China's LineShine supercomputer has taken the top spot on the 67th-edition TOP500 list, posting 2.198 exaflops on the High Performance Linpack benchmark and pushing the AMD-powered El Capitan into second place by more than 20%. The system, installed at the National Supercomputing Centre in Shenzhen (NSCS) and built by the Shenzhen Cloud Computing Center, used no GPUs or accelerators of any kind, and reached the figure with 13,789,440 cores of domestically designed silicon, the first machine on the list to clear two exaflops of double-precision performance on CPUs alone. It’s also the first China-based system to lead the TOP500 since Sunway TaihuLight in 2017.

The fact that a sanctioned country has managed to build an exascale flagship without a single Western accelerator is one thing, but what’s more telling is that China has decided to put it on the list. For years, its fastest machines have stayed off the rankings entirely, and the decision to submit a chart-topper now is a deliberate change of posture.

A domestic stack from core to OS

LineShine is built on what NSCS calls the LingKun platform. Each of its 20,480 compute nodes carries two LX2 processors, Armv9-based parts with 304 cores running at 1.55 GHz, organized as eight clusters of 38 cores. Every core includes Arm's Scalable Vector Extension and Scalable Matrix Extension units covering FP64, FP32, BF16, FP16, and INT8.

Each of those LX2s pairs 32 GB of on-package HBM rated at up to 4 TB/s with as much as 256 GB of off-package DDR5, an arrangement that’s closer to Fujitsu's A64FX in Japan's Fugaku than to a conventional server CPU. Nodes are tied together by the proprietary LingQi interconnect, and the machine runs the homegrown Kylin OS.

It’s not known who designs the LX2 — NSCS names no vendor — but Jon Peddie Research has attributed the chip to Huawei, and the project's pilot phase reportedly ran on Huawei Kunpeng servers. The fabrication node and foundry are likewise unconfirmed. SMIC's 7nm-class process is the obvious domestic candidate by elimination, given that EUV tooling and TSMC capacity are both off the table, but nobody has documented the part to date.

Not an AI crown

LineShine also took first on HPCG, the test that rewards memory- and communication-bound workloads closer to real scientific code, at 22.00 petaflops. But on HPL-MxP, the mixed-precision benchmark that approximates AI training math, it came in only fourth at 7.92 exaflops, a 3.6 times uplift over its FP64 score.

In other words, the accelerator-based machines it beat on Linpack pull far ahead the moment precision drops. Per the TOP500 announcement, El Capitan posts 16.7 exaflops on HPL-MxP, a 9.2 times jump over its standard result, with Aurora and Frontier showing similar multipliers. Reduced-precision throughput is exactly where GPUs and APUs separate from CPUs, and LineShine has nowhere to hide it.

We can see similar issues cropping up in terms of power. LineShine draws 42,220 kW and returns 52.07 gigaflops per watt on its Linpack run. That beats Intel’s Aurora comfortably but trails El Capitan's 60.94 gigaflops per watt, so LineShine produces more total FP64 output than the Livermore system while burning roughly 42% more power to do it.

It’s worth holding onto this distinction because the TOP500 ranking is decided on FP64 Linpack, the one regime where a wide, HBM-fed CPU can still go toe-to-toe with accelerators. LineShine is a genuine double-precision champion, but it’s not a world-leading AI training machine, and its fourth-place HPL-MxP result says so.

So, why did China submit it?

China stopped submitting its fastest systems to the TOP500 around 2021, after a run of entity-list additions hit Sunway's Wuxi center and Sugon. The community has long believed that the country operated exascale hardware well before this entry: the Sunway successor OceanLight and the NUDT-built Tianhe-3 both appeared via Gordon Bell Prize science papers without ever appearing on the list. TOP500 co-founder Jack Dongarra has said for years that Chinese researchers told him they weren’t permitted to submit, and that omissions were about avoiding U.S. attention rather than any lack of capability.

Last June's list, which AMD topped while Chinese HPC remained absent, was especially conspicuous, but putting LineShine forward now reverses that. It has been reported that the system was developed without public funding, which lowers the political exposure of disclosing it, and the all-domestic design means there’s no dependency on Western parts for Washington to choke off after the fact.

Addison Snell, chief executive of HPC analyst firm Intersect360 Research, told Reuters he wasn’t surprised by the performance but by the disclosure itself, noting the surprise was that China submitted the result and wanted recognition for it. Ultimately, submitting a number-one system that runs entirely on indigenous parts is a statement that the sanctions regime hasn’t closed the gap China cares about.

AMD still dominates

The top of the list might have changed hands, but the bulk of it hasn’t. The U.S. still dominates with three of the top five in El Capitan (1.809 exaflops), Frontier (1.353 exaflops), and Aurora (1.012 exaflops), and Germany's JUPITER Booster remains the first and only European exascale system at an even 1.000 exaflops.

AMD’s silicon underpins most of the accelerated field with the company, per its own blog, now powering 191 systems on the list, up 11% year over year, and 41% of this edition's new entries. It holds three top-10 slots — El Capitan, Frontier, and the newly deployed HPC7 at Italian energy firm Eni — and contributes more than 40% of combined top-10 Linpack performance. On efficiency, it powers 56% of the top 50 Green500 systems, and its first Instinct MI355X deployments, two Cambridge Zenith systems in the UK, entered at positions 67 and 68.

None of that is dented by LineShine, not least because the two aren’t competing for the same workload. AMD’s MI300A and MI355X parts are built for mixed-precision AI arithmetic, where LineShine places fourth, and the rest of the Western labs are optimizing for that, not FP64 leaderboard positions.

El Capitan, Frontier, and Aurora all post HPL-MxP scores several times their Linpack results, enabled by hardware that LineShine doesn’t have. So, while it’s true the TOP500 crown moved to Shenzen, it did so on a benchmark that Western labs are no longer chasing with their fastest machines.

China's LineShine supercomputer has dethroned El Capitan as the world's number one supercomputer, going straight to the top of the charts after the National Supercomputer Center in Shenzhen (NSCS) submitted its results.

LineShine hit 2.198 FP64 ExaFLOPS in the Linpack benchmark and became the industry's first machine in the Top 500 list to sustain more than 2 ExaFLOPS of double-precision performance using only CPUs. The system is deployed at the National Supercomputing Centre in Shenzhen and was built by the Shenzhen Cloud Computing Center using semi-custom 304-core LX2 processors based on the Armv9 instruction set architecture and running at 1.55 GHz. The machine employs 13.79 million cores in total, uses proprietary LingQi interconnect, and consumes 42.2 MW of power.

From a performance-per-watt point of view, the LineShine machine delivers 52.07 GFLOPS/W, which is below El Capitan's 60.94 GFLOPS/W. However, LineShine by far outperforms Fugaku — another CPU-only supercomputer that used to be the No.1 HPC system several years ago — that can only deliver 14.78 – 16.84 GFLOPS/W depending on whether its efficiency is optimized or not.

LineShine also moved to the top of the HPCG ranking with 22.00 HPCG-PFLOPS. However, the supercomputer achieved 7.92 mixed-precision EFLOPS in HPL-MxP, which puts it behind El Capitan, Frontier, and Aurora. This limits LineShine's usability for AI training and inference, but this can be justified with its exceptional performance for traditional supercomputer tasks.

Each LX2 CPU relies on two compute chiplets and has a total of 304 CPU cores organized into eight CPU clusters containing 38 cores each. Every core includes Arm SVE (Scalable Vector Extension) and SME (Scalable Matrix Extension) units that accelerate vector and matrix operations used in AI training and scientific computing that support FP64, FP32, BF16, FP16, and INT8 data formats. The chip features a rather unusual memory architecture that pairs 32 GB of on-package HBM, offering up to 4 TB/s of bandwidth with as much as 256 GB of external DDR5 memory to maximize both bandwidth and capacity.

Despite this, the processor only gains 3.6X performance when moving from FP64 to mixed-precision data, which is lower compared to systems that integrate low-precision accelerators, such as AMD's Instinct MI300A or Intel's Ponte Vecchio. While an Armv9 CPU with SVE/SME can accelerate FP16/BF16/INT8 workloads, its mixed-precision uplift remains limited compared to systems with accelerators due to many reasons, including memory bandwidth, software maturity, and interconnect efficiency. That said, it may be too early to make final conclusions about the LX2 and its usability for mixed-precision workloads.

In any case, the very fact that a Chinese supercomputer has achieved extraordinary FP64 performance is remarkable. Furthermore, the fact that NSCS has actually submitted results to Top 500 indicates that the organization is confident that the LineShine supercomputer relies exclusively on domestic technologies and the U.S. government cannot affect the production of these technologies.

A historically significant supercomputer has gone up for auction. The first Cray T3D ever produced, serial number 6001, was recently listed by The Saleroom with opening bids from £60,000 (~$81,000) invited. This T3D has quite some provenance, having been Cray’s internal development machine and then installed at Edinburgh University, earning the moniker the Typhoon. Moreover, the TOP500 list ranked the Typhoon as the fastest supercomputer in Europe in June 1996.

The Cray TD3 marked a significant shift in the supercomputer pioneer’s performance strategy. “As the inaugural machine of the T3D series, it represents a defining step in Cray’s move from traditional vector systems into the era of massively parallel supercomputing,” explain the auction notes. For enthusiasts and collectors it also “stands as a museum-grade survival of exceptional importance.”

Inside the stylish H193cm x W117cm x D193cm (over 6ft tall) ‘Tomato Red’ chassis pictured is a single-cabinet Cray T3D-MC512, equipped with 512 DEC Alpha 21064 150 MHz compute processors, according to the listing. Cooling this beastly throng is Fluorinert liquid-based. This auction lot includes the main cabinet and HEU first-stage cooling system (also over 6ft tall and weighing 0.85 tons).

The reserve / starting price of ~$81,000 is a big ask for this type of machine, as its size and weight will limit the appeal among enthusiasts. We note that, at the time of writing, there are no bids yet and just 10 watchers (auction ends May 31). The Saleroom appears to be framing the Typhoon as a bargain, though, mentioning examples like this would have originally cost $15M when new.

Two other Cray supercomputers are being auctioned at the same time as this TD3. On the listing source page you will also see that auctions for a Cray Triton T-932 Supercomputer and a Cray Y-MP4E Supercomputer are ongoing and scheduled to end on May 31.

Elon Musk has confirmed on X that Tesla has restarted work on the Dojo3 supercomputer following the new success of its AI5 chip design. The billionaire stated in a recent X post that the AI5 chip design is now in "good shape", enabling Tesla to shuffle resources back to the Dojo 3 project. Musk also added that he is hiring more people to help build the chips that will inevitably be used in Tesla's next-gen supercomputer.

This news follows Tesla's decision that it was cancelling Dojo's wafer-level processor initiative in late 2025. Dojo 3 has gone through several iterations since Elon Musk first chimed in on the project, but according to Musk's latest thoughts on it, Dojo 3 will be the first Tesla-built supercomputer to take advantage of purely in-house hardware only. Previous iterations, such as Dojo2, took advantage of a mixture of in-house chips and Nvidia AI GPUs.

According to Musk, the Dojo3 will use AI5/AI6 or AI7, the latter two being part of Musk's new 9-month cadence roadmap. AI5 is AI5 is almost ready for deployment and is Tesla's most competitive chip yet, yielding Hopper-class performance on a single chip and Blackwell-class performance with two chips working together using "much less power". Work on Dojo 3 coincides directly with Musk's new nine-month release cycle, where Tesla will start producing new chips every nine months, starting with its AI6 chip. AI7, we believe, will likely be an iterative upgrade to AI6; building a brand new architecture every 9 months would be extremely difficult, if not impossible.

It will be interesting to see whether or not Dojo3 will prove to be successful. Dojo 1 was supposed to be one of the most powerful supercomputers when it was built, but competition from Nvidia prevented that from happening, among other problems. Dojo 2 was cancelled mid-way through development. If Tesla can deliver competitive performance with Nvidia GPUs consistently, Dojo 3 has the potential to be Tesla's first truly successful supercomputer. Elon also hinted that Dojo 3 will be used for "space-based AI compute".

AMD and Eviden, an Atos Group-controller maker of supercomputers, have announced their first project — the Alice Recoque supercomputer — that will use AMD's next-generation EPYC 'Venice' CPUs and Instinct MI430X accelerators to achieve performance beyond 1 ExaFLOPS. The machine will cost over half a billion euros and will be accessible to researchers working on demanding AI and scientific projects.

The Alice Recoque supercomputer will be based on AMD's upcoming EPYC 'Venice' processors with up to 256 cores; the company's next-generation Instinct MI430X accelerators based on the CDNA 5 architecture and equipped with 432 GB of HBM4 memory that is tailored for HPC, but also supports FP4 and FP8 data formats to make them useful for AI; and AMD's FPGAs to give the system more flexibility and versatility.

The full installation of Alice Recoque will consist of 94 racks based on Eviden's BullSequana XH3500 platform with DDN storage and will use the company’s BXI fabric to link all compute resources and enable high-performance scale-out connectivity. The system's 5th-Generation warm-water liquid cooling for all power-hungry components.

Interestingly, the Alice Recoque supercomputer will also pack sovereign Rhea2 CPUs (note that this is the first time we hear about Rhea2, as previously we only heard about Rhea or Rhea1) for users who prefer to perform their computations on processors developed in Europe. Keeping in mind that Rhea2 has not taped out yet, we can only wonder when these racks will be installed, but perhaps sometime towards the end of the decade.

Alice Recoque will cost €554 billion, which will come from EuroHPC JU, the Digital Europe Programme, and the Jules Verne consortium, which includes France (GENCI, CEA), the Netherlands (SURF), and Greece (GRNET). The machine will be installed in France under the responsibility of GENCI, which will host the system, while CEA will operate it once deployed. The system will ultimately serve researchers and industrial users across Europe and will support a variety of projects, including climate research, materials and energy science, personalized medicine, AI model development, and analysis of massive data streams from satellites, telescopes, and IoT sources.

The press release does not specify the commissioning or operational date of Alice Recoque, but given the fact that AMD is set to introduce its EPYC 'Venice' CPUs and Instinct MI430X accelerators in 2026, it is reasonable to expect the machine to enter service in 2027 or 2028.

Coming on the heels of the Vera Rubin-based supercomputers for Los Alamos National Laboratory, Nvidia announced on Tuesday at GTC 2025 that, together with partners, it would build seven ExaFLOPS-class AI supercomputers for Argonne National Laboratory. Two out of five systems will be built by Oracle and will use over 100,000 Blackwell GPUs, delivering a combined performance of up to 2,200 ExaFLOPS.

