A man suspected of attempting to smuggle SK hynix tech secrets to China was apprehended at South Korea’s Incheon International Airport earlier this month. DigiTimes reports that the former employee subcontractor aimed to leak the firm’s proprietary high-bandwidth memory (HBM) packaging technology to entities based in China.

According to the source report, authorities had been aware of the activities of suspect ‘Mr Kim’ for several months. The Seoul Metropolitan Police Agency's Industrial Technology Security Investigation Unit made the dramatic airport arrest, moments before flight boarding, and charged Kim with violating South Korea's trade secret and unfair competition laws. Its investigators say that Kim stole critical data about HBM at around the same time he resigned as a subcontractor at SK hynix, at the beginning of the year.

Investigators found 11,000 document images

Korean prosecutors allege that Kim kept printed and photographed records from his time at SK hynix. He went as far as removing SK hynix branding and ‘confidential’ markings from any material he captured, it is claimed. In total, reports say Kim was caught with a haul of 11,000 images from his time at SK hynix. Boldly, Kim is even said to have cited some of these stolen documents in job applications to Chinese firms, including Huawei’s HiSilicon.

Above, we mentioned that the stolen proprietary data was mostly about HBM technology. More precisely, the almost-leaked SK hynix secrets were focused on the backend packaging stage of HBM, and hybrid bonding techniques, says reports.

HBM is widely used in AI accelerators, so it is understandable that Chinese rivals would covet SK hynix’s work on fine-tuning the production of this kind of computer memory.

The case of Mr Kim is possibly only the tip of the iceberg, and there have been several other high-profile tech leaks to China involving rival companies like Samsung, for example.

South Korea already has reasonably strong deterrents to stem semiconductor IP theft. Individuals can face fines up to the equivalent of US$71,000 and up to 10 years behind bars. Harsher penalties and sentences can be imposed where strategic sectors fall victim to IP theft. Nevertheless, Korean authorities are making efforts to bolster legal deterrents, say reports.

Unable to use leading-edge process technologies to produce its high-end processors for AI, Huawei has to rely on brute force – install more processors than its industry competitors to achieve comparable performance for AI.

To do this, Huawei took a multifaceted strategy that includes a dual-chiplet HiSilicon Ascend 910C processor, optical interconnections, and the Huawei AI CloudMatrix 384 rack-scale solution that relies on proprietary software, reports SemiAnalysis. The whole system provides a 2.3X lower performance per watt than Nvidia's GB200 NVL72, but it still enables Chinese companies to train advanced AI models.

At glance

Huawei's CloudMatrix 384 is a rack-scale AI system composed of 384 Ascend 910C processors arranged in a fully optical, all-to-all mesh network. The system spans 16 racks, including 12 compute racks housing 32 accelerators each and four networking racks facilitating high-bandwidth interconnects using 6,912 800G LPO optical transceivers.

Unlike traditional systems that use copper wires for interconnections, CloudMatrix relies entirely on optics for both intra- and inter-rack connectivity, enabling extremely high aggregate communication bandwidth. The CloudMatrix 384 is an enterprise-grade machine that features fault-tolerant capabilities and is designed for scalability.

In terms of performance, the CloudMatrix 384 delivers approximately 300 PFLOPs of dense BF16 compute, which is nearly two times the throughput of Nvidia’s GB200 NVL72 system (which delivers about 180 BF16 PFLOPs). It also offers 2.1 times more total memory bandwidth despite using HBM2E and over 3.6 times greater HBM capacity. The machine also features 2.1 times higher scale-up bandwidth and 5.3 times scale-out bandwidth thanks to its optical interconnections.

However, these performance advantages come with a tradeoff: The system is 2.3 times less power-efficient per FLOP, 1.8 times less efficient per TB/s of memory bandwidth, and 1.1 times less efficient per TB of HBM memory compared to Nvidia.

Comparison between Nvidia's GB200 NVL72 and Huawei's CloudMatrix CM384

But this does not really matter, as Chinese companies (including Huawei) cannot access Nvidia's GB200 NVL72 anyway. So if they want to get truly high performance for AI training, they will be more than willing to invest in Huawei's CloudMatrix 384.

At the end of the day, the average electricity price in mainland China has declined from $90.70 MWh in 2022 to $56 MWh in some regions in 2025, so users of Huawei's CM384 aren't likely to go bankrupt because of power costs. So, for China, where the energy is abundant, but advanced silicon is constrained, Huawei's approach to AI seems to work just fine.

HiSilicon Ascend 910C: Huawei goes dual-chiplet

When we first encountered Huawei's HiSilicon Ascend 910C processor several months ago, it was a die shot of its compute chiplet, presumably produced by SMIC, which had an I/O that was supposed to connect it to its I/O die. This is why we thought it was a processor with one compute chiplet. We were wrong.

Apparently, the HiSilicon Ascend 910C is a dual-chiplet processor with eight HBM2E memory modules and without an I/O die that resembles AMD's Instinct MI250X and Nvidia's B200. The unit delivers 780 BF16 TFLOPS compared to MI250X's 383 BF16 TFLOPS and B200's 2.25 - 2.5 BF16 TFLOPS.

Comparison between Nvidia's B200 and Huawei's Ascend 910C

The HiSilicon Ascend 910C was designed in China for large-scale training and inference workloads. The processor is was designed using advanced EDA tools from well-known companies and can be produced using 7nm-class process technologies. SemiAnalysis reports that while SMIC can produce compute chiplets for the Ascend 910C, the vast majority of Ascend 910C chiplets used by Huawei were made by TSMC using workarounds involving third-party entities like Sophgo, allowing Huawei to obtain wafers despite U.S. restrictions. It is estimated that Huawei acquired enough wafers for over a million Ascend 910C processors from 2023 to 2025. Nonetheless, as SMIC's capabilities improve, Huawei can outsource more production to the domestic foundry.

The Ascend 910C uses HBM2E memory, most of which is sourced from Samsung using another proxy, CoAsia Electronics. CoAsia shipped HBM2E components to Faraday Technology, a design services firm, which then worked with SPIL to assemble HBM2E stacks alongside low-performance 16nm logic dies. These assemblies technically complied with U.S. export controls because they did not exceed any thresholds outlined by the U.S. regulations. The system-in-package (SiP) units were shipped to China only to have their HBM2E stacks desoldered to be shipped to Huawei, which then reinstalled them on its Ascend 910C SiPs.

In performance terms, the Ascend 910C is considerably less powerful on a per-chip basis than Nvidia's latest B200AI GPUs, but Huawei's system design strategy compensates for this by scaling up the number of chips per system.

More processors = more performance

Indeed, as the name suggests, the CloudMatrix 384 is a high-density computing cluster composed of 384 Ascend 910C AI processors, physically organized into a 16-rack system with 32 AI accelerators per rack. Within this layout, 12 racks house compute modules, while four additional racks are allocated for communication switching. Just like with Nvidia's architecture, all Ascend 910Cs can communicate with each other as they are interconnected using a custom mesh network.

However, a defining feature of the CM384 is its exclusive reliance on optical links for all internal communication within and between racks. It incorporates 6,912 linear pluggable optical (LPO) transceivers, each rated at 800 Gbps, resulting in a total internal bandwidth exceeding 5.5 Pbps (687.5 TB/s) at low latency and with minimal signal integrity losses. The system supports both scale-up and scale-out topologies: scale-up via the full-mesh within the 384 processors, and scale-out through additional inter-cluster connections, which enables deployment in larger hyperscale environments while retaining tight compute integration.

With 384 processors, Huawei's CloudMatrix 384 delivers 300 PFLOPs of dense BF16 compute performance, which is 166% higher compared to Nvidia's GB200 NVL72. However, all system power (including networking and storage) of the CM384 is around 559 kW, whereas Nvidia's GB200 NVL72 consumes 145 kW.

As a result, Nvidia's solution delivers 2.3 times higher power efficiency than Huawei's solution. Still, as noted above, if Huawei can deliver its CloudMatrix 384 in volumes, with proper software and support, the last thing its customers will care about is the power consumption of their systems.

Last year, Huawei and SMIC made quite a breakthrough with the 7nm-class HiSilicon Kirin 9000S processor which powers Huawei's Mate 60 Pro smartphone. However, this year's Mate 70 Pro handset and its system-on-chip demonstrate limited progress in semiconductor manufacturing in China, as the processor continues to use a 7nm-class process technology reports Bloomberg citing TechInsights. While this may not be a problem for a smartphone, the slow semiconductor progress will likely become an issue for Huawei's AI processors.

Still 7nm

The Huawei Mate 70 Pro is based on the HiSilicon Kirin 9020 processor manufactured by SMIC using its 2nd Generation 7nm-class process technology (also known as N+2). The Kirin 9020 is an enhanced version of its predecessor, with a 15% larger die size (136.6mm²) and a modified layout, aimed at boosting performance and power efficiency. It retains design elements like the 'Hi36C0' and 'GFCV110' packaging markings while adding unique identifiers such as 'WH231203' on the die, which may suggest that this is a revamped version of the previous-generation architecture, not something brand new.

Given the fact that Huawei used to make the original HiSilicon Kirin 9000 processor on TSMC's EUV-enabled N7+ node, that 2020 chip may still have an edge over the 2024 application processor in certain areas.

