC workloads could consume a whopping 70 megawatts of power. Since these GPUs need servers and cooling to operate, it's safe to say that a datacenter with 100,000 Nvidia H100 processors will consume around 100 megawatts of power. That's comparable to the power consumption of a small city.

Musk stressed that while the compute GPU supply has been a significant obstacle so far, the supply of electricity will become increasingly critical in the next year or two. This dual constraint underscores the challenges of scaling AI technologies to meet growing computational demands.

Despite the challenges, advancements in compute and memory architectures will enable the training of increasingly massive large language models (LLMs) in the coming years. Nvidia revealed its Blackwell B200 at GTC 2024, a GPU architecture and platform that's designed to scale to LLMs with trillions of parameters. This will play a critical role in development of AGI.

In fact, Musk believes than an artificial intelligence smarter than the smartest human will emerge in the next year or two. "If you define AGI as smarter than the smartest human, I think it is probably next year, within two years," Musk said in an interview on X Spaces. That means it's apparently time to go watch Terminator again, and hope that our future AGI overlords will be nicer than Skynet. ☺

Intel announced it's finally bringing LGA1851 socket-compatible Core Ultra processors based on the Meteor Lake architecture to the market. But before you get your hopes up, these chips are targeted at edge computing rather than consumer desktops. Intel's new edge-focused CPUs will be the first-ever Meteor Lake CPUs to come in an LGA socket format.

The chips are optimized for on-premise edge computing. Intel reports that its Core Ultra processors for edge will offer up to 5X better image classification inference performance compared to its outgoing 14th Gen Raptor Lake Refresh desktop processors. The new chips incorporate all of Intel's bleeding-edge Meteor Lake architecture, including an Arc-based iGPU and an NPU for graphics and AI-based workloads.

The chips target customers in retail, education, industrial, kiosk, and smart point-of-sale-systems for brick-and-mortar stores, to name a few. Edge computing is a large business for Intel and encompasses more than 90,000 deployments alone. These new Meteor Lake parts, announced at the company's Vision 2024 event, will likely sell like hotcakes as the industry continues to shift computing strategies from mainstream cloud solutions to edge-focused applications.

Edge computing is nothing new, but its capabilities have become extremely popular over the past few years. Edge computing is a strategy that localizes most of the computing power that is usually provided by cloud services. In a practical application, devices within a theoretical business will connect directly to a small mainframe located inside the building for data processing. This is different from traditional cloud computing where business devices (like kiosks) connect directly to a cloud service provider.

Edge computing has massively grown in popularity over the past few years as data consumption has ballooned beyond what the internet can handle. According to TechTarget, 75% of enterprise-generated data is expected to be created outside centralized data centers, including through edge-based solutions.

So far Intel has unveiled five Meteor Lake PS 15W SKUs and four Meteor Lake PS 45W models. The 15W parts include Core Ultra 3, Core Ultra 5, and Core Ultra 7 SKUs with core counts ranging from 8 to 12. All of these chips have just two P-cores while the rest of the cores are E-cores and LPE-cores. Boost clocks range from 4.2 GHz to 4.9 GHz on the P-cores, depending on the model.

The Meteor Lake PS 45W chips have far more cores, ranging from 14 to 16 total. The Core Ultra 5 models come with four P-cores, eight E-cores, and two LPE-cores, while the higher end Core Ultra 7 models have six P-cores, eight E-cores, and two LPE-cores. Clocks range from 4.5 GHz to 5 GHz, depending on the model.

While these are edge-focused CPUs, the fact that they use socket LGA1851 means they could be compatible with future LGA1851 desktop boards. LGA1851 is the successor to LGA1700 and is expected to arrive on the consumer side with Intel Arrow Lake CPUs. Compatibility will come down to the firmware availability and whether or not Intel wants to include support for the processors on all platforms and chipsets, or if support will be restricted to specific boards.

Regardless, if you're an enthusiast or a desktop user, Intel's edge-focused chips shouldn't be on your radar. Intel is already preparing to launch an enhanced desktop version of Meteor Lake (codenamed Arrow Lake) later this year, sporting a more advanced Intel processing node and 5x the AI performance, and likely with more CPU cores. Arrow Lake will be optimized for gaming and desktop computing and will be the first Intel desktop CPU lineup to incorporate an NPU.

Following last year's reports of a "potential" NAND shortage, and with the recent Taiwan Earthquake further disrupting the tech industry's supply chain, Western Digital has officially informed its partners that it will "continue to implement price increases on flash and hard drive products this quarter." These increases started on April 8th [h/t Trendforce, original source TechNews.TW].

The following screenshot shows the official letter that was sent to partners and provided to TechNews.TW. The original coverage posits that the booming AI market is most likely responsible for the surge in demand that WD refers to in the message.

Scott Davis, Senior VP of WD's Channel and Regional OEM Sales, likely sent this notification to direct business partners. However, it's certainly just as important for consumers moving into 2024. Higher-than-expected demand combined with supply constraints have pushed us out of the era of dirt-cheap storage now, as was rumored last fall. Hopefully, the increased prices won't last too long.

TechNews.TW's sources point toward HDD pricing increasing by as much as 10%, with a minimum 5% pricing increase expected. There are no hard numbers for NAND flash pricing, but NAND prices were already heading up and we expect further pricing increases could be in line with HDDs, if not worse.

Trendforce's coverage points to its own Q3 2023 revenue data, where WD is the third-largest NAND flash manufacturer, with Samsung and SK hynix taking the top two spots. Phison's CEO warned back in March that the ever-rising prices of SSDs could "hobble NAND flash industry with reduced demand." That doesn't seem to be the case for now, but it will take time for the effects of the price increases to propagate through the chain.

Last year saw prices on the best SSDs fall to under $100 for decent quality 2TB M.2 PCIe 4.0 models, with even higher-end options like the WD Black SN850X and Samsung 990 Pro falling below $125. Currently, the WD Black SN850X 2TB costs $156 while the Samsung 990 Pro 2TB costs $176. Larger 4TB drives were going for as little as $160, compared to today's $209 for the Silicon Power UD90 4TB.

Hopefully you were able to take advantage of the low prices while they lasted. We warned in November that Black Friday / Cyber Monday might be the last chance for a a dirt-cheap SSD, and that has proven to be the case. This statement from Western Digital only points toward pricing getting worse before it gets better. While we wait for prices to come back down, check our routinely updated SSD and HDD Deals, where we'll try to help you make the most of these market conditions.

Intel CEO Pat Gelsinger flashed a Lunar Lake processor, the company's next-gen laptop chip, at its Vision 2024 event, saying the chip provides 100+ TOPS of performance in AI workloads, with 45 of those TOPS coming from the NPU alone. That's 3X the AI performance of Intel's current-gen chips and meets the bar of 45 TOPS from the NPU that the company recently discussed during its AI Summit in Taipei as the requirement for next-gen AI PCs. At that event, Intel executives, in a question-and-answer session with Tom's Hardware, also said that elements of Microsoft's Copilot will soon run locally on Windows PCs. 

Intel's current-gen Meteor Lake chips only provide 10 TOPS of performance from the NPU, falling below the bar for next-gen AI PCs. Lunar Lake's 45 TOPS from the NPU meets that bar exactly. Gelsinger didn't elaborate on how much of the remaining 55+ TOPS comes from the CPU and GPU, but it's reasonable to expect in the range of 50 TOPS from the GPU and 5 to 10 FLOPS from the CPU cores.

AMD's current-gen Ryzen Hawk Point platform has an NPU with 16 TOPS of performance, which also falls below the bar for next-gen AI PCs. However, the company recently sent us a statement regarding its next-gen products: 

“We believe an AI PC requires strong CPU, GPU and NPU engines, which is what AMD has been delivering for more than a year with our Ryzen 7040 and now 8040 Series. At our December Advancing AI event, we disclosed our next-gen “Strix Point” mobile processors with XDNA 2 architecture would have up to 3x the generative AI performance of the current generation. We believe this performance will position us to remain the leading choice for next-gen AI PCs." — AMD representative to Tom's Hardware.