The first of five AI supercomputers for Argonne National Laboratory is Equinox, which will pack 10,000 Blackwell GPUs and serve as the first phase of the project, coming online in 2026. The second phase of the project — called Solstice — will be a 200 MW system packing over 100,000 Blackwell GPUs. The two systems will be connected to deliver an aggregate performance of 2,200 FP4 ExaFLOPS for AI computations.

"We are proud to announce that Nvidia, the U.S. Department of Energy and Oracle are partnering to build two AI factories at Argonne National Laboratories featuring Blackwell," said Dion Harris, the head of data center product marketing at Nvidia. "This collaboration aims to significantly boost America's scientific research and development productivity and establish U.S. leadership in AI. Phase one features the Equinox system, which is 10,000 Blackwell GPUs; phase two, providing 200 MW of AI infrastructure, totaling 2,200 ExaFLOPS of AI performance."

The systems will be used to build three-trillion-parameter AI simulation models as well as for classic scientific computing.

One interesting thing to note about the Equinox and Solstice supercomputers is that they will be built by Oracle, a company that nowadays is not widely known as a vendor that designs and builds completely bespoke supercomputers for customers, as traditional HPC vendors like Atos, Dell, or HPE do. Oracle's primary business emphasis is on cloud infrastructure enabling AI/HPC workloads rather than custom HPC system integration from the ground up. While Oracle has its Cloud@Customer option, these machines also run Oracle's software and are managed by the company. Whether Equinox and Solstice will be managed by Oracle remains to be seen.

In addition, the Argonne National Laboratory will expand its Argonne Leadership Computing Facility — which will be available to researchers and scientists through competitive national programs — with Nvidia-based supercomputers, including Tara, Minerva, and Janus. For now, it is unclear which platform these systems will use or whether they will be built by HPE or Oracle, but we can be sure they will deliver formidable performance.

"Argonne's collaboration with Nvidia and Oracle represents a pivotal step in advancing the nation's AI and computing infrastructure," said Paul K. Kearns, director of Argonne National Laboratory. "Through this partnership, we are building platforms that redefine performance, scalability and scientific potential. Together, we are shaping the foundation for the next generation of computing that will power discovery for decades to come."

Nvidia announced at GTC that it would team up with HPE to build two new supercomputers based on its Vera Rubin platform for Los Alamos National Laboratory. The machine will be used for national security and scientific research using AI simulation and scientific computing, thus following the latest trends in supercomputing. In all, Nvidia said it has won seven supercomputer contracts with the Department of Energy (DoE).

Nvidia's announcement comes on the heels of AMD's announcement yesterday of two new supercomputer wins with the Department of Energy.

Los Alamos National Laboratory has contracted with HPE to build its Mission and Vision supercomputers based on Nvidia's Vera Rubin platform, which comprises the company's next-generation Vera CPUs and Rubin GPUs. The machines will scale up using NVLink Gen6 technology and scale out using Nvidia's QuantumX 800 Infiniband networking.

The Mission supercomputer — designed for the National Nuclear Security Administration and therefore used to ensure the safety, reliability, and performance of the American nuclear stockpile without conducting live nuclear tests — is scheduled to go online in 2027. The Vision computer will build on the achievements of the earlier Venado supercomputer and serve open scientific and AI research.

"Mission is the fifth advanced technology system in Los Alamos' AI for national security science program," said Dion Harris, the head of data center product marketing at Nvidia. "It is expected to be operational in 2027 and designed to run classified applications. The Vision system builds on the achievements of the LANL Venado supercomputer and is designed for unclassified AI and open science research. The Mission and Vision systems represent a significant investment in the US national security and open science capabilities."

Nvidia did not disclose the expected performance of Mission and Vision. However, Vision is projected to take up the baton from Venado, the world's 19th-fastest supercomputer with an Rmax FP64 performance of 98.51 PFLOPS, so it is reasonable to expect Vision to offer at least twice the compute throughput for scientific computing. The comparison also implies that Nvidia's Vera Rubin platform will not sacrifice HPC-oriented FP64 performance for AI-oriented low-precision performance.

" We will share more details on the specific configurations [of supercomputers] later," said Harris. "The great thing about this [platform], looking at how these systems will be used for both open science and for national security research, we think it will bring both AI capabilities as well as traditional simulation capabilities to, to the scientific research endeavors."

Tens of billions seem to fly in all sorts of directions these days in the AI world. A "mere" $1 billion deal would arguably not even register, but the one announced today is arguably far more important than all the data centers for enabling AI chatbot sexting. According to a Reuters report, the U.S. Department of Energy (DoE) and AMD have announced a partnership for building two supercomputers at the Oak Ridge National Laboratory (ORNL) as a foundation for future nuclear fusion and medical research.

The partnership has the DoE and ORNL on the blue corner, and AMD, HPE, and Oracle on the other. The deal is that ORNL will host the datacenters, thus presumably providing the energy to run them, and the private companies will foot the bill for the hardware and software. When built, both sides will share the computing power.

The supercomputers themselves will predictably be an all-AMD affair for the major bits of hardware. The first one is called Lux and is set to be functional within six months, with AMD Instinct MI355X accelerators, to the tune of 1400 W board power each. ORNL director Stephen Streiffer says Lux will be three times as powerful in AI over current supercomputers, while Lisa Su states it was the fastest deployment of this size of supercomputer.

The second supercomputer will be called Discovery and is scheduled for delivery in 2028 and an operational kick-off in 2029. Discovery will use AMD's upcoming Instinct MI430 parts, a design with one Epyc CPU and four MI430X-HPC dies. The 430X and 450X are variations of the same design, with the former focusing on high-precision FP32 and FP64 performance, while the latter goes in the exact opposite direction and bets all its chips on FP8 and FP16.

Energy Secretary Chris Wright says this project will "supercharge" research on multiple fronts and tackle "large scientific problems ranging from nuclear power to cancer treatments to national security". He seems particularly bullish on fusion energy, stating he believes that with the help of these systems, [the U.S.] will have "practical pathways to harness fusion energy in the next two or three years." He also hopes that cancer will become a manageable disease in a timeframe of five to eight years.

Normally, simulations for scientific research are performed on supercomputers as they require tremendous compute throughput. There are also types of research — such as simulation of quantum behavior of molecules with exponentially more interacting states — that require quantum computers to simulate them, or simplifications to make the task suitable for modern supercomputers. However, Chinese scientists from Sunway have successfully used an AI model and an existing Oceanlite supercomputer to model complex quantum chemistry at the scale of real molecules, which is both a scientific and technological breakthrough, reports VastData.

A quantum state in quantum mechanics — described by a wavefunction (Ψ) — determines all possible configurations of a quantum system, such as the positions, spins, or energy levels of particles like electrons in a molecule, along with their probabilities. Modeling it is challenging because the state space grows exponentially with the number of particles, making it impossible (and not feasible) to simulate on classic supercomputers that we use today. To that end, scientists use a variety of approximation methods to simplify the quantum equations while preserving accuracy to describe molecular structures, reactions, and energies. However, the scaling of existing methods that approximate the wavefunction is limited to small molecules.

To study many-body quantum systems with strong electron correlations (e.g., tens of electrons, over 100 spin orbitals, etc.), several years ago, physicists proposed using modern machine-learning surrogates, such as neural-network quantum states (NNQS), to approximate all the possible configurations and motions of electrons within a molecule. This method promises to wed AI scalability with quantum accuracy for research that is currently impossible using traditional methods.

To conduct their experiment with 120 spin orbitals, researchers developed their own NNQS framework. Their simulation trained a neural network to approximate the molecule's wavefunction, determining where electrons are most likely to be. For each sampled electron arrangement, the system calculated the local energy and adjusted the network until its predictions matched the molecule's true quantum energy pattern.

The proprietary NNQS that was tailored for China's Oceanlite supercomputer, based on a 384-core Sunway SW26010-Pro CPU, which supports FP16, FP32, and FP64 data formats and features a very unique architecture tailored for HPC rather than for AI. In particular, they had to keep in mind how the SW26010-Pro parallelizes workloads and handles data.

The researchers have designed a hierarchical communication model where management cores handled coordination between processors and nodes, while millions of 'lightweight' 2-wide compute processing elements (CPEs) featuring a 512-bit vector engine performed local quantum calculations. In addition, they created a dynamic load-balancing algorithm to ensure that uneven computational loads did not leave any cores idle.

The scientists ran their code on 37 million CPE cores and achieved 92% strong scaling and 98% weak scaling, a high level of efficiency for such a scale, which highlights that the developers have found a near-perfect synchronization between software and hardware, a major accomplishment for China's supercomputer community. Also, to date, the simulation of molecular systems with 120 spin orbitals is the largest AI-driven quantum chemistry calculation ever performed on a classical supercomputer, marking a breakthrough for PRC's AI and quantum industries.

Without any doubt, the achievement demonstrated that NNQS can be used for quantum physics research on modern supercomputers. However, it is impossible to find out whether it is efficient to use an exascale supercomputer like Oceanlite for AI-based quantum physics research, both in terms of effort and power.

Japan is preparing to reclaim its position at the forefront of high-performance computing with FugakuNEXT, a hybrid AI-HPC supercomputer designed to surpass exascale and venture into uncharted zetta-scale territory. Announced at a ceremony in Tokyo on August 22, the project marks a major collaboration between RIKEN, Fujitsu, and, for the first time, Nvidia, bringing together Japan’s sovereign infrastructure and U.S. GPU technology in a system built for the next decade of scientific and industrial challenges.

Fugaku, installed in 2020, held the world’s top spot for two years and played a crucial role during the COVID-19 pandemic, where its simulations helped guide research and response efforts. Today, it still ranks seventh globally, but Japan is already looking ahead. FugakuNEXT is expected to come online around 2030 at RIKEN’s Kobe campus on Port Island, with a development budget exceeding 110 billion Yen ($740 million USD). This marks the latest in a lineage of Japanese flagship supercomputers—from the Earth Simulator in 2002 to the K computer in 2011 and Fugaku itself—each of which once held the crown for computational speed.

Unlike its predecessors, FugakuNEXT will integrate GPUs as a central design element. Fujitsu will develop the new MONAKA-X CPUs, while Nvidia will supply GPUs and co-design the interconnect fabric through NVLink Fusion, a high-bandwidth bridge between CPU and GPU. The result will be a system purpose-built for both traditional simulation workloads and next-generation AI applications. RIKEN has emphasized that this is Japan’s first flagship system to adopt GPUs, reflecting the growing global trend of unifying HPC and AI computing.

RIKEN has set quite ambitious goals for FugakuNEXT, projecting a fivefold increase in hardware performance over the original Fugaku, combined with a twentyfold software and algorithmic improvements through optimizations such as mixed-precision computing and physics-informed neural networks. Together, these advances are expected to deliver a massive hundredfold increase in real-world application performance compared to today’s machines.

Early targets already place the system’s peak at around 600 exaFLOPS in FP8 sparse precision, with the potential to become the world’s first zetta-scale supercomputer. Crucially, RIKEN has also confirmed that this leap in performance will be achieved within the same forty-megawatt power envelope as Fugaku, underscoring energy efficiency as a core principle of the project.

The institution also frames FugakuNEXT not just as a technical upgrade but as an integrated AI-HPC platform and a model for “AI for Science.” By automating various aspects of the research process—from hypothesis generation to experiment simulation and validation—the machine aims to accelerate discovery across diverse fields, including climate modeling, drug development, disaster resilience, and advanced manufacturing.

More importantly, Nvidia will serve a crucial role as its full software stack will be integrated into the supercomputer, from CUDA-X libraries for quantum simulation and data science to TensorRT for inference and NeMo for large language model training, reflecting the growing role of AI as a scientific instrument.

In today’s Linux patch series, AMD engineer Sameul Zhang highlighted an unusual issue where Linux servers are failing to hibernate due to excessive VRAM and a high number of AMD Instinct accelerators per system. For context, Instinct accelerators are powerful AMD GPUs designed specifically for data centers handling AI, high-performance computing, scientific workloads, and other demanding tasks.

Part of what makes these GPUs so powerful is that they come with massive amounts of VRAM, like 192GB in some, which might sound huge to gamers but is fairly standard for modern data center chips. In fact, this AMD AI Linux-powered server is equipped with a total of eight Instinct cards that bring the total VRAM to around 1.5TB. However, while more VRAM is generally a good thing, in cases like this, it can lead to unexpected issues.

But while VRAM capacity does play a part, the root cause of the hibernation failure isn’t the number of Instinct cards, but rather how Linux handles GPU memory during the hibernation process. When the system initiates hibernation, all GPU memory is first offloaded to system RAM, typically through the Graphics Translation Table (GTT) or shared memory (shmem). From there, the kernel creates a hibernation image by copying all system memory content, which also includes the evicted VRAM, into a second memory region before writing it to disk.

Sounds confusing? Well, in simple terms, if your server has 1.5TB of total VRAM, this duplication can push the memory usage up to 3TB, which easily exceeds the capacity of servers equipped with only 2TB of system memory. The spill-out ultimately causes the hibernation process to fail.

Fortunately, Zhang has been working to address this hibernation issue and suggests two main changes. The first is aimed at reducing the amount of system memory needed during hibernation, which would allow the process to succeed. However, doing so introduces a new issue, as the "thawing" stage (when the system resumes from hibernation) could take nearly an hour due to the large amount of memory. To fix this, a third patch was added to skip restoring these buffer objects during the thaw stage, significantly reducing the resume time.

Now, most high-end AI servers run continuously, so it's fair to ask why anyone would hibernate them. One common reason is to reduce power consumption during downtimes and help stabilize the electrical grid. Since large-scale data centers consume massive amounts of power, this can help lower the risk of blackouts, like the one we recently saw in Spain.

Sandia National Laboratories has allegedly activated its own SpiNNaker 2 server, a brain-inspired supercomputer comprising thousands of ARM-based CPU cores that don't require SSDs, hard drives, or GPUs. Blocks and Files reports the new supercomputer will rank among the top five brain-inspired platforms worldwide.

Sandia is reportedly utilizing a 24-board version of the SpiNNaker 2 architecture, featuring a total of 175,000 cores. Each board reportedly carries 96GB of LPDDR4 memory, giving the system 2.3TB of memory and 23GB of integrated on-chip SRAM storage.