TechInsights expected Huawei to use its HiSilicon Kirin 9100 processor made on its 5nm-class fabrication process for its new flagship smartphone. However, the company decided to be conservative and continues to use its N+2 instead of N+3 (which is believed to be a 6nm-class production node) or a 5nm-class manufacturing technology.

Without any doubt, the very fact that SMIC can still make chips using its 2nd Generation 7nm-class process in high volume demonstrates that it can do so despite extensive restrictions. However, the lack of significant progress indicates that the pace of SMIC's innovations is slowing down.

SMIC's pace of innovation is slowing down

Huawei had been rumored to be developing a 5nm processor for release this year, according to Bloomberg and TechInsights. However, its reliance on SMIC, which lacks access to advanced lithography tools from ASML due to export restrictions, has stalled progress. As a result, Huawei is unlikely to achieve 5nm production before 2026 (assuming that N+3 will be adopted in 2025), Bloomberg claims.

By that time, TSMC will be ready with its 2nd Generation 2nm-class fabrication technology called N2P with gate-all-around (GAA) nanosheet transistors, as well as with its all-new A16 production node with GAA transistors and backside power delivery network. Essentially, this means that TSMC will be three to four generations ahead of SMIC.

Big problem for AI

For now, Huawei can make its Huawei Mate 70 Pro smartphone based on the Kirin 9020 processor competitive as the SoC delivers decent performance for a handset. However, limitations imposed by SMIC's slow progress with process technologies will play a bigger role in Huawei's Ascend 900-series processors for AI.

The company's next-generation HiSilicon Ascend 910C is projected to rely on SMIC's N+3 node (6nm-class). That processor is expected to be available in 2025 and given the limited advantages of 6nm nodes over 7nm nodes, it is not expected to be significantly faster than the Ascend 910B, which is a simplified version of the Ascend 910 released in 2020 to compete against Nvidia's A100. As a consequence, Huawei's AI capabilities on the processor level will be two generations behind those of Nvidia in 2025, which will ramp up its Blackwell-series processors next year.

Of course, Huawei could try and build enormous AI datacenters using hundreds of thousands or millions of Ascend 910-series processors that it will have in 2025 and train large language models that will be competitive against those trained in the U.S. But for how long Huawei could be competitive this way?

The US recently revoked Huawei's access to Intel processors for its products, and shortly thereafter, a leaked document from Huawei shared by Chinese media revealed that the company had a new plan called the 'Taishan Battle' for PC processors, reports Cailian News Agency. However, in a rare move, Huawei has denied the report and called it 'fake news.' The supposed plan included launching Huawei's HiSilicon Kirin X-series processors and supporting platform this year, and Huawei's rare rebuttal gives us some insight into the company's actual plans. 

The report originally emerged from the Huawei Pollen Club, stating that He Tingbo and Yu Chengdong of Huawei HiSilicon Semiconductor and the Consumer Business Group had announced the plan. This initiative was supposedly aimed at speeding up the development and release of PC processors in the wake of the revoked Intel export licenses that left Huawei without a steady chip supply for its PC lineup.

Huawei officials categorically denied the report, describing it as unfounded. Journalistic investigations supported this by confirming that multiple sources within the company had not received any communications related to PC processors. 

For now, Huawei says it will use processors it has already procured to build its PCs, meaning it will leverage its existing stockpile. However, some of Huawei's computers aimed at the domestic market already use processors from China-based chipmaker Phytium. Furthermore, the company is working with Phytium to unify the software and hardware infrastructure of the China-native Kunpeng and Phytium CPUs. 

Oddly, Huawei strongly denies that it is developing its higher-end CPUs for laptops that would compete with Intel or AMD. That isn't too surprising, though—Huawei has a history of hiding its suppliers, and even its own chip models, from prying Western eyes that might bring about more sanctions on its products.

There are many reasons for Huawei to develop its own PC processors. On the one hand, this will ensure that the company can differentiate itself from other PC makers. On the other hand, this will reduce risks for Huawei's PC division amid tightening U.S. export controls. On May 8th, the U.S. government withdrew export licenses from semiconductor giants Intel and Qualcomm, effectively barring them from selling to Huawei. The Department of Commerce confirmed the revocation of these licenses, impacting undisclosed American companies. Intel has already cut its quarter forecast due to the revocation of the license.

Last month, it was reported that Huawei was developing a next-generation Kirin processor for PCs. The new PC chip is expected to feature eight Arm cores — four high-performance next-gen Taishan V130 cores and four energy-efficient cores. It also features a significantly enhanced Mailiang 920 GPU with 10 clusters, promising a major upgrade over the previous Kirin 9000s. This chip aims to rival Apple's M3 in multi-threading performance and near M2 levels in graphics, supporting up to 32GB of memory, which hints at a 128-bit interface. Moreover, Huawei is said to be considering expanding its Kirin lineup with "Pro" and "Max" versions, including additional cores, a more powerful GPU, and expanded memory capacity. 

Despite having a smaller share of the global PC market than Apple, which means that it may have no economic reasons to develop its own processors, Huawei's development of high-performance Kirin CPUs aligns with China's goal of achieving self-sufficiency in semiconductors. These processors are poised to replace AMD and Intel components in PCs used by government agencies, enhancing national technological independence.

Huawei's latest Pura 70 smartphone, launched earlier this month in China, is powered by the company's latest HiSilicon Kirin 9010 application processor made by China fab SMIC. The chip is based on SMIC's 7nm-class process technology, according to findings by TechInsights reported by Bloomberg. The findings demonstrate that despite the U.S. government's efforts to limit Huawei's ability to develop advanced processors and SMIC to produce them, new chips just keep coming. 

Huawei's HiSilicon Kirin 9010 system-on-chip is produced by SMIC, China's largest contract maker of semiconductors, on its 2nd Generation 7nm-class process technology that is sometimes called N+2, according to TechInsights analysis. This is the very same process technology that SMIC used to make the Kirin 9000s SoC released last August inside the Huawei Mate 60 Pro smartphone. 

TechInsights has yet to make details about the Kirin 9010 processor public, but GSMArena reports that the Kirin 9010 has the same general-purpose CPU core count as the Kirin 9000s: one ultra-high-performance core, three high-performance cores, and four low-power Arm Cortex-A510 cores. The only noticeable difference between the Kirin 9010 and Kirin 9000c is that the former clocks its UHP core at 2.30 GHz, whereas the latter manages to run it at 2.62 GHz. 

The new Pura 70 smartphones also have 16 GB of memory, whereas the flagship Mate 60 Pro has 12 GB. This indicates that the new handset has a reworked memory subsystem, although we have no idea whether the Kirin 9010 application processor is involved. 

The Kirin 9010 name suggests that this chip is a version of the Kirin 9000s, though it is hard to say whether this is a more advanced version (given the lower clock, it does not seem to be more advanced) or just a re-spin aimed at improving yields and maybe refining some features. 

In fact, premium Huawei tends to introduce all-new processors with its Mate-series smartphones that launch in the fall to coincide with the launch of new iPhones. In spring, Huawei used to introduce its P-series high-end handsets (now called Pura) that relied on the SoCs already adopted by Mate-branded smartphones. Therefore, it is unlikely that the Kirin 9010 is HiSilicon's new flagship application processor for mobile phones. 

One of the important questions about the Kirin 9000s and Kirin 9010 SoCs is whether SMIC can produce enough of them. We do not know, but based on market checks data from Jefferies (cited by Bloomberg), Huawei's Pura 70 smartphones were sold out in a couple of days, so apparently, Huawei could not build enough handsets.

Baidu recently told its investors that it would not be able to procure leading-edge GPUs for AI applications in the future due to U.S. export restrictions, reports The Register. However, the company believes its advanced software stack will ensure its competitive position despite using less sophisticated processors.

"In the long run, we may not have access to the most cutting-edge GPU," said Robin Li, chief executive of Baidu, at the company's earnings call this week (via SeekingAlpha). "But with the most efficient homegrown software stack, […] the user experience will not be compromised."

Baidu has not been significantly impacted by the U.S. export restrictions on high-performance AI GPUs and other processors, so it will not be impacted in the short term, according to chief executive Robin Li. The company says its existing reserves of Nvidia's GPUs and current procurement capabilities are sufficient to power its existing AI services, AI-native applications, and at least some future projects. This includes support for ERNIE, Baidu's answer to ChatGPT, which is touted as China's most powerful foundation model.

Baidu has traditionally used Nvidia's high-performance GPUs for training and inferencing of its AI models. Several years ago, the company introduced its own Kunlun II chip, which could rival Nvidia's A100 in AI workloads. However, whether it currently powers a significant portion of the company's services is unclear. The company also reportedly adopted Huawei's Ascend 910-series processors late last year to reduce its reliance on Nvidia.

Now that Baidu cannot procure Nvidia's A100 and H100 processors and will not have access to Nvidia's B100 GPUs when they are released this year, it will have to rely on processors developed and made in China, which are not as fast as those from Nvidia, of course. This is where the company pins a lot of hope on its software stack and innovation.

"There is ample room for innovation in the application layer, the model layer and the framework layer," said Li. "Our end-to-end self developed four-layer AI architecture along with our strong R&D team will support us in using less advanced chips for efficient model training and inferencing. This provides Baidu with a unique competitive advantage over our domestic peers."