AMD wouldn't elaborate further and break out the amount of performance delivered from each unit, but simple math puts AMD's Strix Point NPU at 48 TOPS if NPU performance triples.

Qualcomm's Snapdragon X Elite chips are the elephant in the room — these Arm chips will debut with 45 TOPS of performance from its NPU (75 TOPS total) in the "middle of the year," setting up a tight race between Intel, AMD, and Qualcomm for the lead in the next generation of AI PCs. 

Intel appears to have a running start with its Core Ultra Meteor Lake chips. Gelsinger says the company has already shipped five million AI PCs to date and plans to ship 40 million units by the end of the year. That's a solid start towards the company's goal of shipping 100 million AI PC chips by the end of 2025. 

The roll out of Intel's first generation Arc Alchemist GPUs back in 2022 was a slow process that took several quarters, which ultimately hurt sales and adoption rates and relegating them to the budget end of the best graphics cards. With its second generation Arc Battlemage products, the company doesn't want to repeat the same mistake and intends to introduce them this year ahead of the critical holiday season. This is according to a report from ComputerBase, citing sources at Embedded World.

Intel aims to launch its second generation Arc 'Battlemage' graphics processors this fall, ideally capturing the holiday sales surge that kicks off around Black Friday, with plans for a release by November at the latest, according to the report. This timing would let Intel capitalize on the lucrative holiday shopping season, a critical sales period in the retail calendar, especially in the U.S.

The timeline suggests that Intel's Battlemage launch should occur by November at the latest in a bid to let Intel's add-in-board partners to have enough time to ramp up production of their graphics cards and ship actual products to retailers worldwide. However, the initial launch could occur earlier — exact details are not yet fully nailed down.

Intel is in the process of bringing up two new Battlemage discrete graphics processing units, known as Battlemage-G10 and Battlemage-G21, based on recent leaks. The naming convention suggests that Battlemage-G10 will be the more powerful of the two, targeting the midrange market, while Battlemage-G21 will cater to entry-level systems that still need a standalone graphics processor.

Neither Battlemage-G10 nor Battlemage-G21 GPUs are finalized yet, as G21 is rumored to be in the pre-qualification stage. This stage involves testing the chip's functionality, reliability, and performance, but it does not guarantee readiness for mass production. However, if these pre-qualification tests are successful and the GPUs meet the necessary criteria for performance, power, and yields, they could move forward to mass production. It's unclear whether Battlemage-G10 is at a similar state or if it's further out — note that the previous Alchemist generation launched the smaller ACM-G11 first, in the Arc A380.

Intel's first generation Arc Alchemist only barely managed to compete against mainstream GPUs from AMD and Nvidia — the Arc A750 matched up decently against the RTX 3060 12GB and RX 6600 XT, both of which were over a year old at the time. In fact, Nvidia and AMD were in the midst of launching their newer generation parts that would widen the performance deficit. This was all largely thanks to Alchemist being delayed by several quarters due to driver readiness. With Battlemage, Intel will have more or less stable drivers and an established driver development team, so the new family of chips won't face the same challenges.

What remains to be seen is what market segments will Intel be able to address with its Battlemage GPUs. Intel potentially faces a similar problem as last time, as potentially Nvidia's Blackwell-based RTX 50-series and AMD's RDNA 4-powered RX 8000-series could also start showing up before the end of the year. Battlemage should deliver plenty of improvements, and hopefully Intel will be better prepared for the fight against AMD and Nvidia.

Intel announced at its Vision 2024 event that the next-gen Granite Rapids and Sierra Forest processors will be branded ‘Xeon 6,’ thus dispatching with its old nomenclature that used generational designators, like the preceding ‘Fifth-Gen Xeon Scalable’ models. This new branding somewhat mirrors Intel's new simplified branding for its consumer chips, which it announced last year. Intel also provided an update on the launch schedule, indicating the Sierra Forest and Granite Rapids chips remain on track. The company also briefly mentioned new software support for the MXFP4 data format, which the company says allows the CPUs to run 70 billion parameter Llama-2 models and reduce next-token latency by up to 6.5x compared to fourth-gen Xeon. 

Intel’s existing 'Emerald Rapids' Fifth-Gen Xeon models won’t see a rebrand, so the new branding nomenclature will apply only to Xeon 6 and newer processors. The company says the new branding makes it “easier to tell a unified Intel Xeon story, eases customer navigation with intuitive organization, and amplifies the performance signals.”

As you can see in this table that outlines the current-gen Xeon product stack, Intel’s branding can be more than a little confusing. Intel says this new branding will be simpler and easier to understand, and it will discard some of the sub-brands, though the company didn’t provide further details.

Intel’s current product stack has 32 models divided into six primary swim lanes (second slide). These include processors designed for the cloud, networking, storage, long-life use, single-socket models, and processors designed specifically for liquid-cooled systems. The stack is also carved into Platinum, Gold, Silver, and Bronze sub-tiers. 

It isn’t yet clear how Intel will arrange the Xeon 6 lineup to reduce confusion, but with plans to launch Sierra Forest this quarter, we expect to learn more details soon.

Intel says its Xeon 6 Sierra Forest processors, the company’s first data center chips to have an architecture comprised entirely of efficiency cores (E-cores), will arrive as scheduled this quarter. To that effect, Supermicro also announced the early availability of Sierra Forest chips in its Early Ship program and free remote access for testing and validation through its JumpStart program. 

Intel has already detailed both its Sierra Forest and Granite Rapids processors. The Sierra Forest chips have an E-core-based architecture designed to deliver the ultimate threaded heft for scale-out, cloud-native, and containerized environments. You can read the deep-dive architectural details here. The chips employ CPU chiplets with the Intel 3 process combined with twin I/O chiplets based on the Intel 7 node to provide a flexible architecture that can scale to higher core counts by adding more chiplets.

As before, Intel claims Sierra Forest delivers a 2.4X improvement in performance per watt and a 2.7X improvement in performance per rack. Intel also claims that 72 Sierra Forest server racks will provide the same amount of performance as 200 server racks equipped with the aging second-gen Xeon models, thus saving a megawatt of power. The company has also announced a 288-core Sierra Forest model that will arrive later this year.

Intel's Granite Rapids processors slot in for general-purpose workloads and feature a more standard performance-core (P-core) based design. Intel's press release mentions that the processors now support the MXFP4 data format, a new narrow-precision format that's championed by OCP with broad support from other vendors, like Nvidia, AMD, and Arm, among many others. 

Intel says MXFP4 software support enables Granite Rapids to reduce next token latency by up to 6.5x over its older fourth-gen Xeon using FP16, and that it allows the CPU to run a 70 billion parameter Llama-2 model. Intel hasn't shared any more detail about this caveat beyond a bullet point on the press release, but we'll be sure to follow up as more details emerge.

Intel says the Granite Rapids processors will follow ‘soon after’ Sierra Forest but hasn’t committed to delivering them in a specific quarter.

Intel made a slew of announcements during its Vision 2024 event today, including deep-dive details of its new Gaudi 3 AI processors, which it claims offer up to 1.7X the training performance, 50% better inference, and 40% better efficiency than Nvidia’s market-leading H100 processors, but for significantly less money. Intel also announced new branding for its data center CPU portfolio, with the Granite Rapids and Sierra Forest chips now branded as the new ‘Xeon 6’ family. Those chips are on track to come to market this year and add support for the new performance-boosting standardized MXFP4 data format, which you can read about here.

Intel also announced that it is developing an AI NIC ASIC for Ultra Ethernet Consortium-compliant networking and an AI NIC chiplet that will be used in its future XPU and Gaudi 3 processors while also being made available to external customers through Intel Foundry, but it didn’t share more details on these networking products. 