SpiNNaker 2, or Spiking Neural Network Architecture 2, is a computer architecture inspired by the human brain — mimicking the actions of neurons and synapses (and is a form of Neuromorphic computing). The architecture relies on dozens of motherboards, each featuring nearly 50 ARM CPUs, all communicating via a vast web of interconnects. This design choice prioritizes boosting computational efficiency over maximizing raw performance through brute force methods, which is common in today's mainstream AI processors/GPUs (such as Nvidia's B200).

SpiNNaker's unique design also affords it so much memory and SRAM storage that it doesn't require any dedicated storage. Anything and everything the server is working on is stored permanently in the server's LPDDR4 memory and SRAM.

The second version of SpiNNaker represents a significant upgrade over its predecessor, achieving a 10-fold increase in core count compared to SpiNNaker 1. Other upgrades include the transition to a 22nm FD-SOI (22FDX) manufacturing process, adaptive body biasing for near-threshold voltage operation, workload dynamic voltage and frequency scaling, and accelerated and customized chip-to-chip interconnects.

SpiNNaker is capable of processing large-scale hybrid AI models, as well as biological neural simulations and whole-brain modeling, with its unique design. It is also highly scalable, with a machine in Dresden allegedly projected to incorporate over 720 motherboards, totaling 5.2 million CPU cores once complete.

A highly significant China tech industry merger looks set to go ahead, reports the South China Morning Post (SCMP). The Hong Kong-based organ says that the merger plan between chip designer Hygon Information Technology and supercomputer maker Sugon marks “a major move to consolidate two of the leading players in China’s computing supply chain.” We may be seeing the forming of a highly impactful vertically integrated supercomputing giant that has blossomed in the shadow of U.S. sanctions.

The proposed deal involves Hygon absorbing Sugon shares in a stock-swap agreement. Should the process complete successfully, with both companies’ shares being taken off the open market for up to 10 trading days, the newly consolidated entity will appear on the Shanghai stock exchange.

To give this merger some context, regular Tom’s Hardware readers will be aware that Hygon chips leverage the AMD Zen processor architecture. However, the firm says it has moved on from those days. In a recent report we published highlighting an extraordinary Hygon C86-5G, a 128-core, 512-thread CPU with AVX-512 and 16-channel DDR5-5600 support, we quoted a company exec asserting it uses a "new self-developed microarchitecture” in its latest designs.

Our previous reports on Sugon made clear its existing close relationship with Hygon. In recent years, the supercomputer maker leaned heavily on Hygon x86 chips to develop high-performance platforms. Sugon is backed by the Chinese Academy of Sciences, explains the SCMP, and has managed to push China into the “global top three for supercomputing,” it is claimed.

The merging partners were both on the U.S. Entity List

MSI has confirmed it will take the covers off a host of exciting new products at Computex 2025 later this month, including a brand new desktop AI supercomputer powered by Nvidia's DGX Spark platform.

The company confirmed in a press release that it will unveil its EdgeXpert MS-C931, a new desktop AI supercomputer built on the Nvidia DG Spark platform. The MS-C931 is powered by Nvidia's GB10 Grace Blackwell Superchip and is capable of 1,000 AI TOPS FP4 performance. The GB10 SoC features Nvidia Blackwell GPU architecture and fifth-generation Tensor Cores, as well as NVLink-C2C connection to Nvidia's Grace CPU, an Arm architecture core featuring 20 power-efficient chips.

It will also feature ConnectX 7 networking, 128GB unified memory, and LLM support, which Nvidia has previously promised can support up to 4TB of NVMe storage, and can run up to 200-billion-parameter LLMs, or 405-billion parameter models when running two linked chips.

Nvidia's DGX Spark platform promises compact and efficient performance, and comes pre-installed with Nvidia's AI software stack so that developers can run AI models from all the major players, including DeepSeek, Meta, and Google.

Alongside the MS-C931, MSI also says it will unveil a new lineup of Industrial Motherboards, as well as systems powered by Intel Twin Lake, Raptor Lake Refresh, Bartlett Lake, and Arrow Lake processor families.

Specifically, the company highlighted three new products. The MS-C926 is an ultra-slim fanless box PC with applications in smart retail and digital signage.

The Ms-927 is an ultra-compact box PC featuring Intel Core Ultra processors for high-performance edge computing.

Finally, the MS-CF20 is a new next-generation ATX motherboard featuring 16th Gen Intel Arrow Lake-S processors.

MSI says it will also have live demonstrations for solutions for smart retail and digital signage. There will also be a new fanless palm box for "space-constrained industrial environments, remote control management solutions utilizing SysLink, and edge AI innovations, including LLM and chatbot applications enabled by AI smartLink software.

The Fugaku supercomputer was again one of the unlikely stars of the recent Nico Nico Super Conference in Japan. In a re-run of last year, Fugaku was the brains behind the ‘Super Keisan Battle’ competition, where humans ritually humiliate themselves in a calculation race with a petascale supercomputer. Up for grabs was a free full day to utilize Fugaku however you might want, a ‘Fugaku One-Day Unlimited Ticket.’ However, the competition page contains an admission that this is a game “that no one can win” against Fugaku (machine translation).

Human 13, Fugaku 442,010,000,000,000,000,000,000,000

The Fugaku Face Off is pretty simple, and kind of stacks things even more in the supercomputer’s favor. According to the instructions we saw, a competition entrant must solve as many numerical additions as possible in 10 seconds. Most seemed to achieve 10 or 11, but Tomooo_108, for example, managed to solve 13 sums in the allotted time. This pales into insignificance vs the supercomputer, though. Fugaku managed to complete 442,010,000,000,000,000,000,000,000 sums during the 10 second battle. That’s four sextillion four hundred twenty quintillion one hundred quadrillion calculations.

Competitors mostly found the scale of their defeat against Fugaku quite amusing, in social media comments we have seen. Some joked that if you could beat Fugaku, you wouldn’t need to use it for the free day. Others pondered over the cosmic collapse that would occur if a human could move a pen fast enough to record four sextillion sums in 10 seconds. Good news - even if you lost against Fugaku, there were consolation prizes such as a miniature Fugaku to take home (expand the Tweet embedded above to see them).

At the Nico Nico Super Conference, which seems to be dominated by tech entertainment phenomena (e.g. VTubers and Cosplay), the Fugaku Face Off is another fun diversion, but with a serious twist. The supercomputer makers remind entrants that Fugaku, and its brethren, are used for tasks like research on disaster prevention – flood damage simulation, optimal evacuation plans, and so on.

Fugaku by the numbers

The Colossus Supercomputer may be a colossal problem for the world's richest man.

Last July Elon Musk fired up the Colossus, a supercomputer that uses 100,000 Nvidia H100 GPUs on a single fabric. The site could draw only 7 MW at its launch, however, enough to power just 4% of its GPUs. To solve this, Musk deployed several massive mobile generators to deliver the site’s electrical demands.

This was supposed to be a temporary solution. At the same time, the Memphis facility awaited approval from the Tennessee Valley Authority (TVA) for at least 50 MW of supply and the completion of its 150 MW substation (which was supposed to be finished by 4Q24).

It’s already 2Q25, and Ars Technica reports that residents say the site is still using over 30 gas turbines, which “release harmful pollution that is tied to asthma, respiratory illnesses, and certain types of cancers.”

Currently, xAI has an ongoing application with the Memphis authorities for 15 turbines. But when the Southern Environmental Law Center (SELC) partnered with South Wings to photograph the site with thermal imaging cameras, they discovered over 30 hotspots, indicating the number of generators operating on the site.

These power sources operate without permits because a legal loophole allows generators to be used for 364 days without one. But with July 2025 coming quickly, xAI needs to have its applications approved. Otherwise, it risks slowing down its operations (or shutting down completely) without access to the electricity it needs for all its GPUs.

Thermal Imaging Shows At Least 35 Portable Methane Gas Turbines

AMD processors were instrumental in achieving a new world record during a recent Ansys Fluent computational fluid dynamics (CFD) simulation run on the Frontier supercomputer at the Oak Ridge National Laboratory (ORNL). According to a press release by Ansys, it ran a 2.2-billion-cell axial turbine simulation for Baker Hughes, an energy technology company, testing its next-generation gas turbines aimed at increasing efficiency. The simulation previously took 38.5 hours to complete on 3,700 CPU cores. By using 1,024 AMD Instinct MI250X accelerators paired with AMD EPYC CPUs in Frontier, the simulation time was slashed to 1.5 hours. This is more than 25 times faster, allowing the company to see the impact of the changes it makes on designs much more quickly.

Frontier was once the fastest supercomputer in the world, and it was also the first one to break into exascale performance. It replaced the Summit supercomputer, which was decommissioned in November 2024. However, the El Capitan supercomputer, located at the Lawrence Livermore National Laboratory, broke Frontier’s record at around the same time. Both Frontier and El Capitan are powered by AMD GPUs, with the former boasting 9,408 AMD EPYC processors and 37,632 AMD Instinct MI250X accelerators. On the other hand, the latter uses 44,544 AMD Instinct MI300A accelerators.

Given those numbers, the Ansys Fluent CFD simulator apparently only used a fraction of the power available on Frontier. That means it has the potential to run even faster if it can utilize all the available accelerators on the supercomputer. It also shows that, despite Nvidia’s market dominance in AI GPUs, AMD remains a formidable competitor, with its CPUs and GPUs serving as the brains of some of the fastest supercomputers on Earth.

“By scaling high-fidelity CFD simulation software to unprecedented levels with the power of AMD Instinct GPUs, this collaboration demonstrates how cutting-edge supercomputing can solve some of the toughest engineering challenges, enabling breakthroughs in efficiency, sustainability, and innovation,” said Brad McCredie, AMD Senior Vice President for Data Center Engineering.

Even though AMD can deliver top-tier performance at a much cheaper price than Nvidia, many AI data centers prefer Team Green because of software issues with AMD’s hardware.

One high-profile example was Tiny Corp’s TinyBox system, which had problems with instability with its AMD Radeon RX 7900 XTX graphics cards. The problem was so bad that Dr. Lisa Su had to step in to fix the issues. And even though it was purportedly fixed, the company still released two versions of the TinyBox AI accelerator — one powered by AMD and the other by Nvidia. Tiny Corp also recommended the more expensive Team Green version, with its six RTX 4090 GPUs, because of its driver quality.

If Team Red can fix the software support on its great hardware, then it could likely get more customers for its chips and get a more even footing with Nvidia in the AI GPU market.

Argonne National Laboratory this week said that its Aurora supercomputer is now fully operational and is available to the scientific community. The machine, which was announced in 2015 and faced massive delays, now offers over 1 FP64 ExaFLOPS performance for simulation as well as 11.6 mixed precision ExaFLOPS for artificial intelligence and machine learning. 

"We are ecstatic to officially deploy Aurora for open scientific research," said Michael Papka, director of the Argonne Leadership Computing Facility (ALCF), a DOE Office of Science user facility. "Early users have given us a glimpse of Aurora's vast potential. We are eager to see how the broader scientific community will use the system to transform their research." 

The availability of the Aurora supercomputer for open scientific research may be considered a formal acceptance of the system by ARNL, which marks an important milestone for the troubled machine. Initially planned for 2018, Aurora missed this target due to Intel's decision to discontinue its Xeon Phi processors. After the machine was re-architected, the project faced further setbacks due to Intel's 7nm process technology delay, pushing the completion date to 2021 and then again to 2023. 

Even after the hardware was installed in June 2023, it took several months for the system to be fully operational and achieve exascale performance, which it finally reached in May 2024. Yet, for well over half a year, the system was only available to select researchers.

While Aurora is not the most powerful supercomputer for simulations, as its FP64 performance barely exceeds one ExaFLOPS, it is the most powerful system for AI as it can achieve 11.6 mixed precision ExaFLOPS according to the HPL-MxP benchmark. 

"A big target for Aurora is training large language models for science," said Rick Stevens, Argonne associate laboratory director for Computing, Environment and Life Sciences. "With the AuroraGPT project, for example, we are building a science-oriented foundation model that can distill knowledge across many domains, from biology to chemistry. One of the goals with Aurora is to enable researchers to create new AI tools that help them make progress as fast as they can think — not just as fast as their computations." 

Some of the first research projects using Aurora are detailed simulations of intricate systems, such as the human circulatory system, nuclear reactors, and supernova explosions. The machine's overwhelming performance is also instrumental in processing data from major research centers, such as Argonne's Advanced Photon Source (APS) and CERN's Large Hadron Collider. 

"The projects running on Aurora represent some of the most ambitious and innovative science happening today," said Katherine Riley, ALCF director of science.
"From modeling extremely complex physical systems to processing huge amounts of data, Aurora will accelerate discoveries that deepen our understanding of the world around us." 

On the hardware side, Aurora clearly impresses. The supercomputer comprises 166 racks, each holding 64 blades, for a total of 10,624 blades. Each blade contains two Xeon Max processors with 64 GB of HBM2E memory onboard and six Intel Data Center Max 'Ponte Vecchio' GPUs, all cooled by a specialized liquid-cooling system. 

In total, Aurora has 21,248 CPUs with over 1.1 million high-performance x86 cores, 19.9 PB of DDR5 memory, and 1.36 PB of HBM2E memory attached to the CPUs. It also features 63,744 GPUs optimized for AI and HPC equipped with 8.16 PB of HBM2E memory. Aurora uses 1,024 nodes with solid-state drives for storage, offering 220 PB of total capacity and 31 TB/s of bandwidth. The machine relies on HPE's Shasta supercomputer architecture with Slingshot interconnects.

Chinese Society of Computer Science has published a list of the Top 100 highest-performing supercomputers in the country, but the country's intentional obfuscation of its true computing power seems to have been taken to a new level. Just like last year, Chinese entities were too cautious to include their alleged ExaFLOPS-class supercomputers in the list. However, even more surprising is that the list contains no new systems. The only difference between the 2023 and 2024 Top 100 lists of China's most powerful supercomputers is their cumulative power, which indicates minor upgrades. Likely, Chinese entities are deliberately withholding information about their most powerful systems to avoid provoking more restrictions from the U.S. government.