Baidu claims to have China's most cost-efficient AI infrastructure for model training and inferencing. This efficiency, coupled with the company's innovative AI architecture, positions Baidu favorably against its domestic competitors. All Chinese tech companies face similar challenges with access to Nvidia's highest-end products. In contrast, international cloud services like AWS, Azure, and Google Cloud can buy tons of Nvidia processors (literally). To that end, how the shortage impacts Baidu and its domestic peers against global competitors remains to be seen.

A Geekbench 6 result features what is likely the first-ever look at the single-core performance of the Taishan V120, developed by Huawei's HiSilicon subsidiary (via @Olrak29_ on X). The single-core score indicates that Taishan V120 cores are roughly on par with AMD's Zen 3 cores from late 2020, which could mean Huawei's technology isn't that far behind cutting-edge Western chip designers.

The Taishan V120 core was first spotted in Huawei's Kirin 9000s smartphone chip, which uses four of the cores alongside two efficiency-focused Arm Cortex A510 cores. Since Kirin 9000s chips are produced using SMIC's second-generation 7nm node (which may make it illegal to sell internationally according to U.S. lawmakers), it would also seem likely that the Taishan V120 core tested in Geekbench 6 is also made on the second-generation 7nm node. 

The benchmark result doesn't really say much about what the actual CPU is, with the only hint being "Huawei Cloud OpenStack Nova." This implies it's a Kunpeng server CPU, which may either be the Kunpeng 916, 920, or 930. While we can only guess which one it is, it's almost certain to be the 930 given the high single-core performance shown in the result. By contrast, the few Geekbench 5 results for the Kunpeng 920 show it performing well behind AMD's first-generation Epyc Naples from 2017.

This appearance of what is likely the Kunpeng 930 is the first we've seen since it was announced in 2019. A report from The Next Platform in 2021 expected the 930 to launch that year on TSMC's 5nm, followed by the Kunpeng 950 on TSMC's 3nm node in 2023. Obviously, that didn't happen, and even if the 930 and 950 were going to be made on TSMC's cutting-edge nodes, those plans had to be shelved since Huawei is banned from fabbing at TSMC.

Additionally, the benchmark was only run on one core, and since the last-generation Kunpeng 920 capped out at 64 cores, this result was likely run on a virtual machine or some configuration where only one core was tested. That in turn means the multi-core score isn't really useful and can't tell us too much about how performant the full chips are.

Per Geekbench's processor rankings, the Taishan V120-powered server CPU is about on par with AMD's Zen 3-based Epyc Milan CPUs, specifically the midrange Epyc 7413. It's also about equivalent to Intel's Xeon E-2136, a Coffee Lake-based Xeon from 2018. Considering the Geekbench 6 result says the Huawei CPU operated at 2.9 GHz, it's not a bad result; the Epyc 7413 boosts to 3.6 GHz and the E-2136 to 4.5 GHz.

However, it is clearly behind modern Epyc Genoa based on Zen 4, as well as Intel's Sapphire Rapids, which are both around 25% faster. To be fair to the Taishan V120 chip, it is likely fabbed on a less performant node, as Genoa uses TSMC's 5nm and Intel uses its own Intel 7 process.

Unfortunately, without a look at multi-threaded performance, power consumption, and efficiency, it's hard to say how competitive Taishan V120 cores will be. For servers in particular, power and efficiency are key due to the cost of electricity, and even if Huawei's latest server CPUs are fast, that won't mean much if they consume tons of power.

Last August Huawei surprised the industry by launching its HiSilicon Kirin 9000S smartphone processor presumably made by Semiconductor Manufacturing International Corp. (SMIC) using its 2nd generation 7nm-class process technology. Since then, another Kirin 9006C chip showed up, though this one was allegedly made by TSMC years ago. This week a yet another HiSilicon system-on-chip — the Kirin 9000W — transpired.

The eight-core HiSilicon Kirin 9000W powers Huawei's MatePad Pro 13.2-inch tablet, which is evidently aimed at the high-end of the market, given its display size and a relatively high price of over $1000. Huawei's Kirin 9000W-based tablets are currently available in Malaysia, Indonesia, and Saudi Arabia, as reported by HuaweiCentral. Meanwhile, there are versions of the MatePad Pro 13.2-inch device that are equipped with the Kirin 9000S processor. 

The main difference between Huawei's Kirin 9000W-based tablets and Kirin 9000S-based tablets is the lack of 4G/LTE connectivity on the former and lack of cellular capabilities of the latter. So, it is possible that the Kirin 9000W application processor comes with disabled built-in modem.  

Huawei does not reveal the specifications of its Kirin 9000W system-on-chip, so we can only wonder how it differs from the Kirin 9000S SoC. Keeping in mind that the Kirin 9000W has the same number of general-purpose cores as the original Kirin 9000 (made by TSMC on its 5nm-class fabrication technology in 2020) and the Kirin 9000S (made by SMIC on its 2nd generation 7nm-class production node in 2023), we are probably dealing with a processor that derives from either of the said chips.

Given the fact that the Huawei MatePad Pro 13.2-inch tablet is clearly aimed at demanding users, we would not expect the HiSilicon Kirin 9000W to be a significantly cut-down version of the Kirin 9000S or the Kirin 9000 (apart from the supposedly disabled modem hardware). In fact, it is reasonable to expect the Kirin 9000W to actually have a higher thermal envelope than the Kirin 9000S or the Kirin 9000 as at the end of the day it is designed for tablets (which tend to have better heat dissipation than handsets), not smartphones. Meanwhile, if this is the case, it does not explain why some versions of the Huawei MatePad Pro 13.2-inch tablet are powered by the Kirin 9000S made by SMIC.

In a bid to lower its reliance on Nvidia's hardware for artificial intelligence applications, Chinese cloud giant Baidu acquired hundreds of servers based on Huawei's HiSilicon Ascend 910B processors, reports Reuters citing sources with knowledge of the matter. These application specific integrated circuits (ASICs) were developed to compete against Nvidia's A100 AI and HPC GPUs, but many would consider them outdated. 

In total, Baidu ordered 200 servers running 1600 Ascend 910B processors in August, demonstrating that Baidu is serious about using Huawei's technology on a large scale. Huawei's Ascend 910B is meant to compete with Nvidia's A100-based offerings that are used primarily for inference, and Baidu's order is a big vote of confidence for Huawei. By October, Huawei had already delivered about 1,000 of the processors, proving it can handle relatively big orders pretty quickly despite restrictions by the U.S. government.

Speaking of volume, it should be noted that Baidu uses thousands of Nvidia GPUs, so Baidu is not wholesale substituting Ampere or Hopper GPUs with Huawei's hardware. However, Baidu's decision to use Huawei's ASICs marks a notable shift in its procurement strategy, signaling that the company is willing to use Chinese processors instead of those developed by Nvidia for its AI application. Given the fact that Nvidia can no longer ship its AI and HPC GPUs to China-based entities without an export license, Baidu and other Chinese companies will have to choose AI processors from domestic vendors. 

Interestingly, Baidu has its own Kunlun AI processors. By choosing to deploy Huawei's HiSilicon Ascend 910B, Baidu may be pursuing a multifaceted AI strategy that relies on AI processors from multiple vendors and in-house designed silicon. It naturally indicates a strategic shift for Baidu, potentially setting a precedent for other tech giants to follow. 

As far as Huawei is concerned, if it manages to ensure a stable supply of its HiSilicon Ascend 910B ASIC, it could earn a significant revenue stream from Chinese companies that are starving for AI and HPC processors due to the US government's ever-tightening sanctions. 

Russian processor developer Baikal Electronics wants to show the world that its chips can compete with Intel, which makes some of the best CPUs on the market, and Huawei. In a series of benchmarks, the company conducted and shared with Russian news outlet Cnews, the fabless semiconductor put its Baikal-S server processor up against Intel's Xeon Gold 6230 and Huawei's Kunpeng 920, and the Russian chip's performance wasn't that good, but wasn't awful either. The chip was way behind Huawei's processor but beat Intel's outdated offering in some tests.

The Baikal-S features 48 Arm Cortex-A75 cores on a 16nm process node with a 2 GHz base clock and 2.5 GHz boost clock. The Kunpeng 920, specifically the 920-4826 model number, wields 48 TaiShan v110 cores with a 2.6 GHz clock speed. Baikal's processor is on an older process node than the Kunpeng 920's newer 7nm TSMC HPC manufacturing process.

Having launched in 2019, Intel's Xeon Gold 6230 CPU is a bit outdated and not necessarily a fair competitor to choose for the Baikal-S. It sports just 20 cores (40 threads) and base and boost clock speeds up to 2.1 GHz and 3.9 GHz, respectively. It's an upgrade over Baikal Electronics' previous comparison, which used an older 20-core Xeon Gold 6148 (Skylake) for the confrontation.

If Baikal wanted a fairer comparison with Intel's server chips, the company would have benchmarked one of Intel's Xeon Platinum products, many of which have 48 or more cores.

For some strange reason, the Russian vendor omitted an AMD chip from its latest comparison. It's a surprising move since Baikal Electronics had enthusiastically claimed that the Baikal-S was comparable to the 16-core EPYC 7351 from the Zen 1 period.