Nvidia’s dominance in AI infrastructure and software is undisputed. Still, Intel, like AMD, is looking to carve out a position as the premier alternative to Nvidia as the industry continues to struggle with Nvidia's crushing AI GPU shortages. To that end, Intel also outlined the full breadth of its AI enablement programs, which stretch from hardware to software, as it looks to gain traction in the booming AI market that Nvidia and AMD currently dominate. Intel’s efforts focus on developing its partner ecosystem to deliver complete Gaudi 3 systems while also working to build an open enterprise software stack to serve as an alternative to Nvidia’s proprietary CUDA.

Intel provided deep-dive details of the Gaudi 3 architecture along with plenty of convincing benchmarks against existing Nvidia H100 GPUs (data for the upcoming Blackwell systems isn’t available yet). First, let’s take a closer look at the Gaudi 3 architecture.

Intel Gaudi 3 Specifications

Intel’s Gaudi 3 marks the third generation of the Gaudi accelerator, which was the fruit of Intel’s $2 billion acquisition of Habana Labs in 2019. The Gaudi accelerators will enter high-volume production and general availability in Q3 of 2024 in OEM systems. Intel will also make Gaudi 3 systems available in its Developer Cloud, thus providing a fast on-ramp for prospective customers to test the chips.

Gaudi comes in two form factors, with the OAM (OCP Accelerator Module) HL-325L being the common mezzanine form factor found in high-performance GPU-based systems. This accelerator has 128GB of HBM2e (not HBM3E), providing 3.7 TB/s of bandwidth. It also has twenty-four 200 Gbps Ethernet RDMA NICs. The HL-325L OAM module has a 900W TDP (higher TDPs are possible, ostensibly with liquid cooling) and is rated for 1,835 TFLOPS of FP8 performance. The OAMs are deployed in groups of eight per server node and can then scale up to 1,024 nodes.

Intel claims Gaudi 3 provides twice the FP8 and four times more BF16 performance than the prior generation, along with twice the network bandwidth and 1.5X the memory bandwidth.

The OAMs drop into a universal baseboard that houses eight OAMs. Intel has already shipped OAMs and baseboards to its partners as it prepares for general availability later this year. Scaling to eight OAMs on the HLB-325 baseboard brings performance to 14.6 PFLOPS of FP8, while all other metrics, such as memory capacity and bandwidth, scale linearly.

Intel has a Gaudi 3 PCIe dual-slot add-in card with a 600W TDP as well. This card also has 128GB of HBMeE and twenty-four 200 Gbps Ethernet NICs — Intel says dual 400 Gbps NICs are used for scale-out. Intel says the PCIe card has the same peak 1,835 TFLOPS of FP8 performance as the OAM, which is interesting given its 300W lower TDP (this likely won't hold in long-duration workloads). However, scaling is more limited inside the box, as it is designed to work in groups of four. Intel says this card can also scale out to create larger clusters but didn’t provide details.

Dell, HPE, Lenovo, and Supermicro will provide systems for the Gaudi 3 launch. Gaudi air-cooled models have already been sampled, with sampling of liquid-cooled models to follow in Q2. These will be generally available (mass production) in Q3 and Q4 of 2024, respectively. The PCIe card will also be available in Q4.

Gaudi 3 leverages the same architecture and underlying fundamental principles as its predecessor but comes with a more advanced TSMC 5nm process than the TSMC 7nm node Intel uses for the Gaudi 2 accelerator.

The OAM design has two central 5nm dies with 96MB of SRAM split between the two, providing 12.8 TB/s of bandwidth. The dies are flanked by eight HBM2E packages, totaling 128GB, that deliver up to 3.7 TB/s of bandwidth. A high-bandwidth interconnect between the two dies provides access to all memory present on both dies, thus allowing it to look and act as a single device (at least as far as the software is concerned - latency might vary). Gaudi 3 also has a x16 PCIe 5.0 controller for communication with the host processor (CPU) and different ratios of CPUs and Gaudi accelerators can be employed.

Compute is handled by 64 fifth-gen Tensor Processing Cores (TPC) and eight matrix math engines (MME), with workloads orchestrated between the two engines by a graph compiler and the software stack. The Gaudi 3 chip package also includes twenty-four 200 Gbps RoCE Ethernet controllers that provide both scale-up (in-box) and scale-out (node-to-node) connectivity, doubling the 100 Gbps connections on Gaudi 2.

Gaudi 3 Scalability

At the end of the day, the key to dominating today’s AI training and inference workloads resides in the ability to scale accelerators out into larger clusters. Intel’s Gaudi takes a different approach than Nvidia’s looming B200 NVL72 systems, using fast 200 Gbps Ethernet connections between the Gaudi 3 accelerators and pairing the servers with leaf and spine switches to create clusters.

Nvidia’s system-level architecture utilizes NVLink over the PCIe interface for both in-box connectivity between GPUs and stretching out to connect entire racks with passive copper cabling via its NVLink switches. AMD also has its own approach of using the PCIe interface and its Infinity Fabric protocol between GPUs residing in the server while using external NICs for communication with other nodes, but this adds more networking cost and complexity than Intel’s approach of having networking NICs built right into the chip.

Thanks to the doubled network bandwidth, Gaudi 3 scales from a single node with 8 OAM Gaudis to clusters with up to 1,024 nodes (servers) housing 8,192 OAM devices.

Each server is comprised of eight Gaudi 3 accelerators communicating with each other via twenty-one 200 Gbps Ethernet connections apiece. The remaining three Ethernet ports on each device are used for external communication with the cluster via a leaf switch. The switch aggregates these connections into six 800 Gbps Ethernet ports with OFSP connectors to facilitate communication with other nodes.

Each rack typically contains four nodes, but this can vary based on rack power limits and cluster size. Up to 16 nodes form a single sub-cluster with three Ethernet leaf switches, which then connect to spine switches, typically with 64 ports, to form even larger clusters. Half of the 64 ports on the 800 Gbps leaf switch are connected to the 16 nodes, while the remaining half connect to the spine switches.

Varying numbers of spine switches are used based on the size of the cluster, with Intel providing an example of three spine switches used for 32 sub-clusters comprised of 512 nodes (4,096 Gaudi’s). Intel says this configuration provides equal bandwidth for all server-to-server connections (non-blocking all-to-all). Adding another layer of Ethernet switches can support up to tens of thousands of accelerators.

Gaudi 3 performance vs Nvidia H100

Intel shared performance projections for Gaudi 3, but as with all vendor-provided benchmarks, we should take these with a grain of salt. As you’ll see in the last image in the above album, Intel is now simply providing a QR code for information about its benchmarks instead of the line-by-line details of test configurations that it has provided in the past. This code doesn’t provide us with any meaningful way to look at the closer details of these test results and configurations, so add a shovelful of salt to any of these benchmark claims.

Intel compared to publicly available benchmarks for H100 systems but didn’t compare to Nvidia’s upcoming Blackwell B200 due to a lack of real-world comparative data. The company also didn’t provide comparisons to AMD’s promising Instinct MI300 GPUs, but that's impossible because AMD has continued to avoid publishing public performance data in the industry-accepted MLPerf benchmarks.

Intel provided plenty of comparisons in both training and inference workloads compared to the H100 with similar cluster sizes, but the key takeaway is that Intel claims Gaudi is from 1.5X to 1.7X faster in training workloads. Comparisons include LLAMA2-7B (7 billion parameters) and LLAMA2-13B models with 8 and 16 Gaudi’s, respectively, and a GPT 3-175B model tested with 8,192 Gaudi accelerators, all using FP8. Interestingly, Intel didn’t compare to Nvidia’s H200 here, which has 76% more memory capacity and 43% more memory bandwidth than the H100.

Intel did compare to the H200 for its inference comparisons but stuck to performance with a single card as opposed to comparing scale-out performance with clusters. Here we can see a mixed bag, with five of the LLAMA2-7B/70B workloads coming in 10 to 20% below the H100 GPUs, while two match and one slightly exceeds the H200. Intel claims Gaudi's performance scales better with larger output sequences, with Gaudi delivering up to 3.8 times the performance with the Falcon 180 billion parameter model with a 2,048-length output.