China's Top 3 supercomputers in 2024 are the same heterogeneous CPU + GPU systems that topped the list in 2023. The range-topping machine in the official list was deployed in 2023; it features 15,974,400 CPU cores and a maximum Linpack performance of 487.94 PFLOPS. The machine is more powerful than Japan's Fugaku supercomputer (442 FP64 PetaFLOPS), but it lags significantly behind American ExaFLOPS-class machines El Capitan (1.742 ExaFLOPS), Frontier (1.353 ExaFLOPS), and Aurora (1.012 ExaFLOPS). 

The second highest-performing supercomputer in China was launched in 2022, it has 460,000 CPU cores, and a maximum Linpack performance of 208.26 PFLOPS. The third most powerful machine in China (well, officially) is a system with 285,000 CPU cores and a maximum Linpack performance of 125.04 PFLOPS. 

The difference in the aggregated performance of China's official Top 100 lists of supercomputers in 2023 and 2024 is negligible. Back in 2023, they offered an aggregated compute power of 1.398 ExaFLOPS, whereas in 2024, their performance increased to 1.406 ExaFLOPS. 

However, the official Top 100 list of supercomputers may not accurately reflect China's supercomputing capabilities. Jack Dongarra, a supercomputer industry luminary and co-founder of Top500.org, last year said that China has at least three ExaFLOPS-class machines with performance ranging from 1.3 ExaFLOPS to 1.7 ExaFLOPS, powered by hardware designed in China, as well as a 2 ExaFLOPS machine featuring x86 processors from Hygon. This information has never been confirmed, but the words of Jack Dongarra are usually taken seriously by the industry. 

For China's top-ranked supercomputers, the Chinese Society of Computer Science does not publish their exact specifications, possibly to obscure the suppliers of hardware for these systems. Analysts tend to be keen enough to figure out what hardware might be used and who can supply it. It is speculated that these machines are powered by industry-standard CPUs and GPUs, with most of their crucial components obtained through grey channels. Yet, there are systems in the Top 10 powered by China's domestically developed processors and accelerators.

The Oak Ridge National Laboratory (ORNL) Summit supercomputer is being decommissioned this month after spending several years churning through data 24/7. As part of its retirement, HPC Guru said on X that the ORNL planned a virtual farewell on Zoom for the once most powerful computer in the world. And although it’s set to be retired this month, it’s still running at almost full power, with only about 0.5% of its nodes running idle.

The supercomputer has been used by thousands of customers worldwide for various reasons. It even ran during the 2020 Coronavirus Global Pandemic and helped find a protein that would bind with the virus and prevent it from infecting people. Even as newer, more powerful supercomputers emerged, Summit was still able to retain the ninth spot in the Top 500.

Still, the Frontier supercomputer, which currently holds the top spot as the most powerful supercomputer, is already running in ORNL since 2022. Although it consumes over two times the power that Summit needs (22,768kW versus 10,096kW), it delivers over eight times the computing performance, making it far more efficient. The national laboratory is also working on diamond-based quantum accelerators with Quantum Brilliance (QB) as part of its research on quantum computing.

Power efficiency is the number one problem for many supercomputers and AI data centers, especially as the latest generations of processors, both AI and otherwise, consume more and more power. This has reached the point that many companies, like Google, Microsoft, Amazon, and Oracle, are investing in nuclear energy to deliver the projected power needs of future processors.

However, even current data centers are asking for more power, with Elon Musk’s xAI Colossus just getting approval for 150MWs of electricity from the Tennessee Valley Authority (TVA). All this meant that many companies were missing their climate targets, with Google’s carbon footprint jumping by more than 40% from 2019 and Microsoft’s greenhouse gas emissions increasing by nearly 30% from 2020. A former Google CEO even suggested suspending climate targets to alI technologies to go full tilt and then use that technology to catch up in the future.

While we cannot stop humanity’s desire for more computing power, ORNL’s move to retire its supercomputer for a more efficient one is a step in the right direction. Hopefully, chip makers like Nvidia, AMD, and Intel could make processors as efficient as they are influential in the future, thus reducing humanity’s demand for more electricity.

NEC this week announced that it had been selected to develop a next-generation supercomputer for Japan's National Institutes for Quantum Science and Technology (QST). The machine will use Intel's Xeon 6900P processors, AMD's Instinct MI300A accelerators and will offer performance of around 40 PetaFLOPS. It will be mainly tasked with advancing nuclear fusion research.

The system is set to be installed at QST’s Rokkasho Institute for Fusion Energy in Aomori, Japan, and will include 360 NEC LX 204Bin-3 units powered by 720 Intel Xeon 6900P processors equipped with MRDIMM DDR5 memory and 70 NEC LX 401Bax-3GA units with AMD Instinct MI300A GPUs, reaching a combined theoretical performance of 40.4 PetaFLOPS. The CPUs and GPUs will take advantage of Giga Computing-developed liquid cooling to ensure consistent performance and high reliability.

"By integrating the Intel Xeon 6900P Series, the first server CPU to support MRDIMMs, and the first in Japan, we are delivering a leap in memory performance and bandwidth, an ideal choice for complex calculations and simulations required in fusion research," said Ogi Brkic, Vice President & General Manager, Go-To-Market Builders & Technology Acceleration Office, Sales & Marketing Group, Intel.

For storage, the supercomputer will feature DDN's ES400NVX2 solution, which has a total capacity of 42.2PB and features the Lustre ExaScaler file system. As for network infrastructure, the machine will use an InfiniBand setup with Nvidia's QM9700 switches. On the software side, the machine will use Altair PBS Professional software for workload management and a scheduler optimized for AMD's Instinct MI300A accelerators.

"We appreciate that NEC has selected AMD's Instinct MI300A and this choice is further proof that AMD's MI300 Series GPUs offer a compelling supercomputer accelerator solution," Jon Robottom, Corporate Vice President, Representative Director & General Manager, AMD Japan. "We believe that AMD innovation, together with NEC's advanced technological capabilities, will continue to make a significant contribution to research conducted at the QST and NIFS."

A performance level of 40.4 PetaFLOPS is 2.7 times greater than the two current systems at QST and the National Institute for Fusion Science (NIFS), which will provide a significant boost for supercomputer-based simulations for fusion science research as well as AI, and Big Data applications.

The new supercomputer is set to be operational starting from July 2025.

The Russian government plans to boost supercomputer usage by offering grants to companies that rent them, reports CNews. This initiative will be instrumental in encouraging businesses and research institutions to adopt high-performance computing. With new supercomputers coming online in the coming years, Russia could increase computational power tenfold by 2030. However, it is unclear how Russia plans to build supercomputers as companies like AMD, Intel, and Nvidia cannot sell their highest-performance AI and HPC processors to Russian entities. 

The Ministry of Digital Development will provide financial support to organizations using supercomputers for tasks like artificial intelligence training and simulating intricate processes, such as molecular modeling. Currently, Russian supercomputers are highly sought after by research facilities, particularly in fields that require loads of compute. According to experts, this grant program will be most valuable to scientific labs that rely on supercomputers for various simulation workloads. Businesses will also benefit from these supercomputer grants, especially in specialized industries that involve complex design and process optimization.  

The high cost of acquiring and operating supercomputers is a significant hurdle for small companies. AI projects require compute GPU resources worth millions of dollars, which most startups cannot afford. These grants aim to lower financial barriers, allowing smaller businesses to access advanced technology and scale their innovations. 

Despite existing investments, Russia has massive unmet demand for computational power to develop products like polymers and composites, potentially costing the market millions of dollars each year. This lack of resources hinders innovation in areas that require advanced simulations and calculations. 

Institutions like ITMO University, even with newly installed supercomputing systems, face challenges in meeting their growing needs. The demand for training fundamental AI models significantly exceeds their current capabilities, confirming the need to enhance HPC capabilities in the Russian academic sector. 

Major Russian companies like Yandex, Sber, Moscow State University, and MTS currently own the country's top supercomputers. Machines like Lomonosov 2 and Christofari are already running at full capacity as companies like Sber and Yandex train their AI models, whereas the MSU uses its machine for scientific tasks. 

During his address to the Federal Assembly in late February 2024, the Russian president called for a tenfold performance increase of Russian supercomputers by 2030. To support this growth, the government also plans to reimburse companies building supercomputers for AI training for the costs associated with connecting them to the power grid. 

One thing that is unclear is how Russian entities plan to get new supercomputers, considering U.S. restrictions on selling supercomputers and AI parts to Russian and Chinese entities. One way for Russian entities to get processors like Nvidia's H100 or H200 is to smuggle them through third countries, such as China or the UAE. However, smuggling thousands of processors is a hard and expensive task.

Isambard 2, one of the first Arm-based supercomputers, is finally set to retire after six years. First deployed in May 2018, this 10,000-core machine used 64-bit Armv8 ThunderX2 processors developed by Cavium and manufactured by TSMC, plus a few Nvidia P100 GPUs.

The Great Western 4 (GW4) Alliance, a group of universities from Bristol, Bath, Cardiff, and Exeter, was the primary operator of the Isambard 2, with Professor Simon McIntosh-Smith, Head of the Microelectronics Group at the University of Bristol, leading the project. He also broke the news about Isambard 2’s retirement in an X post, saying that its successor, Isambard 3, will come online to take over its duties. The Arm supercomputer will go offline at 9 am on Monday, September 30, so users must move all data off it by then.

The new supercomputer, Isambard 3, is still powered by Arm processors, but this time, it will have Nvidia Grace CPU Superchips with 34,272 cores. Although it’s currently not in the Top500’s list of most powerful supercomputers, it achieved second place in the Green500 in June 2024, making it one of the most efficient supercomputers available today. Furthermore, Isambard 3 is set to expand its processor count by an additional 5,280, increasing its performance by up to 32 times and potentially land it in the top 10 of the Top500 list in its next run.

Isambard 2 isn’t the first high-profile supercomputer to retire in 2024. Oak Ridge National Laboratory’s Summit supercomputer is also set to be decommissioned in November this year. This supercomputer was built in 2018, the same year as Isambard 2, but has since been replaced by the far more powerful Frontier supercomputer.

Supercomputers cost millions of dollars to acquire and even more to operate. However, as technology advances, institutions need to keep up, thus requiring the retirement of older models, even if they still operate well. Newer silicon offers better performance and efficiency, allowing researchers to achieve breakthroughs faster. As such, these massive computer advancements in science and technology will make the enormous investment worth it.

Oak Ridge National Laboratory has decommissioned its Alpine storage system, a 250-petabyte storage system that held data for the Summit supercomputer and its other support systems. The Summit supercomputer, currently the world's ninth-fastest supercomputer, will be retired on November 1st, but its aging Alpine storage didn't survive that long.

As ORNL prepares for the Discovery system—a computer set to be the world's fastest with an estimated 8.5 exaFLOPS of performance—the time has come to decommission its predecessor, Summit. Summit was initially set to be shut down in 2023, but its high productivity rates led the Department of Energy to keep it in operation for another year. Unfortunately, its storage could not last so long.

Alpine was part of ORNL's storage solution for Summit and its peripheral systems, holding scratch data from the supercomputer and its external nodes, which pre- and post-process Summit's calculations. The Alpine storage system held 250 petabytes of capacity inside 32,494 10TB NL-SAS drives. Made up of 77 IBM Elastic Storage Server (ESS) nodes, the system could have 2.2 TB/s random read and write speeds at its peak. Still, in recent years, drive failure rates have hit unacceptable levels, necessitating installing a stopgap replacement storage system, Alpine2.

When the time came to dismantle Alpine, the ORNL team could fully dismantle the data servers in under two months, thanks to an industrial disk drive shredder. An outside vendor brought a mobile shredder, a four-foot-wide, three-horsepower unit that can eat one hard drive every 10 seconds. ORNL gives an estimated 12,000 clients access to Summit’s computing power, so data security was seen as essential.

“Even though we’re not dealing with classified data, the data still belongs to the users, and we have a responsibility to make sure it’s protected,” said Paul Abston, group leader for HPC infrastructure at ORNL. “The teeth of the shredder tear the drives into tiny pieces, making it impossible to reconstruct into a functioning drive.”

Thanks to help from outside vendors, ORNL fully shredded and recycled all 32,000 of Alpine’s drives, plus an additional 10,000 drives from other Summit support systems. The effort far outpaced Oak Ridge’s previous significant decommissioning effort of dismantling the Atlas storage system in 2019, a 20,000-hard drive job that took nine months. ORNL also purchased its heavy-duty data shredder, making future jobs even more efficient.

Summit once stood as the world’s most powerful supercomputer, a title it held for almost a year after its initialization in 2018. It runs on 4,356 compute nodes, each powered by two IBM Power9 22-core 3.07 GHz CPUs and six Nvidia Tesla GV100 GPUs. Summit uses the Alpine2 storage system, a 50-petabyte filesystem across 16 IBM ESS nodes. This data-lite storage system will support Summit until its end of life, transitioning to read-only mode on November 19th and eventually phasing into the bullpen of ORNL’s supplementary parallel filesystems.

Summit’s next-gen replacement, Discovery, is not expected to come online until 2027; the application period for potential vendors ended on August 30th, but no winning bid has yet been announced. In the meantime, ORNL still operates Frontier, the world’s fastest supercomputer per the Top500 leaderboards.

Oak Ridge National Laboratory’s (ORNL) Summit supercomputer is set to be decommissioned in November this year, after serving almost six years and delivering 200 million node hours to researchers. The Oak Ridge Leadership Computing Facility (OLCF) made this announcement today on X (formerly Twitter) after the last day of the 2024 OLCF User Meeting, where its attendees signed a piece of the supercomputer to commemorate its faithful service.  

Summit once stood as the most powerful supercomputer in the world, taking the top spot on the Top500 list during 2018 and 2019. It has 4,356 nodes, each one powered by two IBM Power9 22-core 3.07 GHz CPUs with six Nvidia Tesla GV100 GPUs. It was dethroned in 2020 by the Arm-based Fugaku supercomputer, but ORNL regained the top spot in 2022 when it introduced the AMD-powered Frontier supercomputer.

Despite being six years old, which is a relative eternity in terms of computer development, Summit never left the top ten of Top500’s most powerful supercomputers during its lifetime. However, it seems that the ORNL has deemed that it’s time for the Summit supercomputer to retire. After all, Summit’s theoretical peak of 200.79 PFlop/s pales against Frontier’s 1,714.81 PFlop/s, the first supercomputer to break the exascale barrier. Aside from this, ORNL is working with Quantum Brilliance (QB) to integrate the latter’s quantum accelerators, helping the former’s researchers evaluate their viability for quantum computing.