Unfortunately, Baikal didn't disclose the test systems' specifications or the testing environment's conditions in sharing its benchmarks. So take these results with a giant grain of salt.

Although there were only three processors to test, Baikal Electronics didn't run all the benchmarks on each. Whether or not the company was trying to cherry-pick results to help the Baikal-S stand out is uncertain. For us, the challenging part in evaluating these was combing through all the results and finding the most relevant metrics for comparison since some data from specific benchmarks were missing.

Admittedly, CoreMark is far from being a comprehensive benchmark for evaluating processors. Nonetheless, the Kunpeng 920 was up to 13% faster than the Baikal-S in the CoreMark single-threaded test. The Kunpeng 920 also beat the Baikal-S by 23% on the multi-threaded test. Meanwhile, the Baikal-S outperformed the Xeon Gold 6230 by 43% in the same benchmark.

The Stream benchmark helps measure sustainable memory bandwidth. While we know the number of memory channels supported per processor, we don't know the speed or capacity of the DIMMs that Baikal Electronics used for its tests. According to the results, the Baikal-S delivered 34% higher bandwidth than the Xeon Gold 6230 in the Stream benchmark. However, the Russian chip was overshadowed by the Kunpeng 920, which put up a 33% higher score.

The Linpack benchmark is something that many will be familiar with since it's the default test for ranking the TOP500 list of supercomputers. In Baikal Electronics' case, the vendor used version 2.3 of Linpack.

The Xeon Gold 6230 was the best-performing chip in Linpack, annihilating the Baikal-S and Kunpeng 920 by 140% and 160%, respectively. The Baikal-S scored a small victory over the Kunpeng 920, beating the Chinese chip by 8%.

Baikal Electronics also shared some SPEC CPU 2017 benchmark results for the Baikal-S running at 2 GHz and 2.5 GHz. The company didn't compare Baikal-S to the Xeon Gold 6230 or the Kunpeng 920.

Regarding 7-Zip compression workloads, the Kunpeng 920 was 73% faster than the Baikal-S. The company didn't benchmark the Xeon Gold 6230's performance in this metric. On the other hand, the Baikal-S defeated the Xeon Gold 6230 by a 67% margin in the 7-Zip decompression workloads. However, the Baikal-S was no match for the Kunpeng 920, which posted a whopping 78% difference over the Russian processor.

Geekbench 5 is a benchmark that a lot of us can relate to. Although it's not the best test for comparing processors, it's a more mainstream benchmark. Baikal Electronics didn't benchmark the Kunpeng 920. As expected, the Xeon Gold 6230 delivered 112% higher single-core performance than the Baikal-S. However, the Russian chip achieved an 80% higher multi-core score than the Xeon Gold 6230, which is unsurprising considering that Intel's processor has less than half as many cores.

Baikal Electronics products are far from competing with Intel, AMD, or even Huawei, and the company's results back it up. However, given the multi-socket support, the company is optimistic that the Baikal-S can get on par. A two-socket configuration is reportedly ready, while the company's working on a quad-socket design.

According to representatives, the company has already commenced its work on the Baikal-S2, a next-generation 6nm chip with 28 Arm Neoverse-N2 cores ticking at 3 GHz and supporting up to eight channels of DDR5 memory. Baikal Electronics expects to release the Baikal-S2 between the second to third quarter of 2025, allegedly offering performance uplifts up to 6X that of the Baikal-S.

The US has blacklisted both Huawei and Phytium, a China-based CPU maker, but business still continues apace. Huawei last week said that it had teamed up with the state-owned China Electronics Corporation (CEC) to build a unified hardware-software ecosystem over its own Kunpeng processors and the US-sanctioned Tianjin Phytium Technology's Phytium CPUs, reports HuaweiCentral. The joint ecosystem will be called Penteng and will enable CEC to use both Kunpeng and Phytium chips in its devices without making major changes. In addition, the companies plan to invite new partners to the project.

Contemporary Kunpeng and Phytium processors use an Armv8 instruction set architecture and are compatible with Chinese cryptography algorithms, so it should not be too tricky to ensure that both CPUs run the same programs just fine. On the hardware side of matters, things will get more challenging since modern chips use different process technologies and different packaging. Therefore, it remains to be seen how these differences will be addressed by engineers working on the Penteng project. 

Both Huawei's chip design arm HiSilicon and Tianjin Phytium Technology are blacklisted by the U.S. government and cannot access leading-edge production capacities. As a result, both companies have to find a way to make competitive chips domestically. Therefore, it is probable that the companies will unify the form factors and packages of their next-generation CPUs. While the partnership includes a focus on Phytium, purportedly a subsidiary of CEC, Phytium has not yet formally announced its participation in the Penteng project.

This collaboration not only incorporates the refinement of software as well as the unification of some hardware parts but also plans to enlarge the ecosystem by inviting more partners through a certification program. This joint work provides the opportunity to engineer a wide array of products that can accommodate all modern workloads, including cloud data centers, edge computing, and client PCs. Moreover, Huawei and CEC have committed to address emerging business opportunities with Penteng.

Building up an ecosystem encompassing hardware and software developed in China seems like quite a significant move as it will make the Chinese IT industry generally more competitive. Programs for x86 platforms running Windows or Linux work on appropriate PCs from all vendors. By contrast, this is not always the case with Chinese programs and PCs from different suppliers.

Huawei Technologies, which is banned from using U.S.-originated technologies, is planning its return to flagship smartphone market by the end of the year, reports Reuters citing three research firms. The company will reportedly use Chinese fabs and technologies to produce advanced chips required for 5G handsets, although the details remain vague.

Reports from several research companies covering the Chinese smartphone market indicate that Huawei will likely be able to secure domestic 5G chips, thanks to its own developments in semiconductor design tools and chip manufacturing by the Semiconductor Manufacturing International Co. In particular, Huawei is rumored to make use of SMIC's N+1 manufacturing process, which belongs to 7nm-class production nodes, though it is likely that the high-tech giant will use a refined version of N+1.

Huawei's current flagship P60 Pro smartphone relies on Qualcomm's platform as well as the company's own Harmony OS operating system. Using SMIC's production node, the company could probably produce a cheaper handset for the mass market, increase sales of its smartphones, and gain experience with domestic technologies for 5G smartphones. The company has reportedly upped its smartphones sales projections for 2023 from 30 million to 40 million.

What remains to be seen is whether Huawei can produce a rival for the Qualcomm SM8475 Snapdragon 8+ Gen 1 platform using one of SMIC's nodes.

Earlier this year Huawei announced that it had developed electronic design automation (EDA) software to design chips based on 14nm-class and thinner process technologies, which represented a significant advancement for the Chinese EDA industry. Huawei expected to complete testing on this software in 2023 and indicated plans to use its EDA software to design its proprietary HiSilicon chips.

While Huawei does not have its own fabs and cannot access leading-edge process technologies at TSMC, IFS, and Samsung Foundry, it has been collaborating with SMIC, Chinese foundry champion, over building a fab and developing fabs tools for a while. The two companies could refine SMIC's N+1 and other advanced nodes to make them suitable for Huawei's HiSilicon SoCs, but this is our speculation at this point.

The U.S. government restricted Huawei's access to advanced process technologies in 2020, but despite this, the company continues to advance its technologies, particularly in such fields as artificial intelligence. Apparently, the company is using outdated hardware obtained years ago for its current AI projects, reports DigiTimes. At its Huawei Developer Conference this week, Huawei also unveiled its new Pangu Models 3.0 AI Large Model for a wide range of industrial and scientific applications. 

Before Huawei was hit by severe sanctions in 2020, the company's HiSilicon division was among the largest clients of TSMC and obtained plenty of chips made on the foundry's N7 (7nm-class) process technology. The chips in question are Kunpeng 920 processors as well as Ascend 910 AI accelerators that are used for Huawei's Atlas 900 Pod A2 AI training cluster that Huawei's using to advance its AI efforts.

Despite encountering challenges in obtaining access to advanced processing technologies, Huawei is still establishing a strong presence in the AI market through the strategy of import substitution, DigiTimes reported without elaborating. Earlier this year, the company demonstrated new server motherboards and used tape and coolers to hide the names of its component suppliers.

At the Mobile World Congress 2023 event held in Shanghai, Huawei displayed AI devices that perform inferencing in both cloud and edge computing environments, the report claims. At another trade show, the World Artificial Intelligence Conference (WAIC), the company demonstrated an AI robotic arm powered by its Atlas 200IDK A2 and the Ascend developer kit that could recognize colors and move. That was followed by the company's release of the Pangu Models 3.0 AI models earlier this week.

Keeping in mind that the AI industry in China is estimated to reach a value of $150 billion eventually, it is just too attractive of a market for Huawei to miss. As a result, Huawei is expected to keep investing in its AI software and ecosystem (including its own compute framework) while it is trying to either get potent hardware either by using advanced packaging and mature process technologies or gaining access to sophisticated production nodes.

It is unclear whether Huawei can catch up with rivals that are not hit by restrictions, but at least for now, it looks like the company can continue the development of its AI prowess.

Semiconductor Manufacturing International Corp., China's largest contract maker of chips, has quietly removed 14nm fabrication technology from the list of its services on its website. The company also did not talk about advanced FinFET-based technologies at its earnings call last week, reports DigiTimes.