Intel also claims up to a 2.6X advantage in power consumption for inference workloads, a critical consideration considering restrictive power limits in data centers, but it didn’t provide similar benchmarks for training workloads. For these workloads, Intel tested a single H100 in a public instance and logged the H100’s power consumption (as reported by the H100) but didn’t provide examples of inference with a single node or larger clusters. With larger output sequences, Intel again claims better performance and, thus, efficiency.

Gaudi 3 software ecosystem

As Nvidia’s dominance with CUDA has illustrated, the software ecosystem is just as critical of a consideration as the hardware. Intel touts its end-to-end software stack and says that “most” of its engineers are currently working on bolstering support. Intel’s current focus is on supporting multi-modal training and inference models and RAG (retrieval augmented generation).

Hugging Face has over 600,000 AI model checkpoints available, and Intel says its work with Hugging Face, PyTorch, DeepSpeed, and Mosaic has eased the software porting process to enable faster turnaround times for deploying Gaudi 3 systems. Intel says most programmers are programming at the framework level or above (i.e., simply using PyTorch and scripting with Python) and that low-level programming with CUDA isn’t as common as perceived.

Intel’s tools are designed to ease the porting process while abstracting away the underlying complexity, with OneAPI serving as the underlying kernel and communication libraries. These libraries adhere to the specifications outlined by the Unified Accelerator Foundation (UXL), and industry consortium that includes Arm, Intel, Qualcomm, and Samsung, among others, that's intended to provide an alternative to CUDA. PyTorch 2.0 is optimized for using OneAPI for inference and training with Intel CPUs and GPUs. Intel says its OpenVino also continues to enjoy rapid adoption, with over one million downloads so far this year.

Thoughts

As we covered above, Intel, Nvidia, and AMD are all taking different paths to providing the enhanced cluster scalability that is key to performance in both AI training and inference workloads. Each approach has its own respective strengths, but Nvidia's proprietary NVLink is the most mature and well-established solution, and its extension to rack-scale architectures is a significant advantage. That said, Intel's approach with Ethernet-based networking brings an open solution that affords plenty of customization options by supporting networking switches from multiple vendors, and its built-in NICs also offer cost advantages over AMD's competing Instinct MI300 series. 

However, both Nvidia's Grace-based products and AMD's MI300A offer sophisticated merged CPU+GPU packages that will be hard to beat in some workloads, while Intel continues to rely on separate CPU and accelerator components due to its cancellation of the merged CPU+GPU version of Falcon Shores. There have been reports of Nvidia's new GB200 CPU+GPU servers comprising the bulk of the company's Blackwell orders, highlighting the industry's voracious appetite for these types of tightly-coupled products.

Intel's future Falcon Shores product will arrive as an AI accelerator-only design, so it will still be able to compete with GPU-only Nvidia and AMD clusters. We also see room for a refresh generation of Gaudi 3 that moves from HBM2E to HBM3/E—both AMD and Nvidia employ the faster memory in their AI products. Although it hasn't shared hard data, Intel says it also plans to compete aggressively on pricing, which might be a powerful recipe as Nvidia continues to grapple with shortages due to crushing demand for its GPUs.

Falcon Shores will also be compatible with code optimized for Gaudi, providing forward compatibility. Intel also cites a 3X improvement to its Gaudi 2 platform over the last several quarters as an example of increased adoption of its platform. 

Notably, Intel didn't promote its Ponte Vecchio GPUs for the event, which isn't surprising given its cancellation of the next-gen Rialto Bridge GPUs, so we expect the company's AI efforts to coalesce solely on Gaudi 3 as it prepares Falcon Shores for launch next year.

The air-cooled Gaudi 3 models are already sampling to partners, with general availability in Q3. Liquid-cooled models will follow in Q4. We're watching the Intel Vision webcast for further details and will update as neccesary.

IDC's latest data shows positive growth in global PC sales during the first quarter of 2024, with a return to pre-pandemic sales volumes. After declining for the past two years, worldwide PC sales amounted to 59.8 million during Q1 2024, comparable to Q1 2019 sales when 60.5 million PCs were sold worldwide.

The global PC shipments include desktops, notebooks and workstations. Though tablets are usually considered personal computing devices, they're not included in IDC's stats. Among the top five OEM PC makers, Lenovo continues to enjoy its dominant position with 23% of the market share, followed by HP with 20%.

Growth in the global markets does not necessarily mean every company has benefited. Dell, Asus, and the "others" category all show a decrease in yearly sales during this quarter. Apple on the other hand shows the highest growth of 14.6%, likely thanks in part to the new M3 laptops.

Here's the look at the top five companies as tracked by IDC, with comparisons with Q1 2023:

* — IDC declares a statistical tie when there's less than a 0.1% difference between vendors, which is why Acer and Asus are listed at number five.

"Despite China's struggles, the recovery is expected to continue in 2024 as newer AI PCs hit shelves later this year and as commercial buyers begin refreshing the PCs that were purchased during the pandemic," explained Jitesh Ubrani, a research manager for IDC's Worldwide Mobile Device Trackers. "Along with growth in shipments, AI PCs are also expected to carry higher price tags, providing further opportunity for PC and component makers."

Intel, AMD, and others have been talking a lot about the "AI PC" over the past few months. Definitions are still rather nebulous, and so far there's no killer app — Microsoft's Copilot isn't there yet, in other words. But things can change rapidly in the AI space, as the last year has aptly demonstrated.

Two years ago, IDC predicted that the market would begin to recover in 2024, despite 2022 and 2023 showing no signs of recovery. The market situation began to fall in 2020. During one quarter, all major companies experienced losses except Apple, Inc. Which isn't to say that Apple hasn't had bad quarters at other times.

US Sanctions Affecting Global PC Shipment Volumes

While global PC sales indicate positive growth, sanctions against China and Russia have likely affected these numbers despite the recovery. Many companies have pulled out of Russia and China, leading those countries to invest in domestic alternatives and giving smaller, local competitors the chance to gain a market share.

IDC notes that China is the largest market for desktop PCs, and before trade sanctions, there was already weak demand there, resulting in lower PC shipments in the region. However, such reports calculate statistics using shipments to distributors and consumers directly, and obviously smuggling operations and the like aren't part of the data gathering.

Overall, PC sales are strong, with new technologies potentially spurring an earlier upgrade cycle in some businesses. Q1 tends to be a weaker quarter for sales, with Q4 being the strongest, so it will be interesting to see how the rest of the year develops.

The U.S. government has announced plans to provide TSMC financial support totaling $11.6 billion, comprising $6.6 billion in grants and up to $5 billion in loans. TSMC will use the funding to build three semiconductor production facilities in Arizona, with a total investment exceeding $65 billion. This is the largest financial package the U.S. government approved under the CHIPS and Science Act.

TSMC will build the third fab module near its Fab 21 phase 1, which should start the production of chips using 4nm and 5nm-class process technologies in the first half of 2025. Fab 21 phase 2 is scheduled to begin operations in 2028, making chips on 2nm and 3nm process technologies. The third plant will likely use an even more advanced fabrication process (TSMC says to expect something at 2nm and sub-2nm), though it is unknown when it will start producing chips. 

TSMC's investment should have a significant economic impact, creating 6,000 high-tech manufacturing jobs and over 20,000 construction jobs. The project also includes a $50 million allocation for training local workers. Arizona is expected to benefit significantly from this investment, with the semiconductor industry's expansion being a central element of President Joe Biden's economic agenda.

TSMC's projects in Arizona have faced challenges, including delays due to labor disputes and uncertainties regarding government support. The second facility's production timeline has been pushed back from 2026 to 2028. Additionally, Bloomberg reports that at least one TSMC supplier has abandoned its planned Arizona project, citing workforce difficulties. 