The newer Frontier supercomputer is also far more efficient, as it only consumes a little over twice the power (22,786 kW) that Summit needed (10,096 kW) while delivering over eight times the performance. This is crucial especially as power consumption is now the number one concern in high-performance computing. It’s expected that a single modern H100 GPU would consume at least 3.7 MWh of power annually, with all the AI GPUs sold last year alone already accounting for 14,348.36 GWh of electricity use. And with Nvidia’s next-generation Blackwell GPUs expected to consume even more power, we’re in dire need of new processors that are far more efficient while still able to deliver the ever-growing computing power that our data-driven society needs.

Japan's Ministry of Education, Culture, Sports, Science and Technology (MEXT) has announced [PDF] plans to build a successor to the country's Fugaku supercomputer, which was once the world's most powerful HPC machine. The ministry wants RIKEN and Fujitsu to start developing the supercomputer next year, reports Nikkei.

A document by MEXT says this new supercomputer aims to achieve an unprecedented 50 ExaFLOPS of AI performance with Zetta-scale peak performance in mind to use AI for scientific purposes. The Zetta-class designation indicates a system capable of performing one sextillion floating-point operations per second. ZettaFLOPS is 1,000 times faster than ExaFLOPS, so if Japan manages to build such a system by 2030, as planned, it will likely again have the world's fastest supercomputer.  

MEXT wants each computational node of the Fugaku Next supercomputer to have peak performance of several hundred FP64 TFLOPS for double-precision computations, around 50 FP16 PFLOPS for AI-oriented half-precision calculations, and approximately 100 PFLOPS for AI-oriented 8-bit precision calculations, with memory bandwidths reaching several hundred TB/s using HBM-type memory. To put the number into context, the peak performance of a Fugaku computational node is 3.4 TFLOPS for double-precision calculations, 13.5 TFLOPS for half-precision calculations, and 1.0 TB/s for memory bandwidth. 

The Ministry of Education, Culture, Sports, Science, and Technology plans to allocate ¥4.2 billion ($29.06 million) in the first year of development, with total government funding expected to exceed ¥110 billion ($761 million).  

RIKEN, one of Japan's most prominent research institutions, is expected to lead the development of Fugaku's successor. MEXT wants its Fugaku Next supercomputer to use Japanese technology. Considering that RIKEN will likely want to maintain software compatibility with Fugaku, it is possible that Fujitsu will lead the development of hardware for Japan's next big supercomputers.  

MEXT does not envision any particular architecture for the Fugaku Next supercomputer, though its documents suggest that it should use a CPU with special-purpose accelerators or a combination of CPU and GPU. Also, MEXT wants the supercomputer to feature an advanced storage system capable of handling both traditional I/O workloads for data science and large-scale checkpointing and new I/O requirements for AI workloads. 

If the successor of Fugaku uses a processor designed by Fujitsu, it will probably use Fujitsu's processor that will follow Monaka, Fujitsu's next-generation CPUs for data centers and commercial supercomputers featuring up to 150 enhanced Armv9 cores. The CPU will adopt a multi-chiplet configuration, so these cores will be distributed across multiple-core dies, which will be paired with SRAM dies, and I/O dies. The I/O dies will facilitate support for DDR5 memory, PCIe 6.0 connectivity, and CXL 3.0 for linking various accelerators and extenders. The core dies are reportedly being produced using TSMC's 2nm fabrication process. 

Japan's next-generation supercomputer is expected to come online in 2030. It is expected to use processors with even more cores and more advanced connectivity compared to Monaka. It could also be made on a 1nm-class process technology or perhaps an even more advanced node.

The U.S. government has proposed a regulation requiring citizens and permanent residents to report transactions involving the construction of supercomputers with performance exceeding 100 PetaFLOPS in countries of concern, primarily China, reports HPCWire. The regulation is meant to further solidify restrictions imposed in 2022 that prevent selling processors that enable 100 PFLOPS supercomputers to China.

The document, Provisions Pertaining to U.S. Investments in Certain National Security Technologies and Products in Countries of Concern, explicitly identifies 'categories of transactions involving technologies and products that may contribute to the threat to the national security of the United States.' 

In 2022, the U.S. government prohibited chipmakers (as well as distributors) from selling processors that could enable building supercomputers with performance of over 100 FP64 PetaFLOPS or over 200 FP32 PetaFLOPS within 41,600 cubic feet (1178 cubic meters) or smaller space to Chinese entities or entities associated with China. 

The new regulation specifically targets transactions related to building, selling, or producing supercomputers in countries of concern (China, Macau, and Hong Kong) with over 100 FP64 PFLOPS or 200 FP32 PFLOPS in the said dimensions. The new rules are not limited to developers of CPUs, GPUs, or FPGAs but cover a much broader spectrum of hardware. These rules also apply to individuals or entities with controlling interests in foreign organizations or those involved in managing such activities. 

The Semiconductor Industry Association (SIA) and National Venture Capital Association (NVCA) have voiced strong opposition to the proposed law. They argue that the financial burden of compliance could be significantly higher than the government’s estimates, potentially reaching up to $100 million. These groups also worry that the restrictions could erode the US’s competitive edge in the semiconductor industry.

Venture capital firm Andreessen Horowitz (a16z) has suggested that the government should reconsider its focus on computing power thresholds, arguing that these could quickly become outdated. Instead, they propose that restrictions should be based on the type of foreign entities involved rather than the computational capabilities.

As the proposed regulation remains in the comment period, the tech industry continues to push for revisions, emphasizing the need for clearer definitions and a more focused approach that safeguards national security without stifling innovation or harming US companies' global competitiveness.

Six major Russian supercomputer centers located across the country have teamed up to form the 'Distributed Scientific Supercomputer Infrastructure' consortium, reports the Russian Academy of Sciences. This new alliance aims to pool resources for the performance needed to process modern scientific workloads. In total, the six supercomputing centers have around 900 servers and these fleets cannot be upgraded due to the U.S. sanctions. To that end, pooling resources is largely a forced measure. Meanwhile, DSSI has a plan to collectively upgrade their hardware.

Six supercomputing centers join forces

The consortium's consolidated infrastructure includes 900 servers with a total peak performance of 1.5 FP64 PetaFLOPS and scientific data storage systems exceeding 15PB. This infrastructure supports distributed systems for the collection, storage, and analysis of scientific data across various regions. To put the number into context, 1.5 FP64 PLOPS performance trails the slowest supercomputer in the Top 500 list, which has Rmax performance of 2.13 PetaFLOPS. Members of the DSSI are providing computing resources and support to 240 organizations across the country and by working together they will be able to offer more compute performance to their customers as the latter's requirements increase. 

The consortium does not include Russia's most powerful state-owned supercomputer, the Lomonosov 2 located in the Lomonosov Moscow State University, which has Rmax performance of 2.48 FP64 PFLOPS. It should be noted that the MSU introduced its MSU-270 supercomputer with a peak computational power of 400 'AI' PFLOPS last Fall and this one also does not participate in the DSSI project. Commercial supercomputers, such as Yandex's Chervonenkis, Russia's most powerful machine with Rmax Linpack performance of 21.53 FP64 PFLOPS, will also not be a part of the consortium. 

The DSSI consortium brings together multiple players from different regions, including the Far Eastern Federal Research Center in Khabarovsk, the Institute of Automation and Control Processes in Vladivostok, the Matrosov Institute in Irkutsk, the Institute of Computational Mathematics and Mathematical Geophysics in Novosibirsk, the Krasovsky Institute in Yekaterinburg, and the Space Research Institute in Moscow.  

"We need to move away from competition between collective-use centers and instead focus on uniting their resources and expertise," said Alexey Sorokin, director of the computing center of the Far Eastern Branch of the Russian Academy of Sciences and a member of the consortium's Coordinating Council. "There is a growing demand for a distributed infrastructure that connects technological sites across different regions of the country via high-speed communication channels. These sites should be equipped with modern computing systems of various architectures and specialized data storage systems. It is impossible for even the largest organization to tackle this complex task alone. It makes sense to leverage the groundwork laid by existing supercomputer centers, which have already proven their efficiency. This approach will undoubtedly lead to more efficient use of financial resources and, just as importantly, foster the development of scientific teams in the regions, particularly in HPC space."  

And have upgrade plans

Due to sanctions imposed by the U.S. after Russia invaded Ukraine, Russian state-owned entities have a hard time acquiring supercomputer hardware, such as the latest processors from AMD, Intel, and Nvidia. While smuggling exists (and largely prospers), it is likely that Russian institutions cannot get enough new machines to modernize their fleets.  

Pooling in resources will enable Russian scientists to get access to higher performance. Also, the DSSI consortium aims to develop a comprehensive project to upgrade scientific supercomputer centers for collective use with modern compute hardware. After discussions with all interested participants, the project will be submitted to the Ministry of Science and Higher Education of the Russian Federation and the Russian Academy of Sciences. 

Given the sanctions imposed by the U.S., EU, Taiwan, South Korea, and other countries, it will not be easy to upgrade all of the aforementioned datacenters with modern Western hardware. But perhaps the plan is to use parts that can be obtained without any curbs: including usage of cut-down components like Nvidia's HGX H20 or hardware from Chinese developers. However, it is evident that the Russian scientific community wants access to higher performance.

Although the Frontier supercomputer, located at Oak Ridge National Laboratory, is considered the fastest supercomputer in the world because it tops the global Top500 list, it may not be the de facto leader. Some scientific papers suggest that Chinese supercomputers, the Sunway and Tianhe-3, may surpass Frontier in performance, believes Jack Dongarra, a co-founder of Top 500. 

"The Chinese have machines that are faster," Dongarra told the Wall Street Journal. "They just have not submitted the results." 

In recent years, China's government and scientists have become secretive and no longer disclose domestic achievements in the field of supercomputing because they don't want the U.S. government to make it harder for them to obtain high-performance hardware. 

China has better supercomputers than in the U.S.?

The latest version of the Top 500 list was released last month. Despite the latest rankings indicating that the three fastest supercomputers are in the U.S., there is a strong belief that China possesses more powerful machines. According to Dongarra, China has faster supercomputers but has not submitted their results due to concerns over further U.S. restrictions. 

Some experts from the U.S. believe that the Sunway machine has 39 million cores, which is quadruple that of Frontier, which may mean that this machine could be more powerful than the Frontier supercomputer from ORNL. Those processors powering the Sunway supercomputer may be made using an outdated process technology and may not be as energy efficient as modern American CPUs. However, brute force is brute force, and for China, economic or power efficiency may not really matter as it also considers supercomputers a national security matter. 

China dominated the Top500 list by 2017, with 202 machines compared to 143 from the U.S. Then the U.S. restricted Chinese access to Intel processors and other U.S. hardware in 2015, followed by broader export restrictions under the Trump administration in 2019, which have been tightened further by the Biden administration in 2022. As a result, Chinese participation in the Top500 list dwindled, to some degree because access of Chinese entities to the latest hardware got harder and to some degree because Chinese scientists no longer want to share details about their machines with anyone. 

Dongarra told the WSJ that Chinese colleagues told him they were not permitted to submit information about their supercomputers, leading to reduced data sharing with Top500 in particular and with other scientific forums in general. However, based on his conversation with his colleagues, he believes Chinese scientists have very fast machines that surpass the capabilities of America's best supercomputers. 

China has its own HPC Top100 list of supercomputers, and Dongarra believes this list omits some of China's top supercomputers. The number one machine on the latest Top100 and several others is described only in general terms without specifying its name or operating institution. Last December, a month after the latest list was published, the National Supercomputing Center in Guangzhou unveiled a machine named Tianhe Xingyi, claiming it significantly outperformed the earlier Tianhe-2 model from the Milky Way series and that machine could be faster than ORNL's Frontier.

On the one hand, this means that American and Chinese scientists will no longer collaborate on supercomputers, creating a divide that Western scientists believe will hinder technological development. Nations will work on different projects or similar projects, albeit on their own. On the other hand, this means that nobody outside China knows the performance of Chinese supercomputers.  

This growing secrecy poses challenges for the U.S. in determining whether it or China possesses faster supercomputers, a question deemed crucial for national security. Supercomputers play a pivotal role in the U.S.-China technological rivalry, as the nation with superior machines gains an edge in developing advanced military technology, including nuclear weapons. Jimmy Goodrich, a senior adviser at the Rand Corporation, told WSJ that even a slight supercomputing advantage can significantly impact military capabilities.

Last week, the Department of Energy (DOE) issued a request for proposals (RFP) to develop a new supercomputer named Discovery. This supercomputer will replace the current known fastest supercomputer in the world, Frontier, at Oak Ridge National Laboratory. Discovery aims to surpass Frontier's performance, offering three to five times more computational throughput (e.g., 8.5 ExaFLOPS) by 2027 or early 2028.  

ORNL mentions advanced AI, machine learning, improved energy efficiency, and comprehensive system modeling among the workloads that will run on the Discovery supercomputer. Unlike previous RFPs, this one does not specify an exact performance increase but only says that the new supercomputer has to be three to five times more powerful than its predecessor. 

Discovery's computational power will support scientific research in various fields, including AI, climate change, drug discovery, nuclear security, and green energy solutions. Researchers will be able to leverage Discovery’s advanced computational power and capabilities in modeling, simulation, high-performance data analysis, and AI to achieve significant breakthroughs in both scientific and industrial fields. Like Frontier, scholars globally will have the chance to compete for computing time on Discovery to address major scientific challenges. 

"Discovery will enable the scientific community to model real-world situations at new levels of detail. It will help us study challenging problems we cannot easily explore with experiment, observation or theory alone," said Georgia Tourassi, ORNL associate laboratory director of computing and computational sciences. "Using Discovery, scientists will improve the safety and efficiency of nuclear power plants and aerospace engineering, pushing the boundaries of what’s possible in sustainable power generation and aviation. They will accelerate the development of new drugs and advanced materials. They will even gain deeper insights in global climate dynamics to inform critical decisions for our collective future." 

Proposals for Discovery are due by August 30, 2024.  

The ORNL has a history of deploying the world's fastest supercomputers. For example, Jaguar, Titan, and Summit led the world's Top 500 list in different years. Frontier is the world's No. 1 supercomputer today. In fact, over the past decade, the facility has increased its computational power 500-fold while only quadrupling energy consumption.