The U.S. government's latest export control rules prohibit Chinese semiconductor manufacturers from acquiring instruments and technologies required to produce non-planar transistor logic chips with dimensions of 14nm/16nm or less, 3D NAND chips with 128 or more active layers, and DRAM ICs with a half-pitch of 18nm or below. With similar restrictions from the Netherlands, Japan, and Taiwan set to come into effect in mid-2023, Chinese companies like SMIC and YMTC will be cut off from obtaining the equipment needed for fabricating chips on their latest production nodes, including 14nm/12nm and N+1 for SMIC, as well as 128-layer and 232-layer 3D NAND for YMTC.

Without access to advanced equipment and spare parts from suppliers like ASML, Applied Materials, KLA, and Lam Research, SMIC may be unable to build chips for its customers using its latest fabrication technologies, so it is reasonable to remove its 14nm platform from the list of its technologies.

Meanwhile, the company's shuttle services page still mentions 14nm and even states that the company regularly produces shuttle wafers using this node.

"Currently, we provide shuttle services for processes from 0.18μm to 14nm on a regular basis," a statement by SMIC reads 

Furthermore, another page at SMIC's website also mentions its 14nm capability.

"SMIC is a pure-play semiconductor foundry that provides wafer fabrication of 200mm and 300mm wafers at 0.35-micron (μm) to 14-nanometer (nm)," the description of the company's foundry solutions reads. 

SMIC has been using 14nm-class production techology since the end of 2019 at its SN1 facility. One of the 14nm SoCs the company produced is Huawei's HiSilicon Kirin 710A. However, despite technically being in mass production, it looks like 14nm volumes were so limited that the company stopped disclosing the revenue attributed to this node. Instead, it combined it with the earnings from the 28nm node, which has not been a significant revenue contributor either.

Last year, Chinese state media reported that SMIC had initiated high-volume manufacturing on its 14nm-class fabrication method at its Fab SN1 fab near Shanghai, China. That report also suggested that the company is advancing with its 7nm and 5nm-class nodes, even though it is unable to acquire cutting-edge wafer fab equipment.

The Made in China 2025 national strategic plan to develop the domestic semiconductor industry inspired the establishment of tens of thousands of fabless chip designers in the People's Republic in just a few years. However, it seems many of them were unfit to survive intense competition between each other and global players, which is why about 10,000 such chip companies went out of business during 2021–2022, reports DigiTimes.

Market observers now wonder whether the massive shuttering of Chinese IC design companies is a result of tighter U.S. export control imposed in 2020–2022, or the ongoing global semiconductor industry downturn. While both factors contributed to the closures, there were a number of issues specific to China that drove around 10,000 homegrown chip companies out of business.

The Made in China 2025 program implemented several policies to meet its goals, including lower taxes for high-tech companies, encouraging acquisition of foreign tech company, supporting R&D funding by large manufacturers, and direct state R&D funding, among other things. The results were inspiring to say the least. The number of Chinese chip developers increased from 736 in 2015 to 1,780 in 2017, according to China Renaissance Securities. Then in 2020–2021 as many as 70,000 chip companies were registered, according to DigiTimes.

The report admits that the escalation of the U.S.-China trade war in 2018 further inspired the Chinese government to fund high-tech companies, with many of the IC design houses established due to subsidies from federal or local governments. In addition to subsidies, a rush of speculative capital drove the establishment of similar IC design firms developing commodity chips that were poised to crash for multiple reasons.

The semiconductor industry is well known for being capital intensive, but the sector is heavily dependent on talent, skillful management, and knowledge. Even if a company has ample resources, it still needs to attract talent, invest in research and development, and ensure sufficient production capacity from foundries for sustainable long-term development. Without engineering talent and proper management, the chances of success are not high. While there are loads of engineers in China these days, there are not enough managers to run these IC designers successfully.

Meanwhile, an investment bubble emerged. Ding Xing Quantum, a private equity firm based in China, has been investing in domestic IC designers since 2017. The company observed that at the onset, the worth of such a company ranged between RMB 200–300 million ($28–$43 million). However, by 2019, the valuation of startup companies in this sector ballooned to over RMB 1–2 billion ($145–$190 million), which clearly points to an investment bubble, and such bubbles tend to crash.

There was another factor that contributed to the shortfall of Chinese IC design companies. Underperformance of the Chinese consumer market transitioned into a structural imbalance of supply and demand starting Q3 2021, and then the global semiconductor industry faced inventory correction in the second half of 2022 and entered its own downturn. As a result, the demand for chips dropped in general, and China-based developers that produced commodity ICs went bankrupt since they could not offer anything special.

The sanctions against the Chinese semiconductor industry clearly affected development of the sector, as it is evident from the problems faced by companies like Alibaba, Biren, HiSilicon, and YMTC. Meanwhile, the global semiconductor downturn and the inability of many Chinese chip designers to compete played a far bigger role in the dissolution of 10,000 semiconductor entities in the People's Republic, the story by DigiTimes concludes.

Huawei said at an event that it had developed electronic design automation (EDA) software to design chips to be produced at 14nm-class process technologies, Reuters reports. While 14nm nodes are several generations behind 3nm production nodes currently used to make the most advanced chips, this is a major breakthrough for the Chinese EDA industry. 

Huawei is expected to finish testing on EDA software for 14nm and more advanced chips this year, according to a speech by rotating chairman Xu Zhijun on February 28. The software suite was developed in collaboration with domestic EDA companies, "basically realizing the localization of EDA tools above 14nm." The rotating chairman reportedly added that Huawei has already developed 78 tools related to semiconductor hardware and software.

Although Huawei intends to use its EDA software to design its own HiSilicon chips, Caijing, a Chinese financial news magazine, reported that Xu also mentioned that Huawei wants to share these tools with partners and customers.

While China has set its goal for a self-sufficient semiconductor design and production industry, it still lacks many critical pieces. Among the things that the People's Republic lacks is its own electronic design automation software. There are thousands of chip designers in China, and the lion's share of them use EDA tools developed by U.S.-based Ansys, Cadence, and Synopsys, as well as Germany-based Siemens EDA.  

Under the latest export control rules imposed by the U.S. government in October 2022, hardware and software technologies of American origin that enable the development or production of logic chips with non-planar transistors on 14nm/16nm nodes cannot be shipped to China without an appropriate export license from the U.S. Department of Commerce. 

Advanced EDA tools, just like sophisticated wafer fab equipment, are among the items that cannot be shipped to China with approval from the U.S. government. 

HiSilicon and potentially other China-based chip designers can develop chips set to be made on a 14nm-class node using Huawei's software already in 2023 or 2024. Meanwhile, it remains to be seen whether they can produce their chips at China's Semiconductor Manufacturing International Corp. (SMIC), Taiwan's TSMC or South Korea's Samsung Foundry. These companies will have to obtain a license from the U.S. DoC to serve such orders from Chinese customers.

Huawei is also working on its own wafer fab equipment tools, which will likely be among the cornerstones of the Chinese semiconductor industry. However, those tools will not make it to fabs for at least the next several years. 

For Huawei, it is close to impossible to procure chips for its server and communications products that are widely used in China and other countries. But that doesn't mean that the company cannot get chips at all; it just does not want anyone to know who sells them. At Mobile World Congress in Barcelona, the company demonstrated its new server motherboards with all chips covered with disguising tape and coolers to hide the names of its suppliers. 

Huawei still sells boatloads of servers and communication equipment in China and a number of other countries, but to build those devices, it needs chips. Virtually all chips today — whether they are logic or memory ICs — are designed using electronic design automation (EDA) tools developed in the U.S. and produced on equipment containing technologies created in America, so their suppliers need to get appropriate licenses from the U.S. Department of Commerce to sell them to Huawei or its subsidiaries.  

But getting those licenses is tricky, which is why Huawei likely has to buy chips on the gray market or use complicated means to get hardware from its developers. In both cases, the company prefers not to demonstrate what it uses openly and hides its suppliers, as we can see from images published on Twitter by Jay Goldberg, a 5G, IoT, and networking analyst focused on China. Of course, another reason to hide chips that it uses from prying is could be a way to conceal trade secrets from competitors (and we know that there are Chinese companies copying Huawei's products).

You need to expand the above tweet to see the images. One of the boards (marked at GFMPUB Ver. A) not only conceals logic chips using a radiator or tape, but even hides the supplier of the memory ICs. Another board, which looks like a prototype (or even a mockup) of a 4-way server motherboard, not only conceals the markings on some of the chips, but even does not carry processors, perhaps to guarantee that no one could guess their producer even if the board gets stolen.  

After Huawei became a victim of the ongoing trade war between the U.S. and China on the grounds that it has ties with the People's Republic army, the company can no longer procure hardware and software from U.S.-based companies or containing technologies developed in America without getting an appropriate license from the U.S. Department of Commerce. Huawei's HiSilicon subsidiary lost access to advanced semiconductor production outside of China, which to a large degree blocked its advanced chip development. 

As it turns out, there are still ways for Huawei to obtain the chips it needs. Furthermore, the company is reportedly working with China-based SMIC — another company that lost its ability to develop at a rapid pace because of the U.S.-imposed sanctions — to build a fab capable of producing chips and system-in-packages it needs to build its products.