Financially, TSMC plans to apply for U.S. Treasury Department Investment Tax Credits of up to 25% of the qualified capital expenditure at TSMC Arizona. Meanwhile, the payout of the promised funds to TSMC is subject to a due diligence period, followed by the fulfillment of construction and production benchmarks. The final agreement between TSMC and the U.S. government is still pending, and the funds may be subject to return provisions if TSMC fails to meet its commitments on time.

The financial support package for TSMC is part of President Joe Biden's push to revitalize the U.S. semiconductor industry under the 2022 CHIPS and Science Act. This legislation allocates $39 billion in direct grants, alongside $75 billion in loans and guarantees, to encourage semiconductor companies to establish manufacturing operations in the U.S. This move aims to reverse the trend of outsourcing production overseas and strengthen the country's technological independence.

The local government of Shenzhen has launched a program to significantly boost the ecosystem of applications for Huawei's HarmonyOS, in a bid to make the platform more robust and competitive against foreign peers, reports ChinaScope citing the official Xinhua news agency. The story at Xinhua may signal that China has a broader plan for HarmonyOS. However, Huawei may have different goals for its HarmonyOS Next, reckons the Nikkei. 

Huawei's HarmonyOS — largely based on the open-source version of Google's Android — was a way for the company to get around crippling U.S. sanctions in 2019 and keep producing its popular high-end smartphones. Eventually, it was adopted for a whole range of Huawei's devices, including PCs, tablets, smartwatches, and even TVs. For now, HarmonyOS seems to be quite competitive in the smartphone market, though without apps like Gmail, Facebook, Instagram, and YouTube, it is not exactly competitive in Europe or the U.S. But for Chinese authorities and Huawei the HarmonyOS operating system could still be the most viable way to replace Android, iOS, macOS, and Windows-based hardware in China.

The Shenzhen Action Plan for HarmonyOS

This is perhaps why the Shenzhen government recently kicked off the 'Shenzhen Action Plan for Supporting the Development of Native HarmonyOS Open Source Applications in 2024.' The plan sets several goals to boost the development of software for HarmonyOS in the city and province.  

One major goal for native HarmonyOS applications in Shenzhen is to account for more than 10% of China's total by 2024. The plan includes establishing at least two specialized industrial parks for HarmonyOS software development in a variety of applications. It also sets a target for over 1,000 software companies in Shenzhen to have HarmonyOS development talent qualifications.  Furthermore, it encourages eligible companies to expand their outsourcing services for HarmonyOS application development, aiming to reach a scale of 500,000 HarmonyOS developers, which is pretty impressive if it comes to fruition. The initiative also aims for these applications to be fully adopted across Shenzhen's main industries.  

Huawei's Plan for HarmonyOS

Since HarmonyOS is open source, it can be adopted by everyone and for pretty much everything, just like the open-source Android. This will certainly make it a competitor for the closed-source Android and Apple's proprietary iOS according to the Nikkei story. 

Indeed, Huawei is teaming up with app developers and training coders to make apps just for HarmonyOS, aiming to create around three million jobs in the process. However, for now, this work is mostly focused on smartphones and IoT rather than on client PCs. 

Huawei's smartphones, especially the Mate 60 series, have been selling well in China, helping to set the stage for HarmonyOS. The company has grabbed a good chunk of the market share in China, with 16% of smartphones sold there now running HarmonyOS, which is important as it helps to persuade developers to make apps for this OS. 

Huawei plans to add over 5,000 apps to HarmonyOS to cover what most people do on their smartphones every day in China, which is presumably enough to compete against Apple (iOS) and Samsung (customized Android). For now, Huawei has some big names like Bilibili and Alipay on board to make apps for the new system. However, some big apps and smaller developers are still on the fence because of issues like having to share revenue with app stores. Then again, a journey of a thousand miles begins with a single step. This step has been made for sure.

Originally a part of Arm Holdings that licensed its IP in China, Arm China began to develop its own IP specifically for customers in the People's Republic a few years ago. The company does not develop its own high-performance general-purpose cores, but its neural processing unit (NPU) can become a significant building block for inexpensive processors now that its open-source driver will soon be added the Linux kernel, reports Phoronix.

This driver, which is open-source for both the kernel and user space, aims to bring Arm China's Zhouyi AI accelerator technology into the mainstream Linux world. The move will make the IP block easy to program for software developers, making it much more attractive for chip designers. While Zhouyi will hardly directly compete against AMD's Ryzen AI engine, Apple's NPU, or Intel's AI accelerator, it will have a chance to sneak into potentially tens of millions of devices, just like the aforementioned technologies, which will put Arm China's platform on local AI inference map.

Developer boards with Arm China's Zhouyi AI accelerator technology, such as the EAIDK310 development board, have been available at Aliexpress and Amazon for some time, so at least some developers have had a chance to play with this IP. The EAIDK310 uses a Rockwell-designed processor with the Zhouyi AI accelerator, but Phoronix claims that there is also the Allwinner R329 system-on-chip that packs the same IP.

The project started three years ago with a kernel-only driver, and since then, Arm China has developed a full driver stack. This development is a big step forward in making AI accelerator technology more widely available and integrated into the Linux ecosystem, but Phoronix reports that we are not quite there yet.

Currently, the discussion thread on the mailing list is primarily focused on expressing interest and determining the criteria for potentially integrating the Arm China NPU driver into the main Linux kernel. The process will involve reviews by the upstream DRM/accel maintainers, who will check the code to ensure it aligns with the necessary interfaces and other requirements.

We've previously reported on industry rumors that Microsoft's Copilot AI service will soon run locally on PCs instead of in the cloud and that Microsoft would impose a requirement for 40 TOPS of performance on the Neural Processing Unit (NPU), but we had been unable to get an on-the-record verification of those rumors. That changed today at Intel's AI Summit in Taipei, where Intel executives, in a question-and-answer session with Tom's Hardware, said that Copilot elements will soon run locally on PCs. Company representatives also mentioned a 40 TOPS requirement for NPUs on next-gen AI PCs.

Microsoft has been largely silent about its plans for AI PCs and even allowed Intel to officially announce Microsoft's new definition of an AI PC. Microsoft’s and Intel’s new co-developed definition states that an AI PC will have an NPU, CPU, GPU, Microsoft’s Copilot, and a physical Copilot key directly on the keyboard. 

PCs meeting those requirements are already shipping, but that is just the first wave of the AI PC initiative. Intel divulged future AI PC requirements in response to my questions about potential memory criteria. 

"But to your point, there's going to be a continuum or an evolution, where then we're going to go to the next-gen AI PC with a 40 TOPS requirement in the NPU," said Todd Lewellen, the Vice President of Intel's Client Computing Group. "We have our next-gen product that's coming that will be in that category." 

"[..]And as we go to that next gen, it's just going to enable us to run more things locally, just like they will run Copilot with more elements of Copilot running locally on the client. That may not mean that everything in Copilot is running local, but you'll get a lot of key capabilities that will show up running on the NPU."

Currently, Copilot computation occurs in the cloud, but executing the workload locally will provide latency, performance, and privacy benefits. Notably, Intel's shipping Meteor Lake NPU offers up to 10  TOPS for the NPU, while AMD's competing Ryzen Hawk Point platform has an NPU with 16 TOPS, both of which fall shy of the 40 TOPS requirement. Qualcomm will have its oft-delayed X Elite chips with 45 TOPS of performance in the market later this year. 

Lewellen explained that Microsoft is focused on the customer experience with the new platforms. As such, Microsoft insists that Copilot runs on the NPU instead of the GPU to minimize the impact on battery life. 

"We had a lot of discussions over the course of the last year[with Microsoft], and we asked, 'Why can't we just run it on the GPU?' They said they want to make sure that the GPU and the CPU are freed up to do all this other work. But also, we want to make sure it's a great battery life experience. If we started running Copilot and some of those workloads on the GPU, suddenly you're going to see a huge hit on the battery life side," Lewellen explained. 

While AMD holds a slight lead in NPU TOPS performance, and Qualcomm claims a much bigger advantage in its not-yet-shipped chips, Intel says its roadmap includes next-gen processors to address every segment of the AI market.