The Holy Grail of supercomputing chip design is an architecture that combines the versatility and programmability of CPUs with the explicit parallelism of GPUs, and InspireSemi strives to achieve just that. InspireSemi's Thunderbird 'supercomputer-cluster-on-a-chip' packs 1,536 RISC-V cores designed specifically for high-performance computing, but it also supports a general-purpose CPU programming model. It also has incredible scalability — four chips can be placed on a single accelerator card, which comes in a standard GPU-like form factor (AIC), bringing the total of cores per card up to 6,144, with extended scalability to 360,000 cores per cluster. 

InspireSemi's Thunderbird processor packs 1,536 custom 64-bit superscalar RISC-V cores with plenty of high-performance SRAM, accelerators for several cryptography algorithms, and an on-chip low-latency mesh fabric for inter- and intra-chip connectivity. The chip also supports LPDDR memory, NVMe storage, PCIe, and GbE connectivity. It has been taped out and will be fabbed at TSMC, then packaged at ASE. 

InspireSemi aims to install four Thunderbird chips onto a single board to offer developers 6,144 RISC-V cores. The current Thunderbird architecture supports scale-out capability to up to 256 processors connected using high-speed serial transceivers.  

When it comes to performance, InspireSemi says that its solution offers up to 24 FP64 TFLOPS at 50 GFLOPS/W (at 480W), which is a formidable performance. To put this into context, Nvidia's A100 delivers 19.5 FP64 TFLOPS, whereas Nvidia's H100 reaches 67 FP64 TFLOPS. It's unclear if we are dealing with the performance of a single-chip Thunderbird card or a 4-way model. Delivering 1920W to an add-in card is hardly possible, so it is likely that we are dealing with four Thunderbird processors on a card that can offer 24 FP64 TFLOPS per chip. 

The superscalar cores support vector tensor operations and mixed-precision floating point data formats, though there is no word that these cores are Linux-capable, which is why InspireSemi calls Thunderbird an accelerator rather than a general-purpose processor. Still, this processor can be programmed like a regular RISC-V CPU and supports a variety of workloads, such as AI, HPC, graph analytics, blockchain, and other compute-intensive applications. As a result, InspireSemi's customers will not have to use proprietary tools or software stacks like Nvidia's CUDA. The only question is whether industry-standard tools and software stacks will be enough to extract maximum performance from the Thunderbird I in all kinds of workloads. 

"We are proud of the accomplishment of our engineering and operations team to finish the Thunderbird I design and submit it to our world-class supply chain partners, TSMC, ASE, and imec for production," said Ron Van Dell, CEO of InspireSemi. "We expect to begin customer deliveries in the fourth quarter." 

Speaking of customers and partners, InspireSemi has a long list of companies it works with, including Lenovo, Penguin Computing, 2CRSI, World Wide Computing, GigaIO, Cadence, and GUC, just to name a few. 

"This is a major milestone for our company and an exciting time to be bringing this versatile accelerated computing solution to market," said Alex Gray, Founder, CTO, and President of InspireSemi. "Thunderbird accelerates many critical applications in important industries that other approaches do not, including life sciences, genomics, medical devices, climate change research, and applications that require deep simulation and modeling."

Demand for more computing power in the data center is growing at a staggering pace, and AMD has revealed that it has had serious inquiries to build single AI clusters packing a whopping 1.2 million GPUs or more.

AMD's admission comes from a lengthy discussion The Next Platform had with Forrest Norrod, AMD's EVP and GM of the Datacenter Solutions Group, about the future of AMD in the data center. One of the most eye-opening responses was about the biggest AI training cluster that someone is seriously considering.

When asked if the company has fielded inquiries for clusters as large as 1.2 million GPUs, Forrest replied that the assessment was virtually spot on.

Morgan: What’s the biggest AI training cluster that somebody is serious about – you don’t have to name names. Has somebody come to you and said with MI500, I need 1.2 million GPUs or whatever.

Forrest Norrod: It’s in that range? Yes.

Morgan: You can’t just say “it’s in that range.” What’s the biggest actual number?

Forrest Norrod: I am dead serious, it is in that range.

Morgan: For one machine.

Forrest Norrod: Yes, I’m talking about one machine.

Morgan: It boggles the mind a little bit, you know?

1.2 million GPUs is an absurd number (mind-boggling, as Forest quips later in the interview). AI-training clusters are often built with a few thousand GPUs connected via a high-speed interconnect across several server racks or less. By contrast, creating an AI cluster with 1.2 million GPUs seems virtually impossible.

We can only imagine the pitfalls someone will need to overcome to try and build an AI cluster with over a million GPUs, but latency, power, and the inevitability of hardware failures are a few factors that immediately come to mind.

AI workloads are extremely sensitive to latency, particularly tail latency and outliers, wherein certain data transfers take much longer than others and disrupt the workload. Additionally, today's supercomputers have to mitigate the GPU or other hardware failures that, at their scale, occur every few hours. Those issues would become far more pronounced when scaling to 30X the size of today's largest known clusters. And that's before we even touch on the nuclear power plant-sized power delivery required for such an audacious goal.

Even the most powerful supercomputers in the world don't scale to millions of GPUs. For instance, the fastest operational supercomputer right now, Frontier, "only" has 37,888 GPUs.

The goal of million-GPU clusters speaks to the seriousness of the AI race that is molding the 2020s. If it is in the realm of possibility, someone will try to do it if it means greater AI processing power. Forest didn't say which organization is considering building a system of this scale but did mention that "very sober people" are contemplating spending tens to hundreds of billions of dollars on AI training clusters (which is why millions of GPU clusters are being considered at all).

An electrical engineer who helps design supercomputers has built a miniaturized version of the Cray C90 that fits inside a wristwatch. The marriage of 3D printing, an OLED display, and an FPGA development board yielded a tiny Cray that he can wear on his wrist and, with great difficulty and concentration, use to tell time.

To build the 1/25-scale Cray C90 wristwatch, Chris Fenton designed a model of the Cray Y-MP C916 small enough to wear as a watch. He built it with a 3D printer, trimmed the required circuit boards to fit inside, and mounted the OLED display. It also has a battery with a built-in charger, and Fenton says the watch can accommodate a NATO-style wrist strap.

The system is run off a Diligent CMOD-A7 FPGA development board. This contains a Cray CPU core running at 105 MHz, the actual clock speed of Cray’s J90. The C90, mind you, runs at 244 MHz. The front-end processor is a Teensy 3.6 microcontroller, controlling reset signals and the SPI interface going into the Cray CPU. The Teensy MCU also drives the OLED display.

Next, Fenton needed software on the watch. Since it was a for-fun project and part of what he calls his “series in Cray-related computational necromancy,” the electrical engineer pretty much threw usefulness out the window. His work and interests have made him enjoy doing N-body gravity simulations, so that’s what he did with the Cray C90 superwatch. N-body gravitational simulations use mathematical formulas to demonstrate, in this case, how a planet and its moons interact gravitationally.

In Python, he wrote an n-body simulation of Jupiter and 63 of its moons. Mimicking Cray vector instructions, he developed a program for the watch that shows a free-running simulation of Jupiter and 63 of its moons orbiting the gas giant. The coordinates come from NASA’s Horizons system. The software displays each moon based on that data in the ecliptic plane around a circle designating Jupiter.

The maker says that using his supercomputer wristwatch “should be as incomprehensible as my motivation for creating it in the first place,” and it sounds like he hit that mark precisely. To actually tell the time on the wristwatch, you look at the positions of Jupiter’s moons. Because all of us know where those lunar bodies will be at particular times of the day...

The final product is programmable, and it looks like a miniature version of a Cray supercomputer. Fenton says it “pushes the boundaries of usefulness and complexity.” I would tend to agree, adding that I’d be nervous about wearing it — the watch appears to be several inches tall and I can see myself whacking it on doorframes every time I walk through one. Then again, I've got to admit it looks really cool.

Ahead of NASA's upcoming launch of the Grace Roman Space Telescope for the "Roman Mission," researchers are using the Argonne Theta supercomputer to run OpenUniverse-powered simulations of the cosmos. The simulation is being run for the to-be-launched Grace Roman Space Telescope and the grounded Chilean Vera C. Rubin Observatory. According to Jim Chiang, who helped create the simulations, "OpenUniverse lets us calibrate our expectations of what we can discover with these telescopes [...by giving] us a chance to exercise our processing pipelines, better understand our analysis codes, and accurately interpret the results so we can prepare to use the real data right away once it starts coming in."

As cutting-edge and high-concept as this may sound, OpenUniverse is an open-source Solar System simulator leveraging OpenGL that has existed for about 24 years. The classic version of OpenUniverse also inspired other planetarium software.

Of course, NASA's implementation of OpenUniverse in 2024 is much more ambitious than the standard version since it's intended for hard science. The data used by NASA's OpenUniverse 2024 has been released as a 10-terabyte subset of the complete package, with the remaining 390 terabytes still to be processed at the time of writing.

Katrin Heitmann, cosmologist and deputy director of Argonne's High Energy Physics division and the one who managed the project's supercomputer time, stated, "Using Argonne's now-retired Theta machine, we accomplished in about nine days what would have taken 300 years on your laptop. The results will shape Roman and Rubin's future attempts to illuminate dark matter and energy while offering other scientists a preview of the types of things they'll be able to explore using the data from the telescopes."

NASA's official post on the matter refers to the project as "A Cosmic Dress Rehearsal" for the researchers involved, one year ahead of the Rubin Observatory telescope's activation in 2025 and three years ahead of the NASA Roman launch in May 2027. In particular, Roman and Rubin are both meant to help us achieve a fuller understanding of the dark energy that expands our universe and the dark matter that helps fill it.

IEEE Spectrum reports three new supercomputers powered by Nvidia’s Grace Hopper CPU+GPU chips topped the Green500 for June 2024. The list ranks the most efficient supercomputers in the world. The three new supercomputers that broke into the list are Jülich Supercomputing Centre’s JEDI, the 1 ExaFLOPS Jupiter Supercomputer, University of Bristol’s Isambard-AI phase 1, and Cyfronet’s Helios GPU.

The Green500 is the energy efficiency counterpart to the Top500, which lists supercomputers according to their power. Most of the powerful supercomputers in the world remain unchanged, with Oak Ridge National Laboratory’s Frontier, Argonne National Laboratory’s Aurora, and Microsoft Azure’s Eagle still taking the top spots. However, these supercomputers are power-hungry, with the top computer requiring at least 22,000 kilowatts to run — equal to the power requirements of more than 15,700 households.

On the other hand, JEDI has a power rating of 67.31 kW (73 GFlops/watt), Isambard-AI Phase 1 requires 117.08 kW (68 GFlops/watt), and Helios needs 316.88 kW (67 GFlops/watt). These devices aren’t in the top 10 regarding raw computing performance, but they aren’t that far behind. JEDI is ranked at 189, while Isambard reached 128. Helios GPU is the most powerful of the bunch, sitting in the 55th position. However, their combined power requirements are still dwarfed by the Aurora, which requires 38,698 kW (26.15 GFlops/watt) and is the second most powerful supercomputer.

All three new entrants to the Green500 are powered by Nvidia’s Grace Hopper super chips, which use the Arm architecture and are inherently more efficient than traditional x86 systems. Aside from the top three systems, other Grace Hopper-powered supercomputers also made the top 10 in the Green500 list—preAlps (5) and Venado (8) also made it, making Nvidia the preferred supplier of 5 out of the ten most efficient supercomputers. This demand likely helped the company reach a record $26 billion 24Q1 revenue.

The entry of these new power-efficient supercomputers shows the industry’s changing priorities. Instead of adding more powerful chips with greater kilowatt requirements, many institutions are now going for better efficiency—i.e., getting more computing performance per watt. This is a crucial move for the general industry, especially as the Semiconductor Industry Association estimated in 2015 (PDF) that computing’s power requirements will exceed global production by 2040 if efficiency isn’t improved.

Because of this increasing power demand, Microsoft is looking at nuclear reactors to power its data centers. Meta CEO Mark Zuckerberg also said that power constraints will bottleneck AI progress. “A lot of data centers are on the order of 50 megawatts or 100 megawatts, or like a big one might be 150 megawatts,” Zuckerberg said. He then added later, “But then when you start getting into building a data center that’s like 300 megawatts or 500 megawatts, or a gigawatt; I mean, no one has built a single gigawatt data center yet, so I think it will happen. I mean, this is only a matter of time.”

Newer technologies, like Nvidia’s Grace Hopper chips and LPDDR5X memory, are helping make computing more efficient and sustainable. While we need to reduce energy consumption, especially that of non-renewable energy, to help combat the effects of global warming, we also do not want to reduce our computing power. However, with these more power-efficient processors, we can keep up with the demands of data centers and AI computing without needing more energy.

Nvidia said that its Grace Hopper GH200 platforms — consisting of one 72-core Grace processor and an H100 GPU for AI and HPC workloads — have been adopted for nine supercomputers across the globe. These supercomputers collectively achieve a staggering 200 ExaFLOPS of 'AI' computational power, though their FP64 computational performance needed for scientific simulations is significantly lower.

New installations span several countries, including France, Poland, Switzerland, Germany, the United States, and Japan. Among the standout systems is the EXA1-HE in France, developed in collaboration between CEA and Eviden. This supercomputer is equipped with 477 compute nodes that rely on Nvidia's Grace Hopper processors and deliver substantial performance. EXA1-HE is based on Eviden’s BullSequana XH3000 architecture and features a warm-water cooling system to enhance energy efficiency. 

Another important Grace Hopper-based machine is of course the Isambard-AI project at the University of Bristol in the U.K., which was announced in late 2023. Initially equipped with 168 Nvidia GH200 Superchips, Isambard 3 stands as one of the most energy-efficient supercomputers to date. Upon the arrival of an additional 5,280 Nvidia Grace Hopper Superchips this summer, the system's performance is expected to increase approximately 32 times, which will turn the machine into one of the fastest supercomputers for AI-driven scientific research in the world. When fully built, the system will feature over 55,000 high-performance Arm Neoverse V2 cores, which promise to deliver quite formidable FP64 performance.

Other noteworthy machines include the Helios supercomputer at Academic Computer Centre Cyfronet in Poland, Alps at the Swiss National Supercomputing Centre, Jupiter at the Jülich Supercomputing Centre in Germany, DeltaAI at the National Center for Supercomputing Applications at the University of Illinois Urbana-Champaign, and Miyabi at Japan’s Joint Center for Advanced High Performance Computing. These systems vary in specialization but all incorporate the latest Nvidia platform to drive forward their respective scientific agendas. 