Twitter user Fritzchens Fritz has managed to obtain a sample of Baikal Electronics' 48-core BE-S1000 server-grade system-on-chip (SoC) and throw it under an infrared microscope to reveal its internals. In addition, some benchmark results of the SoC have surfaced.

Baikal Electronics has developed several system-on-chips for different devices to replace x86 processors from PCs and various compute appliances made in Russia. However, the pinnacle of the company's design prowess should have been its BE-S1000 server-grade SoC with 48 Arm Cortex-A75 cores, which the company managed to tape out and produce the first sample using TSMC's 16FFC fabrication technology, but which will never be released commercially due to sanctions against Russia for its invasion in Ukraine.

The Baikal BE-S1000 SoC comes in an FCLGA-3467 package to work in uniprocessor, 2-way, and 4-way symmetric multiprocessor (SMP) configurations. The processor's power consumption is about 120W, so it does not require any fancy cooling systems. The die size of the SoC is about 607mm², which is similar to the die size of Nvidia's AD102 graphics processor.

In addition to detailed die shots of the BE-S1000 processor displayed by Fritzchens Fritz, Locuza also published a chip floorplan with annotation.

The Baikal BE-S1000 organizes its 48 Art Cortex-A75 cores (operating at 2 GHz) across 12 clusters, each containing four Arm Cortex-A75 cores with 512KB L2 cache per core and 2MB of unified L3 cache. In addition, the 32MB of L4 cache organized in four blocks accompany the quad-core clusters.

The BE-S1000 processor also has six 72-bit memory interfaces supporting up to 768 GB of DDR4-3200 with ECC memory in total (i.e., 128GB per channel), five PCIe 4.0 x16 (4x4) interfaces (three of which support CCIX 1.0 to organize 2-way and 4-way SMP configurations), one USB 2.0 controller, two 1GbE interfaces, and various general purpose I/O. In addition, the SoC features an interconnect by a coherent mesh network.

Baikal positioned its BE-S1000 against AMD's 16-core EPYC 7351 (2.90 GHz), Intel's 20-core Xeon Gold 6148 (2.40 GHz), and Huawei's 48-core Kunpeng 920 (2.60 GHz) in its marketing materials published by Locuza. As for performance, Baikal's slides indicate that the BE-S1000 that was supposed to hit the market in 2022 ~ 2023 would outperform Intel's Xeon Gold 6148 (from 2017) by a substantial margin in various kinds of benchmarks except for HPLinpack (a supercomputer benchmark), which is not particularly surprising given that Cortex-A75 was not designed to run high-performance computing workloads with an FP64 precision.

In general, while the Baikal BE-S1000 looks like a brave attempt to develop a server-grade SoC that could replace processors from AMD and Intel for some machines, the chip would have arrived too late and been slower than then-contemporary CPUs from the x86 camp. Potentially, this could have been mitigated with the right price (at least for some cases). But because of Russia's bloody war in Ukraine, the BE-S1000 will remain an interesting artifact in the labs and will never become an actual product.

According to Bloomberg, Taiwan authorities have raided the offices of China-based Advanced Manufacturing EDA Co. (Amedac) on allegations that the company had been hiring semiconductor engineers from TSMC without following appropriate regulations. Amedac is co-owned by U.S.-based Synopsys, which has been probed for illegally providing electronic design automation (EDA) tools to Huawei's HiSilicon as well as SMIC. 

Officials from Taiwanese Ministry of Justice (MoJ) raided Advanced Manufacturing EDA Co. in March looking for evidence that the company had allegedly hired personnel from TSMC with the aim to develop its own EDA software. The investigators are looking into Amedac's hiring practices and at this time there are no prosecutions. Taiwan is in process of essentially banning recruiting employees from critical industries, such as production of semiconductors, by Chinese high-tech firms in a bid not to help China to propel its chip production industry. 

Taiwan and Taiwanese semiconductor firms are concerned not only with the fact that China is gaining semiconductor prowess, but with the fact that people hired from companies like TSMC could share trade secrets / proprietary technologies with their China-based employers. Back in the day TSMC sued SMIC for stealing its fabrication processes and while today it is close to impossible to copy complex FinFET-based production nodes, nobody wants to take a risk. 

EDA tools are also a major concern. While China has thousands of chip design firms and local chip producers, it does not have the EDA tools required for chip development. As a result, when the U.S. government essentially left HiSilicon and SMIC without new EDA software in 2020, the companies were badly hit. As a result, it is evident that Chinese firms need locally developed electronic design automation software. 

Established in September, 2019, Amedac develops EDA tools and is based in Hefei, China. Synopsys, one of the world's largest producers of EDA software, owns a 20% stake in the company, whereas the remaining stakes are controlled by SummitView Capital, CEC Huada Electronic Design Co. and the Institute of Microelectronics of the Chinese Academy of Sciences (a government-backed organization). 

Synopsys says that it maintains a distributor relationship with Amedac, which allows it to resell its tools to various customers. Meanwhile, Amedac has not got any actual technologies from Synopsys, the EDA specialist claims.

The U.S. government officials are working with their counterparts from the Netherlands to ban sales of virtually all lithography tools produced in the country to companies in China. If the American politicians succeed, this will significantly blow Chinese efforts to become a world-class semiconductor producer and will thwart the country's Made in China 2025 plan. 

ASML is the world's largest maker of lithography tools used to make all kinds of chips. However, the company cannot sell its extreme ultraviolet (EUV) lithography equipment to its Chinese clients. Still, U.S. politicians propose restrictions under which ASML will not be able to sell its mainstream deep ultraviolet (DUV) lithography tools to Chinese customers as well, reports Bloomberg. DUV is used to make the vast majority of widely used chips that power client PCs, servers, mobile electronics, autonomous vehicles, and robots.  

Persuading the Dutch government to ban sales of virtually all lithography scanners to Chinese customers will not be easy. Chinese fabs run by local companies like Hua Hong,  Semiconductor Manufacturing International Co. (SMIC), and YMTC, or global players like TSMC, Samsung, and SK Hynix accounted for about 16% of ASML's revenue in 2021 (which reached $18.6 billion), and that is a lot of money. 

ASML argues that it is not the world's only maker of DUV scanners (albeit the largest one), and similar machinery is available from Canon and Nikon. Yet, if the U.S. manages to leave China without ASML's tools, other companies will not be able to replace them soon. Furthermore, American politicians argue that advancements in modern chip packaging technologies allow Chinese companies to design and produce pretty sophisticated chips that can advance Chinese supercomputing and, ultimately, military technologies. 

The U.S. has banned several Chinese companies from accessing technologies developed by USA-based companies, nearly driving Huawei's chip arm HiSilicon out of business. In addition, prohibiting sales of ASML's lithography tools to China will devastate the local semiconductor industry.  

Meanwhile, banning ASML's scanners to Chinese companies is not the only way the U.S. can thwart the local chip industry. Semiconductor fabs use hundreds of tools produced by numerous American companies, such as Applied Materials, KLA, and Lam Research. Forbidding them to work with China will have a devastating effect on Tianxia's semiconductor efforts. Moreover, some of their tools cannot be sold to China due to national security concerns. 

But leaving companies from China without semiconductor fabrication tools will impact the rest of the world. For example, a substantial part of DRAM and 3D NAND memory production is in China. Leaving companies like Samsung, SK Hynix, and YMTC without their Chinese fabs will affect the global chip and electronics supply chain. Furthermore, TSMC, SMIC, and Hua Hong produce loads of chips for their international clients; if they wind down production, this will also affect American chip designers.

Huawei has developed (and patented) a chip stacking process that promises to be significantly less expensive than existing chip stacking methods. The tech will help Huawei continue to develop faster chips using older mature process tech, which could theoretically help it avoid US sanctions. 

The only question is whether Huawei can actually take advantage of its innovation, given that foundries cannot produce chips for the company without an export license from the U.S. government. But at least Huawei itself certainly believes that it can, especially given that this tech could provide a performance boost for chips based on older nodes that aren't subject to such harsh US restrictions. 

A Way to Stay Competitive

We'll get to the details of the new tech below, but it's important to understand why Huawei is developing this new tech. Since the U.S. government blacklisted Huawei and its chip design subsidiary HiSilicon and now requires all companies making silicon for them to apply for an export license since all semiconductor production involves technologies developed in the U.S., Huawei can't access any modern node (e.g., TSMC's N5), and therefore has to rely on mature process technologies.

To that end, innovative chip packaging and chiplet interconnection technologies in general as well as 3D stacking in particular is a way for the company to throw in more transistors into its SoCs and get the performance it needs to be competitive, said Guo Ping, Huawei's former rotating chairman, at a recent news conference, reports DigiTimes. Therefore, it makes great sense for the company to invest in proprietary packaging and interconnection methods, such as the one it patented. 

"Micro-nano technology, exemplified by 3D hybrid bonding technology, will be the primary means to extend Moore's Law," said Guo.  

The high-ranking executive from Huawei indicated that since modern leading-edge process technologies are progressing relatively slowly, multi-chiplet designs in 2.5D or 3D packages are a general way for chip designers to keep throwing more transistors into their products and meeting the expectations of their clients in terms of new features and performance.  Therefore, Huawei will continue to invest in area enhancing and stacking technologies for in-house designs, the former chairman stressed. 