"We have our product roadmap and plan on where we're at in mobile with premium and mainstream, but then you also go down into entry. And so we have plans around entry. From a desktop standpoint, we have plans on the desktop side, what we would say [is an] AI PC. And then there's also the next-gen AI PC, the 40 TOPS requirements; we have all of our different steps in our roadmap on how we cover all the different segments."

Intel's Lunar Lake processors will come to market later this year with three times more AI performance on both the GPU and the NPU than the existing Meteor Lake chips. Intel is already sampling those chips to its partners in preparation for a launch later in the year. Those chips will face off with Qualcomm's X Elite and AMD's next-gen processors. 

In the meantime, Intel is working to expand the number of AI features available on its silicon. As we covered in depth earlier today, the company plans to support 300 new AI-enabled features on its Meteor Lake processors this year. 

Many of those features will be optimized specifically for Intel's silicon. The company told me that roughly 65% of the developers it engages with use Intel's OpenVino, which means those applications are optimized specifically for Intel's NPU. The remainder of the developers use a 'mix of splits' between ONNX, DirectML, and WebNN, and Intel says it is happy to work with developers using any framework. 

However, the work with OpenVino could provide Intel with a nice runway of Intel-specific AI features as it heads into the Lunar Lake generation. Those are the types of advantages the company is obviously looking to enable through its AI PC Accelerator Program. The company says it has seen plenty of enthusiasm from the developer community, particularly in light of Intel's stated goal of selling 100 million AI PCs by 2025, which represents a big market opportunity for new AI software.

However, Microsoft's Copilot will run on NPUs from all vendors through DirectML, and having more TOPS will undoubtedly result in higher performance. That means we can expect a TOPS war to unfold, both in silicon and in marketing, over the coming years. 

Update 3/27/2024 6:50am PT: corrected Intel Meteor Lake and Ryzen Hawk Point NPU TOPS specifications. 

Intel announced two new extensions to its AI PC Acceleration Program in Taipei, Taiwan, with a new PC Developer Program that’s designed to attract smaller Independent Software Vendors (ISVs) and even individual developers, and an Independent Hardware Vendors (IHV) program that assists partners developing AI-centric hardware. Intel is also launching a new Core Ultra Meteor Lake NUC development kit, and introducing Microsoft’s new definition of just what constitutes an AI PC. We were given a glimpse of how AI PCs will deliver better battery life, higher performance, and new features.

Intel launched its AI Developer Program in October of last year, but it's kicking off its new programs here at a developer event in Taipei that includes hands-on lab time with the new dev kits. The program aims to arm developers with the tools needed to develop new AI applications and hardware, which we’ll cover more in-depth below.

Intel plans to deliver over 100 million PCs with AI accelerators by the end of 2025. The company is already engaging with 100+ AI ISVs for PC platforms and plans to have over 300 AI-accelerated applications in the market by the end of 2024. To further those efforts, Intel is planning a series of local developer events around the globe at key locations, like the recent summit it held in India. Intel plans to have up to ten more events this year as it works to build out the developer ecosystem.

The battle for control of the AI PC market will intensify over the coming years — Canalys predicts that 19% of PCs shipped in 2024 will be AI-capable, but that number will increase to 60% by 2027, highlighting a tremendous growth rate that isn’t lost on the big players in the industry. In fact, AMD recently held its own AI PC Innovation Summit in Beijing, China, to expand its own ecosystem. The battle for share in the expanding AI PC market begins with silicon that enables the features, but it ends with the developers that turn those capabilities into tangible software and hardware benefits for end users. Here’s how Intel is tackling the challenges.

What is an AI PC?

The advent of AI presents tremendous opportunities to introduce new hardware and software features to the tried-and-true PC platform, but the definition of an AI PC has been a bit elusive. Numerous companies, including Intel, AMD, Apple, and soon Qualcomm with its X Elite chips, have developed silicon with purpose-built AI accelerators residing on-chip alongside the standard CPU and GPU cores. However, each has its own take on what constitutes an AI PC.

Microsoft’s and Intel’s new co-developed definition states that an AI PC will come with a Neural Processing Unit (NPU), CPU, and GPU that support Microsoft’s Copilot and come with a physical Copilot key directly on the keyboard that replaces the second Windows key on the right side of the keyboard. Copilot is an AI chatbot powered by an LLM that is currently being rolled into newer versions of Windows 11. It is currently powered by cloud-based services, but the company reportedly plans to enable local processing to boost performance and responsiveness. This definition means that the existing Meteor Lake and Ryzen laptops that have shipped without a Copilot key actually don't meet Microsoft's official criteria, though we expect Microsoft's new definition to spur nearly universal adoption of the key.

While Intel and Microsoft are now promoting this jointly developed definition of an AI PC, Intel itself has a simpler definition that says it requires a CPU, GPU, and NPU, each with its own AI-specific acceleration capabilities. Intel envisions shuffling AI workloads between these three units based on the type of compute needed, with the NPU providing exceptional power efficiency for lower-intensity AI workloads like photo, audio, and video processing while delivering faster response times than cloud-based services, thus boosting battery life and performance while ensuring data privacy by keeping data on the local machine. This also frees the CPU and GPU for other tasks. The GPU and CPU will step in for heavier AI tasks, a must as having multiple AI models running concurrently could overwhelm the comparatively limited NPU. If needed, the NPU and GPU can even run an LLM in tandem.

AI models also have a voracious appetite for memory capacity and speed, with the former enabling larger, more accurate models while the latter delivers more performance. AI models come in all shapes and sizes, and Intel says that memory capacity will become a key challenge when running LLMs, with 16GB being required in some workloads, and even 32GB may be necessary depending on the types of models used.

Naturally, that could add quite a bit of cost, particularly in laptops, but Microsoft has stopped short of defining a minimum memory requirement yet. Naturally, it will continue to work through different configuration options with OEMs. The goalposts will be different for consumer-class hardware as opposed to workstations and enterprise gear, but we should expect to see more DRAM on entry-level AI PCs than the standard fare — we may finally bid adieu to 8GB laptops.

Intel says that AI will enable a host of new features, but many of the new use cases are undefined because we remain in the early days of AI adoption. Chatbots and personal assistants trained locally on users' data are a logical starting point, and Nvidia's Chat with RTX and AMD's chatbot alternative are already out there, but AI models running on the NPU can also make better use of the existing hardware and sensors present on the PC.

For instance, coupling gaze detection with power-saving features in OLED panels can enable lower refresh rates when acceptable, or switch the screen off when the user leaves the PC, thus saving battery life. Video conferencing also benefits from techniques like background segmentation, and moving that workload from the CPU to the NPU can save up to 2.5W. That doesn’t sound like much, but Intel says it can result in an extra hour of battery life in some cases.

Other uses include eye gaze correction, auto-framing, background blurring, background noise reduction, audio transcription, and meeting notes, some of which are being built to run on the NPU with direct support from companies like Zoom, Webex, and Google Meet, among others. Companies are already working on coding assistants that learn from your own code base, and others are developing Retrieval-Augmented Generation (RAG) LLMs that can be trained on the users’ data, which is then used as a database to answer search queries, thus providing more specific and accurate information.

Other workloads include image generation along with audio and video editing, such as the features being worked into the Adobe Creative Cloud software suite. Security is also a big focus, with AI-powered anti-phishing software already in the works, for instance. Intel’s own engineers have also developed a sign-language-to-text application that uses video detection to translate sign language, showing that there are many unthought-of applications that can deliver incredible benefits to users.

The Core Ultra Meteor Lake Dev Kit

The Intel dev kit consists of an ASUS NUC Pro 14 with a Core Ultra Meteor Lake processor, but Intel hasn’t shared the detailed specs yet. We do know that the systems will come in varying form factors. Every system will also come with a pre-loaded software stack, programming tools, compilers, and the drivers needed to get up and running.