The key thing about the announcement is that Nvidia's Grace Hopper platform powered by the company's own CPU and GPU technologies is gaining traction in the scientific world. While Nvidia almost controls the market of AI GPUs, getting into supercomputers is also very important for the company as HPC is a particularly sizable and lucrative business.

Today, a slew of supercomputing entities submitted their newest benchmark test results to the Top500 committee to compete for the top spot on the list. The Intel-powered Aurora supercomputer was widely expected to take the top spot from the AMD-powered Frontier, the #1 supercomputer on the Top500 list, but it took second place instead. However, Aurora did take the top spot in the AI-centric HPL-MxP mixed-precision benchmark, allowing Intel to lay claim to powering the fastest AI supercomputer in the world with 10.6 AI Exaflops of performance. 

It's noteworthy that Aurora is still not fully operational, so the entire system wasn't used for any of the benchmark submissions. Aurora remains beset by numerous hardware issues like hardware and cooling system failures, operational errors, and network instability, among others (details in the last section below). The continued issues are a bit surprising—the system was first announced nine years ago, the second revision was announced five years ago (the first version was canceled), and the final components were installed eleven months ago. 

The system now houses the full complement of 21,248 CPUs and 63,744 GPUs spread across 10,624 compute blades, but Argonne National Laboratory (ANL), which hosts the system, was again unable to submit a full-system Linpack run for the Top500 list. 

Instead, Aurora placed second with 1.012 Exaflops, breaking the Exaflops barrier with 87% of the system active (9,234 of the full 10,624 nodes). This solidifies Aurora's second-place position — Aurora's first submission (with only half the system) also took second place, reaching 585.34 petaflops six months ago.

Aurora is supposed to be faster than Frontier in the High-Performance Linpack (HPL) benchmark and thus take the lead in the Top500 upon completion, but it's clear the system will need more tuning to live up to its billing. Frontier is ~19% faster than Aurora with 1.206 exaflops of performance, and, assuming linear scaling, Aurora still wouldn't win after adding the remaining 13% of nodes that weren't used for the Top500 benchmark run.

Intel has ballyhooed Aurora's theoretical peak performance of 2 exaflops (Rpeak), but supercomputers are measured by sustained performance (Rmax). Frontier delivers 70% of its peak as sustained performance in Linpack, while Aurora only delivers 51% of its peak. This should hopefully improve over time, and Aurora would easily take the top spot if it delivered a similar 70% of its peak performance (~1.4 exaflops) during sustained workloads.

I asked ANL if Aurora is expected to take the lead over Frontier in the Top500 upon completion. "There's a contractual target number that is faster than Frontier," a representative responded. "So, if we're successful in reaching that number, we'll be faster than Frontier." Notably, the statement says Aurora should beat Frontier, not that it will. We've followed up for a firm confirmation of the actual performance target.

Aurora took first place in the HPL-MxP mixed-precision benchmark with 10.6 exaflops of AI performance with only 89% of the Aurora system active. This benchmark prioritizes lower precision (FP32 and lower, even FP16) than the FP64 used for the Linpack benchmark used for the Top500 ranking. Thus, this benchmark better represents AI workloads and an increasing number of other real-world applications — FP64 is largely relegated to traditional scientific computing, and some argue it is a shrinking portion of that segment, too.

HPL-MxP is becoming much more important to model real-world performance in the age of AI, but Aurora's position at the top will be hotly contested. There has yet to be a submission from a large-scale Nvidia Grace-Hopper-powered system to the leaderboard. The Alps supercomputer, which now promises 20 exaflops of AI performance, is slated to have all of its 10,752 Grace Hopper processors installed by the end of June 2024, so competition for the leadership spot is on the way.  

The High Performance Conjugate Gradients (HPCG) benchmark is also designed to be more representative of real workload applications than Linpack. Aurora performed impressively in this benchmark as well, taking the #3 ranking with a mere 38.5% of the supercomputer active. Aurora also took fifth in the Graph500 benchmark, which is designed to measure performance in data-intensive applications, but ANL didn't specify how much of the system was active for this benchmark run. 

Aurora hasn't placed in the Green500, a list of the most power-efficient supercomputers, and that isn't surprising. Aurora will consume up to 60 MW of peak power, slightly more than double Frontier's 29 MW, but we don't know how its final performance will look. It isn't clear if Aurora can beat Frontier in Linpack performance, but even if it does win, it will be by a small amount—certainly not enough to justify the increased power consumption for that particular workload. However, there are plenty of other applications that operate at lower precisions, and power efficiency comparisons will vary by application. Regardless, Nvidia's Grace Hopper systems now comprise five of the top ten systems on the Green500, so it appears that Nvidia has both Intel and AMD beat in the power efficiency department. 

Aurora facing hardware failures, cooling system malfunctions, among other problems

Ten long months passed between the final Aurora hardware being installed and when ANL submitted its benchmarks, raising questions about the source of the continued delay in standing up the full machine. We followed up with Intel on the matter.

“[...]Since we completed the physical delivery of the last compute node at the end of June 2023 (only 10 months ago), we have been working hand-in-hand with Argonne National Laboratory and HPE to fully stabilize and tune the system, including the compute nodes, storage system, fabric, power delivery, and cooling."

"We are also actively working on addressing stability issues like hardware failures, software bugs, cooling system malfunctions, issues with power supply, networking infrastructure stability, environmental factors, and operational errors,” the Intel representative said to Tom's Hardware.

Argonne National Laboratories and Intel have yet to provide a firm date for when they expect the system to be fully operational, but we do know that Aurora's window to take the lead in the Top500 is closing. The AMD-powered El Capitan, rated for two exaflops of peak throughput (not sustained), is largely expected to beat Aurora and Frontier in Linpack. Lawrence Livermore Labs submitted early results for sub-scale models of El Capitan today, and the system is expected to be completely installed by the end of 2024.

GigaIO and SourceCode have teamed up to bring the world an ultraportable but still supercomputer-class device for AI needs. The Gryf weighs 55 pounds or less and is packed into a TSA-friendly carry-on suitcase.

Despite the small form factor, Gryf can accommodate data collection and processing on a scale that would otherwise require sending the data offsite. This revolutionary development for use cases requires quick processing and analysis turnaround.

Gryf is a suitcase-sized supercomputer that supports disaggregating and reaggregating its GPUs. The user can customize the computer's hardware configuration in the field on the fly. You can create the optimal hardware configuration for one assigned workload and then change it for the next.

Each Gryf contains multiple slots populated with compute, storage, accelerator, and network sleds tailored to the workload. The suitcase-sized supercomputer has six sled slots to insert and remove modules from as needed. 

For AI or ML workloads, for example, you might plug in two compute sleds, an accelerator sled, two storage sleds, and a network sled. Are you moving on to a storage project? Change the configuration to incorporate one compute sled and five storage sleds instead.

The specifications and capabilities for each type of sled and the Gryf platform itself are as follows:

According to GigaIO and SourceCode, a single Gryf can be configured to process over a petabyte of information. Using GigaIO’s FabreX memory fabric, the Gryf can also be stacked with up to four other Gryfs for more demanding workloads.

Once back in the data center, the FabreX memory fabric allows Gryf to connect to the core computer, a GigaPod, for the more demanding processing and analysis tasks. Rather than waiting for days to transmit the data over internet connections, engineers collect and begin processing the data where it was gathered, then cart it off to the data center.

Citing the needs of Department of Defense customers, GigaIO CEO Alan Benjamin pointed out the need to collect and process data where it happened in the field.

“This is true for our Department of Defense customers, who have emphasized the critical need for timely and actionable intelligence in the field. Gryf’s novel architecture, made possible by FabreX, our AI memory fabric, provides those customers the advanced compute, storage, and GPU capabilities they crave in today’s sensor-rich edge environments.”

The companies did not disclose pricing for the Gryf, but the mobile supercomputer is available for purchase now.

Tech gadget builder bitluni recently unveiled his latest build, a 256-core RISC-V megacluster. The YouTube content creator demonstrated the miniature supercomputer in a video that highlighted the design, production, assembly, and testing phases. Not everything went as planned, but the results are actually pretty nerdy and cool.

For the design, bitluni combined 16 RISC-V superclusters into a single megacluster. Each supercluster consisted of 16 CH32V003 RISC-V microcontrollers connected with an 8-bit bus. Each supercluster contains its own LED because bitluni wanted to be able to display a line of text.

To deal with potential problems inherent in building an extremely large (and power-hungry) single PCB for the megacluster, bitluni installed the superclusters in pairs on eight “cluster blade” designs. The inventor used two additional CH32V203 microcontrollers on each blade to serve as a bridge between each supercluster and the main 8-bit megacluster bus.

In the video, you can watch the assembly in time-lapse as bitluni attaches the microcontrollers to the PCB, solders the GPIO headers, and puts the entire megacluster together for testing. The blades all attach to a single main board, and as bitluni demonstrated first the LEDs blinking, he also found one of the mistakes he made in the design.

Without an internal clock source, bitluni discovered his LEDs blinked out of sync with each other. So, even though the lights initially blinked simultaneously, they quickly lost that synchronicity and flashed randomly. (Still, it’s almost hypnotic to watch.)

The creator went on to describe how to programmatically deal with bus collisions as all the microcontrollers talked at once.

Ultimately, bitluni’s megacluster used 256 RISC-V microcontrollers at 48 MHz and 17 RISC-V chips at 144 MHz. It includes 640 GPIO pins and 256 ADC circuits. He describes the 14.7 GHz combined single-core clock rate as "not that impressive but also not too shabby."

Considering bitluni designed and built a supercomputer that is small enough to sit on a TV tray, that description may be a bit of an understatement. Let’s also not forget that what he built cost far less than you would spend on even the used and leaky Cheyenne supercomputer.

One lucky buyer had their day yesterday as the U.S. government's online auction for its Cheyenne supercomputer ended at $480,085. The auction, first covered by Tom's Hardware on Wednesday, has closed after 27 bidders contended for a piece of supercomputing history.

The Cheyenne supercomputer's six-figure sale price comes with 8,064 Intel Xeon E5-2697 v4 processors with 18 cores / 36 threads at 2.3 GHz, which hover around $50 (£40) a piece on eBay. Paired with this armada of processors is 313 TB of RAM split between 4,890 64GB ECC-compliant modules, which command around $65 (£50) per stick online. For a deeper dive into Cheyenne's components and prime performance, check out our initial sale coverage. Unfortunately for buyers, none of the Cheyenne supercomputer's 32 petabytes of high-speed storage are being sold with the lot. Still, a savvy eBay seller could flip the processors and RAM across the machines for around $700,000 (£550,000), making a hefty profit. 

This estimate assumes all of the CPUs and RAM sticks work, which is invalid. The auction site disclaims that "approximately 1% of nodes experienced failure [over the last six months], primarily attributed to DIMMs with ECC errors, which will remain unrepaired." This high failure rate is the reason for the sale, with upkeep made challenging by the fact that Cheyenne "is currently experiencing maintenance limitations due to faulty quick disconnects causing water spray." These two compounding issues have cost the Cheyenne team undue downtime and repair costs, leading them to look for a replacement, the government said. 

The Cheyenne supercomputer array had a rigorous workflow over seven years of operation to earn its wear and tear. The array's services were used by scientists across the state of Wyoming and the rest of the country when needed. The 5.34-petaflop system was mainly used for weather and climate studies, helping the National Science Foundation better study climate change and other Earth-related sciences. 

It isn't easy to provide an analysis of how much the government stands to lose with the sale of Cheyenne. We could not find data for the price of Cheyenne at construction, but its predecessor, Yellowstone, was quoted at $25-35 million before its construction. Some of this cost may be attributed to buildings and infrastructure surrounding the supercomputer. Still, with Cheyenne's replacement, the Derecho, costing $35-40 million from HP, Cheyenne likely initially cost around this 8-figure range as well. At a conservative estimate of $25 million for the array to be constructed, the auction sale is only 2% of the sticker price of Cheyenne. Good luck finding a 98% off sale at the following government surplus sale.

The buyer will have the joy of moving Cheyenne's 30 server racks (28 processing racks, two air-cooled management racks) out of the facility themselves; the government is not providing transport or including any Ethernet or optical cabling needed to get the machine up and running. What use a buyer might have for a malfunctioning and outdated supercomputer array is known only to them, but if you see a flood of Xeon chips hitting the market, you'll have a pretty good idea why.

What was once the 21st most powerful supercomputer in the world is now available to the highest bidder — well, maybe, as the current bid of under $30,000 has not met the required amount. The U.S. General Services Administration opted to put the Cheyenne supercomputer, deployed in 2016, up for auction, in part due to ongoing repair and maintenance problems.

The retired supercomputer is, as the name suggests, a monster. It’s a 5.34 petaflops system, one of the last deployed by Silicon Graphics International after its acquisition by Hewlett-Packard. Since then, it's been a cornerstone of operations at the NCAR-Wyoming Supercomputing Center in Cheyenne, Wyoming.

The Cheyenne supercomputer is a water-controlled installation made up of SGI ICE XA modules with 28 racks holding 8,064 Intel E5-2697v4 CPUs. That totals 145,152 cores, for those keeping count. The main system is spread across 4,032 dual-socket nodes. Here are the specs of the primary components:

Each E-Cell weighs in at 1,500 pounds, and shipping is not included in the winning bid. The purchaser will need to hire a professional moving company to transport the supercomputer from the facility to its new home. The auction notes also state that the supercomputer will be sold as-is, and that it "is currently experiencing maintenance limitations due to faulty quick disconnects causing water spray." Not exactly the pinnacle of supercomputing achievements, then.

Beyond the above hardware, the supercomputer also includes two air-cooled management racks. These consist of 26 1U servers each, 20 of which have 128GB of memory and six with 256GB of memory. That's an additional 8TB of DRAM, if you're wondering. The management racks also include 10 Extreme Switches, and two Extreme Switch power units, and each rack weighs 2,500 pounds.

While the Cheyenne supercomputer has been in operation for the past seven years, the auction notes says the "expense and downtime associated with" fixing the current cooling problems makes it unworthy of continued maintenance. And of course, even though this was a lightning-fast supercomputer when it first launched, it would be considered sluggish by 2024 computing standards. This is a fate shared by many supercomputers, even some of NASA's most powerful ones.