The claim made publicly at a news conference clearly indicates that Huawei aims to use its hybrid TSV-free 3D stacking method (or maybe a similar and more mainstream method) for its upcoming products. The main question is whether the method requires any tools or technologies that the U.S. government may deem cutting-edge and not grant an export license (after all, most of the fab tools use technologies that originated in the USA). That said, whether we'll see a foundry making 3D chiplet packages for Huawei using the latter's patented method remains to be seen. But at least Huawei has a unique technology for inexpensive 3D stacking that can help it to stay competitive even without access to the latest nodes.

Stacking Without Vias

Innovative chip packaging and multi-chiplet interconnection technologies are set to become crucial in the next few years for leading-edge processors, and so all major chip developers and manufacturers now have their own proprietary chip packaging and interconnection methods.

Chipmakers generally use two packaging and interconnection methods: 2.5D packaging that enables a high-density/high-bandwidth in-package interconnect for chiplets sitting next to each other, and 3D packaging that makes processors smaller by stacking different chiplets on top of each other. However, 3D packaging typically requires rather complex wiring since chiplets need to communicate and power has to be delivered using TSVs. 

While TSVs have been used in chip manufacturing for over a decade, they add complications and costs to the packaging process, so Huawei decided to invent an alternative solution without TSVs. What Huawei's specialists designed is essentially a hybrid between 2.5D and 3D stacking, as two chiplets overlap each other inside the package, saving space, but do not exactly sit on top of each other like in classic 3D packages. 

3D Stacking with Overlapping

Huawei's method uses overlapping parts of the chiplets to establish a logical interconnection. Meanwhile, two or more chiplets still have their own power delivery pins connected to their own redistribution layer (RDL) using a variety of methods. But while Huawei's patented technology avoids using TSVs, it does not look easy and cheap to implement.

Huawei's process involves flipping one of the chiplets upside down before connecting to another (or others). It also requires building at least two redistribution layers to deliver power (e.g., two chiplets means two RDLs, three chiplets can still use two RDLs, so do four, see the above gallery for details), which is not particularly cheap as it adds several additional process steps. The good news is that the redistribution layer of one of the chips can be used to connect things like memory, thus saving space.

In fact, Huawei's hybrid 3D stacking method may arguably be more universal than other companies' traditional 2.5D and 3D packaging technologies. For example, it is rather hard to stack two or three power-hungry and hot logic dies on top of each other, as cooling such a stack will be very complicated (which may ultimately mean compromises with clocks and performance). Huawei's method increases the surface size of the stack, which simplifies cooling. Meanwhile, the stack is still smaller than a 2.5D package, which is important for mobile applications like smartphones, notebooks, or tablets.

Not Alone

It is noteworthy that Semiconductor Manufacturing International Corp. (SMIC) with whom Huawei reportedly plans to build a fab, is also betting on advanced packaging and interconnection technologies as on a way to get around sanctions imposed by the U.S. government. The company cannot get access to manufacturing tools required to make chips using sub-10nm fabrication technologies, so advanced packaging and interconnection methods are crucial for SMIC as well.  

Other contract makers of semiconductors (TSMC, GlobalFoundries), integrated design manufacturers (Intel, Samsung), and even fabless chip developers (AMD) that can get access to leading-edge fab tools and process technologies also develop their own 2.5D and 3D chiplet stacking and interconnection methods to offer their clients or for their future products. Thus, Huawei is just following the curve. 

Update 4/16/2022 7:45am PT: Corrected article to clarify Intel's statements and specify that the comments did not occur at a political event. 

Amended article: 

Some interesting comments made by an Intel senior executive in China have come to light. According to a report from Taiwan's DigiTimes, an Intel exec said that Chinese CPU makers could become "strong rivals" in the next three to five years. Rui Wang, SVP of Intel Corporation and chair of Intel China, made the comments to Guancha.cn, a Chinese media outlet.

"So far there has not been any local companies that are able to deal a substantial threat to Intel," stated the Intel SVP. "But in 3-5 years, it will become clear that local companies will emerge as strong rivals," envisioned Ms. Wang (as translated by DigiTimes). Having said this, Wang went a little distance to temper expectations that US tech giant Intel would easily be displaced, dislodged, or usurped as the CPU leader in China. "Intel won't be polite, and will exert its power to compete fairly."

Chinese Companies With Intel in Their Sights

Unfortunately, Intel China's Chair wasn't specific about the avenue from which Intel's business in China would be assaulted. No specific Chinese CPU or IT company was highlighted, and no mention was made of any specific architecture with a chance at becoming an Intel competitor.

China has quite a wide range of CPU design outfits that could potentially surprise us in the next three to five years. We have previously and recently reported on the successes of China-made Zhaoxin CPUs. One of these LuJiaZui microarchitecture quad-core chips features in what looks like a China-only version of the Yoga Slim 7i laptop, dubbed the Lenovo Kaitian N7. The immature architecture combined with a custom China OS means that the N7 isn't a great performer, however, and it looks like it has issues with waste heat and battery life.

We have also posted news in recent months discussing Chinese chip advances by the likes of Loongson (MIPS based), Sunway (supercomputers), and Hygon (AMD Zen 1 based, under license). China also has at least a couple of firms building processors on the Arm architecture. Huawei's HiSilicon and Phytium Technology are the two best know Arm architecture proponents in China. Phytium has a particularly wide portfolio with products for the server, workstation, and embedded markets, as well as five new high performance CPUs penciled in for the coming months.

China has homegrown GPUs and DDR4 (currently planning DDR5 production) too, and hopes to make its semiconductor foundries, especially the state backed SMIC, more competitive in the not-too-distant future.

Some other important data with which to frame the above news includes a recent statement by Xiao Yaqing, China's Minister of Information and Technology. Just ahead of the National Party Congress, Yaqing asserted that the domestic chip industry had grown by a third compared to a year ago.

Huawei is reportedly teaming up with SMIC to build a semiconductor manufacturing facility in mainland China that will serve its needs either exclusively, or almost exclusively, helping it to avoid the impact of US sanctions that have severely limited its access to crucial chip technologies. Huawei intends to spend as much as $10 billion on the fab, according to reports by Huawei Central and UDN.

The Fab

The new fab is expected to be built near Shenzhen, and it will be used to manufacture Huawei's HiSilicon chips, presumably using process technologies developed by SMIC. We can make an educated guess based on Huawei's $10 billion investment that this will be a fairly advanced 300mm fab, so it will likely be at least 28nm-capable (though we would speculate that Huawei is probably looking at 14nm FinFET capabilities). Unfortunately, there is no reliable information about the process technologies planned for the fab. 

At present, SMIC's most advanced fabrication process is N+1, which is believed to be a low-cost 7nm-like node. Meanwhile, the most advanced manufacturing technology that SMIC uses for high volume production (HVM) is its 14nm node. HiSilicon can design leading-edge system-on-chips (SoCs) for TSMC's N5-series nodes, but it also has loads of SoCs for legacy processes, so the fab will undoubtedly be quite busy.

HiSilicon's Needs

Huawei is one of the largest chip consumers in China that also happens to have its own chip design arm (HiSilicon). However, being blacklisted by the U.S. government, Huawei has severe difficulties securing chip production capacities for HiSilicon chips that it uses for tens of millions of products. Currently, foundries must obtain a license from the U.S. Department of Commerce for any contract with Huawei involving an American technology (which means 99.9% of them). 

HiSilicon was founded in 1991, and in 30 years it has developed hundreds if not thousands of SoCs, processors, modems, controllers, and accelerators for the diverse needs of its owner. Given HiSilicon's huge product portfolio, it has very diverse needs for process technologies, so the new fab will have to offer numerous nodes and probably some advanced packaging capabilities (we are speculating) to meet the needs of its key customer. 

For Huawei, one of the ways to ensure a steady supply of HiSilicon chips is to build its own semiconductor production facility that will serve its needs either exclusively, or almost exclusively. As it turns out, it has found a way to get the chips it needs — build a dedicated fab. Yet frankly, there are more questions than answers at this point.

Different Options

At present, the structure of the deal is unclear. Huawei could own the fab, and then SMIC will license its process technologies and lend some engineers and workers. Conversely, it could be co-owned by Huawei and SMIC but operated by SMIC, and therefore inherit the latter's nodes and personnel.

Considering SMIC's typical business model (build a fab in a joint venture with local governments and then operate it), it is likely that China's largest foundry will manage the fab. Still, Huawei will either own a controlling stake or co-own the fab. Here's what that might look like. 

Option 1: Huawei Goes IDM

Suppose the fab is 100% owned by Huawei through a subsidiary (which means that Huawei is set to adopt an integrated device manufacturer [IDM] business model). In that case. In that case, it might be very hard for the company to obtain production tools and then ensure a steady supply of components and consumables. The U.S. government has sent a clear message about limiting Huawei's access to as many advanced technologies as possible.  

Being one of the world's largest high-tech companies, Huawei has very deep pockets, so spending $10 billion on a fab is hardly a problem for the giant. But building it and then ensuring that it operates smoothly is a completely different task.  