Installed tools include Cmake, Python, and Open Vino, among others. Intel also supports ONNX, DirectML, and WebNN, with more coming. Intel’s OpenVino model zoo currently has over 280 open-source and optimized pre-trained models. It also has 173 for ONNX and 150 models on Hugging Face, and the most popular models have over 300,000 downloads per month.

Expanding the Ecosystem

Intel is already working with its top 100 ISVs, like Zoom, Adobe and Autodesk, to integrate AI acceleration into their applications. Now it wants to broaden its developer base to smaller software and hardware developers — even those who work independently.

To that end, Intel will provide developers with its new dev kit at the conferences it has scheduled around the globe, with the first round of developer kits being handed out here in Taipei. Intel will also make dev kits available for those who can’t attend the events, but it hasn’t yet started that part of the program due to varying restrictions in different countries and other logistical challenges.

These kits will be available at a subsidized cost, meaning Intel will provide a deep discount, but the company hasn’t shared details on pricing yet. There are also plans to give developers access to dev kits based on Intel’s future platforms.

Aside from providing hardware to the larger dev houses, Intel is also planning to seed dev kits to universities to engage with computer science departments. Intel has a knowledge center with training videos, documentation, collateral, and even sample code on its website to support the dev community.

Intel is engaging with Independent Hardware Vendors (IHVs) that will develop the next wave of devices for AI PCs. The company offers 24/7 access to Intel’s testing and process resources, along with early reference hardware, through its Open Labs initiative in the US, China, and Taiwan. Intel already has 100+ IHVs that have developed 200 components during the pilot phase.

ISVs and IHVs interested in joining Intel’s PC Acceleration Program can join via the webpage. We’re here at the event and will follow up with updates as needed.

Naver, owner of South Korea's largest search platform and various related services (some AI-powered), has decided to order its next $752 million in AI chips from Samsung rather than Nvidia. This news, as reported by KED Global, follows Naver's October 2023 pivot from Nvidia GPUs to Intel CPUs, and shows that Naver is doing all it can to reduce its reliance on Nvidia, the current leader in the AI hardware space.

Reportedly, Samsung's Mach-1 AI chips are going to be used with Naver's servers for its AI map service, Naver Place. Samsung's Mach-1 AI chips were revealed earlier this week to be set for an early 2025 launch, though this partnership suggests Naver will be getting them before anyone else. 

By supplying up to two million Mach-1 AI accelerators to Naver by the end of the year, Samsung could net up to a full $752 million. News of the Naver-Samsung deal has also caught the ear of other major tech companies, including the likes of Microsoft and Meta, according to KED Global.

By making these changes, it seems that Naver hopes to maximize its performance per dollar to remain competitive with its AI-powered map service. As long as the Samsung Mach-1 chips perform closely enough to the previous Nvidia H100 chips being used, their dramatically lower price ("a tenth of Nvidia's" says an unnamed industry source) and better power efficiency should make this a winning choice. Of course, it's probably also a good PR story for the nation and both companies to have South Korea's largest company providing the next-gen silicon for South Korea's largest search provider.

Previously, we've seen that even Intel CPUs can serve as a viable platform for certain AI workloads. Since Naver has already utilized Intel CPUs for its Naver Map software, it makes sense that the company isn't all-in on Nvidia. Samsung's Mach-1 is an AI accelerator with an SoC design that combines Samsung's own processors and LPDDR memory chips, which supposedly reduces bottlenecks and power consumption, according to industry sources mentioned in KED's reporting.

While this news seems promising for Samsung, only time will tell whether or not ITs Mach-1 AI chips are a valid replacement for Nvidia's H100 AI accelerators. In any case, it seems that big companies looking to deploy massive amounts of AI hardware are yearning for more alternatives to Nvidia's near-monopoly— and a recent ban on CUDA translation layers clearly isn't going to win over new fans for Nvidia.

A pair of SiSoftware benchmark listings have unveiled two new Intel GPUs sporting 10 and 12 Xe cores, with 12GB of VRAM on both results. It's possible these two GPUs are future Battlemage graphics cards, though the relatively low clock speeds and limited core counts leave room for debate. They could also be alternative Alchemist configurations, faked listings, or even potential integrated graphics solutions. But let's talk about what we do know right now.

Battlemage is the upcoming successor to Arc Alchemist that will be targeted at the dedicated graphics card market. It will also be utilized as a future integrated graphics solution for Intel's future CPU architectures (including Lunar Lake). Intel is rumored to be targeting a Q3 2024 release date, which would line up well with this being a potential Battlemage leak.

According to the SiSoftware listings, the specs for both GPUs are seemingly unimpressive. The least-potent offering sports just 1440 shader ALUs, while the more potent SKU has 1728 shader ALUs. To put that in perspective, the Arc A580 has 3456 shaders, while the Arc A380 slashes that figure down to 1024 shaders. But there are some other curiosities to discuss. Depending on how many shaders are in each Xe-core, we end up with either 10 and 12 Xe-cores, or potentially 20 and 24 Xe-cores. Which answer is correct matters quite a lot, as we'll discuss below.

Clock speeds are also unimpressively low at just 1.8GHz, which admittedly could be due to early silicon and driver tests. The only somewhat higher spec is that both have 12GB of VRAM, which is better than the 8GB found on the A580 and A750 midrange GPUs. However, those cards use 1GB chips with a 256-bit interface, so it's still possible that these are Alchemist, just configured with 2GB chips and a 192-bit interface.

And then there's the 8MB of L2 cache. That's a bit of a weird number for a 12GB card, as Intel's Alchemist GPUs have had either 4MB or 2MB per 64-bit controller. The A750 and A770 both have 16MB of L2 cache while the A580 has 8MB, but it's not clear how chips with presumably a 192-bit interface could arrive at 8MB total cache — 2.67MB per 64-bit interface just doesn't seem likely, and such a small amount of L2 also doesn't make much sense for the future Battlemage architecture. Of course, the L2 cache could be decoupled from the memory interfaces, and Intel might have L3 cache that's not showing up in these "leaked" results, but we don't know for certain.

Compared to Intel's existing Arc Alchemist GPUs, these two GPUs offer rather low shader core counts — more than an Arc A380 but far less than an Arc A580. Note also that we can't rule out the possibility of these being early mobile chips for testing and debugging purposes. The system used shows an Intel Z790 motherboard, but it's possible to test mobile GPUs with a desktop if these leaks came from a hardware lab.

In terms of positioning, these appear to be budget to midrange graphics solutions, given the presumed Xe-core counts. We're also accepting SiSoftware's reported data at face value, which lists a 160 CU and 192 CU "capacity" with 1440 and 1728 "shaders." Intel doesn't use CU or Compute Units for its GPUs, as that's AMD, while Nvidia uses "Streaming Multiprocessors," but SiSoftware uses the same CU label for all three. These Intel GPUs seem to be using "CU multiplied by nine" for the shader figures, with the base CU number being 16 times the Xe-core count — or if you want to be old school, it's the Execution Unit counts from older Intel GPUs.

Clear as mud? Let's give a more concrete example with a known GPU configuration. This Arc A380 listing says 128 CU and 1152S. The actual Arc A380 has eight Xe-cores with 1024 shader ALUs spread among sixteen Vector Engines (formerly called Execution Units) per Xe-core. There's also one additional "special purpose" (extended math) ALU per Vector Engine. So, 8 * 16 * 9 = 1152 shader ALUs. But the same calculations from SiSoftware don't seem to apply to the larger Arc A-series GPUs for whatever reason (they use Xe-cores * 16 * 8, without factoring in the extended math ALU). Here's how things break down:

For now, the performance data isn't terribly helpful given we're unsure on some of the specs and how high Battlemage might go. The core configuration and clock speeds suggest performance will be better than an A380 — and it is. It also seems as though Intel is getting far more performance out of fewer shaders with these unknown GPUs.