Cheyenne peaked at number 21 on the Top500 list of the most powerful supercomputers back when it launched. Today, it sits at number 160 — based on an Rmax score of 4.79 petaflops. The paradigm shift to GPU-powered supercomputers over the past decade means that, as an example, you could potentially exceed that level of performance with around 23 Nvidia DGX H100 systems sporting 46 CPUs and 184 GPUs.

Even so, the auction comes with a treasure trove of parts and components for whoever is willing to pony up the cash. The supercomputer will be drained for removal, and it seems it won't necessarily include all the necessary cabling. However, it does include a whopping 313,344GB of DDR4-2400 ECC RAM. That alone could be worth more than $350,000 — not to mention an unspecified amount of storage.

Also of interest is that the supercomputer uses around 1.727 MW of power when fully assembled. Which means that if you want to power it up and run complex simulations on it, the power requirements could cost over $4,000 per day (depending on the price of electricity, naturally).

We presume most bidders would be more interested in parting out the system rather than attempting to get it running again. Besides the missing Ethernet and optical cabling that you'd need to acquire, there's the apparently unresolved issue of the leaking quick-connect liquid cooling components. But who knows? Maybe some enterprising business will find a way to bring Cheyenne back into service, like a phoenix rising from the water-logged ashes.

Updated: Bidding is scheduled to end on May 3, 2024 and is up to $28,085 $120,085 now. There's no longer a "reserve not met" disclaimer, which apparently means someone is going hope with several truckloads of old supercomputer parts. Whether the buyer will try to fix up the system and get it running, sell it for parts, or just grind it up for the raw materials remains to be seen.

AMD plans to introduce upgraded versions of its Instinct MI300-series processors for AI and HPC in the second half of the year, according to market research firm TrendForce. The revamped Instinct MI350-series products are expected to adopt a more sophisticated process node as well as an upgraded memory configuration.

AMD's Instinct MI350-series products will employ chiplets made on TSMC's 4nm-class fabrication process, which is an enhanced version of TSMC's 5nm-class technology, which is used to build CDNA 3 and Zen 4 chiplets for AMD's Instinct MI300-series products. Using TSMC's N4 production node will allow AMD to either increase performance or lower power consumption in its Instinct MI350-series processors compared to Instinct MI300-series processors. This is largely speculation for now. 

AMD's chief technology officer Mark Papermaster said recently that the company was prepping Instinct MI300-series products with a revamped memory configuration. AMD currently has its Instinct MI300A, which has 128GB of memory using 8Hi HBM3 stacks, and the Instinct MI300X AI accelerators, which have 192GB of memory using 12Hi HBM3 stacks. AMD could potentially adopt 12Hi HBM3E stacks to increase the memory bandwidth and expand the memory capacity of the improved Instinct MI350-series products — but, again, this is speculation.

As demand for AI processors and accelerators is quickly increasing, developers such as AMD, Intel, and Nvidia have been feeling more confident about expanding their product lineup with refreshes.  

Nvidia, the undisputed leader in the market of AI accelerators, was the first to refresh its existing compute GPU, the H100 80 GB, with the H200 141 GB HBM3E, which adopted HBM3E memory instead of HBM3, enabling customers to train larger LLMs.  

If the information from TrendForce is accurate, AMD also feels confident enough not only to add faster memory to its CDNA 3-based chiplets, but also to make these chiplets on a more advanced fabrication process (albeit, one that belongs to the same process design kit) to improve their performance potential. As a result, AMD's Instinct MI350-series products will be considerably more competitive against Nvidia's H200 than its Instinct M300 processors are. However, AMD will ship them after Nvidia's H200 hits the market.

NASA works with some of the world's most advanced technologies and makes some of the most significant discoveries in human history. However, according to a special report conducted by the NASA Office of Inspector General and discovered by The Register, NASA's supercomputing capabilities are insufficient for its tasks, which leads to mission delays. NASA's supercomputers still rely primarily on CPUs, and one of its flagship supercomputers uses 18,000 CPUs and 48 GPUs.

NASA currently has five central high-end computing (HEC) assets located at the NASA Advanced Supercomputing (NAS) facility in Ames, California, and the NASA Center for Climate Simulation (NCCS) in Goddard, Maryland. The list includes Aitken (13.12 PFLOPS, designed to support the Artemis program, which aims to return humans to the Moon and establish a sustainable presence there), Electra (8.32 PFLOPS), Discover (8.1 PFLOPS, used for climate and weather modeling), Pleiades (7.09 PFLOPS, used for climate simulations, astrophysical studies, and aerospace modeling, and Endeavour (154.8 TFLOPS).

These machines almost exclusively use old CPU cores. For example, all NAS supercomputers use over 18,000 CPUs and only 48 GPUs, and NCSS uses even fewer GPUs.

"HEC officials raised multiple concerns regarding this observation, stating that the inability to modernize NASA's systems can be attributed to various factors such as supply chain concerns, modern computing language (coding) requirements, and the scarcity of qualified personnel needed to implement the new technologies," the report says. "Ultimately, this inability to modernize its current HEC infrastructure will directly impact the Agency's ability to meet its exploration, scientific, and research goals."

The audit conducted by NASA's Office of Inspector General also revealed that the agency's HEC operations are not centrally managed, resulting in inefficiencies and a lack of a cohesive strategy for using on-premises versus cloud computing resources. This uncertainty has led to hesitancy in using cloud resources due to unknown scheduling practices or assumed higher costs. Some missions have resorted to acquiring their infrastructure to avoid waiting for access to the primary supercomputing resources, which are overwhelmed to a large degree because they do not rely on the latest HPC technologies.

Additionally, the audit found that the security controls for the HEC infrastructure are often bypassed or not implemented, increasing the risk of cyber attacks.

The report suggests that transitioning to GPUs and code modernization is essential for meeting NASA's current and future needs. GPUs offer significantly higher computational capabilities for workloads involving parallel processing, which are very common in scientific simulations and modeling.

Recently, Nvidia launched its H200 with 141GB of HBM3E and stole quite a lot of thunder from AMD; Nvidia's latest came just after AMD had released its Instinct MI300A with 128GB of HBM3 memory and the Instinct MI300X AI accelerators with 192GB of HBM3 memory. However, AMD may be prepping new versions of Instinct MI300-series products with more memory. 

Mark Papermaster, chief technology officer of AMD, implied this possibility during his presentation at the Arete Investor Webinar Conference last week.

"We are not standing still," said Papermaster (via SeekingAlpha). "We made adjustments to accelerate our roadmap with both memory configurations around the MI300 family, derivatives of MI300, the generation next. […] So, we have 8-Hi stacks. We architected for 12-Hi stacks. We are shipping with MI300 HBM3. We have architected for HBM3E." 

It should be noted that AMD's Instinct MI300A is equipped with 128GB of HBM3 memory using 8-Hi known good stack die (KGSD), whereas the Instinct MI300X comes with 192GB of memory and uses 12-Hi KGDSs. For now, none of AMD's products use HBM3E, but it looks like the company could use it for some of the future versions of AMD's Instinct MI300-series products.

By using HBM3E instead of HBM3, AMD can significantly increase the memory bandwidth available to its AI and HPC GPU or APU. Changing memory devices attached to its processors is a fairly easy thing to do, so expect AMD to use HBM3E rather sooner than later. 

According to AMD's performance numbers, the Instinct MI300X outperforms Nvidia's H100 80GB GPU, which is widely used by major hyperscalers such as Google, Meta (Facebook), and Microsoft. The Instinct MI300X is also likely to be a strong contender against Nvidia's upcoming H200 141GB GPU. Meanwhile, it remains to be seen what to expect from AMD's hypothetical Instinct MI300-series with HBM3E memory.

"We have been very closely working with our customers, and what we can tell you this, I willl tell you right now: that race has begun, and it is going to be a competitive race," said Papermaster. "You are going to see that back and forth like you have always seen when you have competition. It is going to be great for the market. It is certainly spurring us at AMD to be at our very best of innovation. […] Stay tuned with us as we will share more details forthcoming on that roadmap because it is indeed a multiyear roadmap that we have laid out."

One of the mysteries surrounding China's Tianhe 3 (Xingyi) supercomputer is what hardware is used to build it. A recent article by The Next Platform has shed light on the MT-3000 processor designed by the National University of Defense Technology (NUDT). As it turns out, the MT-3000 features a unique heterogeneous architecture that includes general-purpose CPU cores, control cores, and matrix accelerator cores.

NUDT's MT-3000 processor features a multi-zone structure that packs 16 general-purpose CPU cores with 96 control cores and 1,536 accelerator cores, according to The Next Platform. The MT-3000 processor reportedly achieves 11.6 FP64 TFLOPS of peak performance and demonstrates a power efficiency of 45.4 GigaFLOPS/Watt at an operational frequency of 1.20 GHz. 

The key aspect of the MT-3000 architecture is that it packs both general-purpose and matrix acceleration cores into the same piece of silicon. To some degree, this integration mirrors the design philosophy behind AMD's Instinct MI300A CPU-GPU hybrid, suggesting a shift away from conventional discrete CPU-GPU systems towards more cohesive and efficient designs. Meanwhile, unlike AMD's multi-chiplet Instinct MI300A, the MT-3000 looks to be a monolithic design.

Being a heavily packed CPU, the MT-3000 has to be made on an advanced process technology, which could range from 14nm to 10nm down and to potentially 7nm technologies, The Next Platform suggests. The chip is likely made by Chinese foundry SMIC, which also produces a rather advanced HiSilicon Kirin 9000S processor for Huawei using its second-gen 7nm process technology. Meanwhile, it is unclear whether SMIC has enough 7nm production capacity to make chips both for Huawei and NUDT. As a result, it is possible that the MT-3000 uses a different production node. 

With the MT-3000 at its core, the Tianhe-3 is believed to achieve unprecedented computational performance, potentially reaching 1.57 ExaFLOPS on LINPACK benchmarks. This projection not only highlights the processor's central role in advancing China's supercomputing capabilities but also signals a significant improvement for the country. In comparison, Frontier, the fastest supercomputer in the US, reaches 1.102 ExaFLOPS of performance.

The MT-3000 processor represents a leap forward in high-performance computing (HPC) technology for China. With its hybrid architecture, high-performance efficiency, and potentially a very sophisticated production node, the MT-3000 looks to be a competitive chip, potentially positioning China at the forefront of global HPC development. 

Intel has teamed up with Japan's NTT and South Korea's SK hynix to develop next-generation chips featuring optical technology that communicate via photons (light) instead of using electrons, according to a report from Nikkei. The companies reportedly obtained financial backing for their joint work, which focuses on integrating silicon photonics and similar technologies into semiconductors. 

The collaboration between Intel, NTT (a leading Japanese telecom carrier), and SK Hynix (a leading memory maker) leverages the specific strengths of each partner. Intel is certainly an expert in processors, whereas SK Hynix specializes in high-performance memory chips. NTT's role is to coordinate the collaboration between the two chipmakers, the report says.

Meanwhile, given the fact that the three companies make completely different products, it looks like they will mostly focus on the silicon photonics optical technology in general (e.g., materials, integration, etc.) as well as aspects like the manufacturability of chips featuring silicon photonics.

The collective goal is to develop production technology for devices that incorporate optical technology into logic chips and memory technology capable of handling terabit-class data speeds by fiscal 2027 — aiming to achieve a 30% - 40% reduction in power consumption compared to conventional products.

Processors and memory with optical connections will be particularly beneficial for artificial intelligence (AI) and high-performance computing (HPC) applications that have to process vast amounts of data. Currently, optical communications are converted into electrical signals before they reach processors or memory. The new technology proposes to integrate optical technology directly into processors to improve performance and save power. 

The Japanese government's substantial financial backing of approximately ¥45 billion ($305 million) certainly indicates the national significance of this initiative, which could help to revitalize Japan's position in the global semiconductor supply chain. It remains to be seen which chipmaker in Japan will actually produce silicon photonics chips in the country.

For now, Intel, NTT, and SK hynix have not commented on the collaboration, possibly because the details have yet to be finalized. It is not uncommon for NTT to work with foreign entities on promising technologies as the company's business lies far beyond telecommunications. 

Intel recently sold its silicon photonics business to Jabil.

China Telecom claims it has built the country's first supercomputer constructed entirely with Chinese-made components and technology (via ITHome). Based in Wuhan, the Central Intelligent Computing Center supercomputer is reportedly built for AI and can train large language models (LLM) with trillions of parameters. Although China has built supercomputers with domestic hardware and software before, going entirely domestic is a new milestone for the country's tech industry.

Exact details on the Central Intelligent Computing Center are scarce. What's clear so far: The supercomputer is purportedly made with only Chinese parts; it can train AI models with trillions of parameters; and it uses liquid cooling. It's unclear exactly how much performance the supercomputer has. A five-exaflop figure is mentioned in ITHome's report, but to our eyes it seems that the publication was talking about the total computational power of China Telecom's supercomputers, and not just this one.

We probably can't expect official performance benchmarks any time soon either, as China is neglecting to submit its supercomputers to TOP500, the organization that tracks the 500 fastest supercomputers in the world. This caginess is apparently down to fears about getting too much attention and inviting even more U.S. sanctions. 

It's hard to guess what might be inside this supercomputer, given the lack of details. On the CPU side of things, it may use Zhaoxin's KaiSheng KH-40000 server CPUs, which are now available in domestically-made servers. There are also other candidates though, like Loongson's 32-core 3D5000 and Phytium's 64-core Feiteng Tengyun S2500. All three chips differ greatly in respect to architecture, with Zhaoxin using x86 like Intel and AMD. Loongson uses a derivative of MIPS, and Phytium runs Arm's architecture.

Similarly, there are plenty of options for Chinese-made GPUs, with possibilities ranging from Moore Threads, Loongson, and Biren. Of the three companies, Moore Threads is the most recent to launch a new GPU in the form of its MTT S4000, which is already planned to see use in the KUAE Intelligent Computing Center. Loongson's LG200 arrived about two weeks before the S4000, though its claimed performance would make for a very slow supercomputer. Biren's BR100 would be a heavyweight champion, but it's unclear if it ever returned to production anywhere after TSMC stopped making it due to U.S. sanctions.

Regardless of the actual hardware inside China Telecom's new supercomputer, that it is reportedly made from top to bottom with Chinese hardware is the most important part. Relying solely on Chinese technology likely means the Central Intelligent Computing Center is disadvantaged in some or many areas. But technological independence is a key goal for China, even if it means swapping out cutting-edge Western hardware for slower but natively-made components. U.S. sanctions won't have much of an impact if China can manage to do everything itself.