By way of comparison, the Arc A580 OpenCL result of 6,121Mpix/s is about the same as the unknown 10 (or 20?) Xe-core chip's 6,031Mpix/s, but the A580 has 24 Xe-cores. And here's an Arc A750 OpenCL getting 7,145Mpix/s while this Arc A770 OpenCL result is 8,229Mpix/s — with the supposed 12 (or 24?) Xe-core GPU getting 7,231Mpix/s. Assuming the math is correct on shader counts, the unknown chips are easily beating similarly configured Alchemist GPUs.

These "new" GPUs land at the level of the A580 and A750, but with far fewer shaders, plus a different memory configuration. If these are Battlemage and can clock 40% higher than the 1.8GHz shown here, and then if we get chips with up to 32 Xe-cores and 4096 shaders, we would end up with a decent level of performance. Three times as many Xe-cores as the presumed 10-core GPU plus higher clocks would potentially put such an Intel chip into RTX 4070 Ti territory. But let's not get too carried away, as the cache and other specs still look odd, and the Xe-core numbers might actually be double what we used.

We've learned from both AMD and Nvidia that a bigger cache on a GPU can greatly improve performance. Intel has learned this as well, and while it was probably too late to add a big cache to Alchemist, we definitely expect Battlemage to come with bigger caches. 8MB doesn't cut it, though it may simply be that Intel has L3 cache and SiSoftware isn't querying or reporting that value.

Ultimately, it's too early to tell how fast Battlemage will be, but these supposed spec leaks look at least somewhat promising. If a 12 Xe-core chip at 1.8 GHz can already match the A750, and if Intel goes as high as 32 Xe-cores with Battlemage, we might actually be looking at reasonably competitive Intel solutions that could maybe take on the high-end AMD and Nvidia solutions. Maybe.

Alternatively, that's a 24 Xe-core Battlemage chip running at 1.8 GHz and matching a 2.4 GHz 28 Xe-core Alchemist chip. We already know where the A750 lands in our GPU benchmarks hierarchy, but while boosting clocks and performance by 50% would certainly help, that would still only result in something on the level of the RTX 3070, and slower than an RTX 4060 Ti. Let's hope this isn't the case.

Intel is already two years behind the competition, and with Intel confirming that Battlemage hardware development is already done and 70% of the team has moved on to Celestial — its GPU architecture after Battlemage — we're hoping these leaked results are for budget-oriented future GPUs and that substantially larger and faster variants are on the way. In the interim, the remaining 30% of Intel's graphics team is working on software and driver support for Battlemage, in preparation for its launch later this year.

Broadcom has demonstrated that it is perhaps the world's largest processor. But for what application? When we visited TSMC's events, we were always shown a deck of multi-chiplet processors that use the company's chip-on-wafer-on-substrate (CoWoS) packaging technology and feature near the reticle limit (858mm^2, 26 mm by 33 mm) compute chiplets. We cannot take photos of the deck, but there are certainly processors that grab attention. One of those devices comes from Broadcom, and it has been shown at the company's recent investor events.

For most observers, Broadcom is a networking and telecommunications giant, but the company also has a significant custom chip design business. For those unfamiliar with this unit of Broadcom, Google is one of the company's most prominent clients in terms of contract chip design.

However, just like TSMC, Broadcom does not announce its clients. For those who want to rekindle its short-term innovations, Broadcom has a list of them in its recent press release. What it does to impress is demonstrate its vast accomplishments to its investors. These are indeed vast, as observed by our friend and colleague Patrick Moorhead of Moor's strategies market analysis company.

"Here is another fun one," Patrick Moorhead wrote in an X post. "The guy who is smiling Frank Ostojic [who] runs Broadcom's custom silicon group. He should be smiling as he announced that he has a third XPU design from a large 'consumer AI company.'

Broadcom officially brands those chips as XPUs so as not to disclose their applications. Meanwhile, the use of high-bandwidth memory pretty much shows its target usage, which might well be artificial intelligence or hardcore AI-infused network switching.

"To the right is a close up of the XPU," Moorhead added. "You can see the two compute units on the center and all the HBM to the left and right. A full up custom SoC with lots and lots of compute, HBM, very high speed intra chip connectivity and, as you would expect, the highest performance external networking."

Developing a chiplet of this scale (i.e., near the reticle size) is already an achievement. Yielding it to a proper level is another dimension of achievement, and it looks like Broadcom's foundry partner, most likely TSMC, has accomplished it as well. Now, it is time for software to catch up and use this processor's might.

Samsung Electronics is gearing up to launch its own AI accelerator chip — the Mach-1 — in early 2025, the company announced at its 55th annual shareholders meeting, reports Sedaily. The move will mean competing against companies like Nvidia, but the South Korean company has no plans (at least for now) to rival Nvidia's ultra-high-end AI solutions, such as the H100, B100, or B200. 

Samsung's Mach-1 is an AI inference accelerator based on an application specific integrated circuit (ASIC) and equipped with LPDDR memory, which makes it particularly suitable for edge computing applications. The report claims that Samsung's Mach-1 has a unique feature that will dramatically reduce memory bandwidth requirements for inference to about 0.125x when compared to existing designs — but does not elaborate on how it will do this, so take this claim with a grain of salt. 

Kye Hyun Kyung, the head of Samsung Electronics Device Solutions Division, overseeing global operations of the Memory, System LSI, and Foundry business units, said that the chip design had passed technological validation on field-programmable gate arrays (FPGAs) and that the physical finalization of the system-on-chip (SoC) is currently underway. Kyung assured that the chip would be ready by the end of the year, allowing for the launch of an AI system powered by the Mach-1 chip early next year. 

Samsung's Mach-1 is not designed to compete against high-performance AI processors such as the AWS Trainium or the Nvidia H100, but it will certainly compete against other inference-oriented solutions, including AWS Inferentia. It is highly likely that Samsung will position its Mach-1 primarily for edge applications that require low power consumption, minimal dimensions, and low costs, which is where it can probably excel. 

Samsung is not only focused on developing the Mach-1 chip but is also actively involved in broader AI semiconductor development efforts, according to Kye Hyun Kyung. The company has established a lab in Silicon Valley dedicated to general artificial intelligence (AGI) research, signaling its commitment to becoming a key player in the future of AI technology. The lab's mission is to create new types of processors and memory that meet processing requirements of future AGI systems.

When Nvidia unveiled its first GPUs based on the Blackwell architecture for AI and HPC earlier this week, it disclosed some of its specifications and performance numbers. But what it did not reveal is when, exactly, it plans to ship these GPUs to clients. On Tuesday the company confirmed that the first Blackwell processors will ship in 2024, but even Meta — one of Nvidia's largest customers — does not expect to get Blackwell this year. 

Nvidia's chief financial officer Colette Kress told financial analysts Tuesday that the company expects its Blackwell-based GPUs 'to come to market later this year,' but admitted that shipments of these processors will only ramp up to significant volumes in 2025, reports Reuters. It looks like the bulk of Nvidia's AI and HPC GPU shipments this year will be based on the Hopper microarchitecture. 

Meta, the owner of Facebook and Instagram, and one of the largest consumers of Nvidia's AI processors, told Reuters that it did not expect to receive shipments of Nvidia's new flagship GPU this year. 

Earlier this year Mark Zuckerberg, CEO of Meta, said that his company would purchase 350,000 H100 GPUs this year, in addition to the 150,000 of these processors that were procured in 2023. In total, Zuckerberg expects to have performance equivalent to that of 600,000 H100 processors by the end of the year. Given the vast performance requirements of Meta, the company would absolutely like to get as many Blackwell B200 processors as possible, but apparently the bulk of the company's purchases from Nvidia will be based on the Hopper architecture. 

It's noteworthy that, starting from its Ampere-based products in 2020, Nvidia has been offering not only GPUs but actual servers. With Hopper, the company offered processors, servers, and even supercomputers based on H100. With Blackwell, the company has introduced its DGX B200 servers and DGX B200 SuperPODs and while it is unlikely that it will not sell B200 processors to traditional server makers or clients like Google, Meta, and Microsoft, Nvidia may start to put even more emphasis on selling machines rather than GPUs themselves.