While Huawei is a mighty company, developing process technologies will take years (assuming that Huawei has appropriate scientists and engineers). But since SMIC is in play, Huawei could license manufacturing processes from its partner, but this might require an appropriate export license from the U.S. government since these nodes were developed using American technologies. In addition, Huawei will need experienced personnel to work at the fab, so it remains to be seen where it plans to get the staff. 

Furthermore, even if the IDM model is the case here, the fab will likely need a license to supply chips to its parent company. Yet, if the fab is dedicated solely to Huawei, the company's chip supply will be somewhat more predictable. Meanwhile, it remains to be seen whether Huawei can utilize the whole fab because an underutilized fab generates substantial losses. 

Option 2: Co-Owning a Fab

If Huawei, SMIC, and other possible parties co-own the fab, it will perhaps be somewhat easier to build and operate it. Yet, it will still have to obtain a license from the U.S. government to produce chips for Huawei. Furthermore, while the fab is probably meant to give priority to HiSilicon chips, it will be natural for SMIC to land additional orders to increase the utilization of the fab. This begs the question of how exactly the parties will share the output of the fab.

Since Huawei and SMIC are both blacklisted, it will be interesting to see how hard it will be for them to obtain necessary export licenses from the U.S. government (hundreds, if not thousands, of licenses will be required). The US reviews license applications involving these two companies with a presumption of denial.

One Important Benefit

All of Huawei's semiconductor options involve obtaining licenses from the U.S. government: whether it is outsourcing production, co-owning a fab, or even owning a fab. So, what benefits do the latter two options bring? 

The location in China and lack of any actual control from the U.S. government. While Huawei and SMIC will have to obtain technology export licenses required by the U.S. government, they might produce some of the chips before they get appropriate authorizations, or even without the said permissions at all since North America's authorities cannot control what happens in Shenzhen. It remains to be seen what happens if the U.S. govt finds out about the violations, but it is highly likely that only the fab (or rather the company owning the fab) will suffer the consequences.

Summary

Rumors about Huawei's plans to build its own fab and even fab equipment have been circulating for quite some time. With information about the fab as well as details like location and planned investment coming from two sources, the rumor now looks far more credible.   

While it will be inherently hard for Huawei and SMIC to build a new fab (remember that even SMIC itself has problems obtaining all the equipment it needs for its fabs), the benefits in the forms of steady chip supply and lack of actual control of the U.S. government will outweigh all difficulties.  

Huawei has capitalized on its Moscow, Russia presence to invest in the installation of an Arm-based server farm in the city. After a series of blows resulting from the U.S.-China trade wars, Huawei has looked towards diversifying its investments both from a market standpoint and a geo-strategic one. Now, Huawei's Intelligent Computing Systems Department has finished construction of its Moscow Arm-based data center - which the company said was the first such facility in Russia. Huawei said that "the computing cluster will enable our partners and customers to test their software products, to join the construction of a new open ecosystem", and help the company itself develop software on Russian soil.

The infrastructure is meant to solve demand from users who want to capitalize on supercomputing and open source software - one of the reasons why the new data center is mostly focused on serving the Moscow academic and scientific communities. The servers run on the company's Taishan infrastructure, which utilizes Kunpeng 920 processors developed by HiSilicon, the Huawei subsidiary responsible for custom-designing Arm SoCs. Kunpeng 920 SoCs feature Armv8 designs scalable at up to 64 cores running at a maximum 2.6 GHz under a 180W TDP. There's also an octa-channel DDR4 -2933 memory controller in the Kunpeng 920 silicon - it truly was Huawei's latest and greatest at time of its introduction.

The Kunpeng 920 Arm SoCs are manufactured on TSMC's 7 nm node, which already stopped deliveries to Huawei last September, following U.S. sanctions on technology exports to the Chinese company. This means that the server farm was most likely built using stockpiled chips. As for Huawei's foundry access after TSMC shut the company off, and without access to Intel or GlobalFoundries due to their ties with the U.S., Huawei has resorted to China's Semiconductor Manufacturing International Corporation (SMIC), which currently is only capable of silicon manufacturing at the 14 nm node.

This isn't the first time an Arm data center has entered the supercomputing fray; in fact, Fujitsu's Fugaku, an Arm-based supercomputer deployed in Japan, is the world's most powerful supercomputer, without ever touching the x86 instruction set. Likewise, Huawei Russia’s Intelligent Computing System Department Director Lyu Lu said that "The tests carried out have shown that in a number of scenarios Arm servers can already compete fully with solutions based on the x86 architecture.” 

Huawei considers that Arm, too, is capable of a more widespread presence in the server ecosystem, porting over the power/performance benefits of a mobile-first architecture into the server business - and the company will do that one server at a time if it has to.

China might be the last country to weigh in on Nvidia’s proposed acquisition of Arm. The Financial Times today reported that Nvidia “submitted an application to Chinese competition regulators” to review the proposal ”in recent weeks” even though the $40 billion acquisition was officially revealed to the public in September 2020.

It’s not clear why Nvidia started the approval process in China almost nine months after it announced its plan to acquire Arm from SoftBank—especially since Nvidia CEO Jensen Huang said at GTC 2021 in April that he expected the deal to close in 18 months. The Financial Times said it could take that long to receive China’s approval.

Assuming “in recent weeks” means some time after GTC 2021, the purportedly late submission to Chinese regulators could delay the acquisition’s closing to 2023, assuming the deal is approved. That‘s later than Huang predicted just a few months ago, and the controversial nature of the deal could stretch the regulatory approval process. Huang has remarked that he expects the merger to close within the expected timeframe. 

Not that China is the acquisition’s only obstacle. The UK Secretary of State for Digital, Culture, Media and Sport asked the Competition and Markets Authority to investigate the deal in April, and companies like Qualcomm, Microsoft, and Google have reportedly complained about the acquisition to US regulators as well.

In this, it seems Chinese tech companies are aligned with their Western counterparts. The Financial Times said that Huawei’s chip company, HiSilicon, as well as Semiconductor Manufacturing International Corporation, have opposed the deal. (Probably at least partly because of US government restrictions on their businesses.)

It’s entirely possible that regulators around the world are going to approve the Nvidia-Arm acquisition faster than most people expect. Nvidia has repeatedly said that it would continue to remain neutral when it comes to licensing Arm technologies to its competitors; maybe that will be enough for the deal to pass.

But it seems unlikely. Arm pioneer Hermann Hauser has already said that technologies like NVLink used in the upcoming Grace CPU show that Nvidia won’t compete fairly with Arm licensees, and Citigroup analysts said in April that they’re giving the acquisition a 10% chance of being approved because of those concerns.

While Intel's revenue dropped in the first quarter, it is still the world's largest supplier of chips by revenue, ahead of Samsung and TSMC, according to IC Insights. Intel's arch-rival AMD significantly increased its sales in Q1 2021 and is now one of the world's Top 15 semiconductor companies. 

Demand for personal computers and chips grew significantly in the recent quarters, so it is not surprising that the sales of the top 15 semiconductor companies increased by 21% year-over-year in Q1 2021. Intel leads the market with $18.676 billion, yet its semiconductor sales were 4% lower than in the first quarter last year. Samsung followed Intel with $16.152 billion, whereas TSMC was No. 3 with $12.911 billion. 

Samsung, which is the world's largest supplier of 3D NAND and DRAM memory, has been challenging Intel for the top position on the market of chips for years. At times, Samsung can dethrone Intel when memory prices are on the rise, but it was not the case in Q1 2021. 

AMD and MediaTek were two new entrants to the top 15 list of semiconductor companies. AMD significantly increased shipments of its CPUs for servers and high-performance desktops, so its revenue in the first quarter totalled $3.445 billion, up 93% year-over-year. In Q1 2021 AMD was ranked 11th largest semiconductor company in the world. The company was ranked 18 in the first quarter 2020. MediaTek's sales reached $3.849 billion, an increase of 90% year-over-year. AMD and MediaTek replaced Huawei's HiSilicon and Sony's chip division in Top 15. 

The list of top 15 semiconductor sales leaders includes CPU, GPU, and SoC developers, makers of memory, suppliers of special-purpose chips, telecom giants, and contract makers of semiconductors. The list includes eight companies headquartered in the U.S., two firms from South Korea, Taiwan and Europe, and one from Japan.

In a bid to overcome US restrictions on its Arm designs, Huawei's HiSilicon has turned to the open-source RISC-V architecture and has even released its first RISC-V board for Harmony OS developers. Due to being blacklisted by the U.S. government, Huawei and its chip division HiSilicon do not have access to development and production technologies designed in America. The restrictions include many Arm processor architectures, including those used in various microcontrollers that Huawei uses widely.

HiSilicon's HiSilicon Hi3861 development board is based on the company's own Hi3861 controller. Huawei's documentation doesn't disclose exactly what the chip does, but it describes it as a main controller chip. The Hi3861 chip is accompanied by a serial port controller as well as a USB-C port. Among the more important aspects, the Hi3861 seems to have all the logic that enables USB-C functionality (e.g., synchronization and port alignment) and GPIO (general-purpose) pins. 

Overall, the HiSilicon Hi3861 development board has rather vast (at least Raspberry Pi-like) capabilities, but not for a world that Huawei's HiSilicon is used to, at least in terms of public opinion.

The Hi3861 is aimed mostly at the IoT market, whereas HiSilicon's development efforts were historically aimed at high-margin smartphones, tablets, PCs, and embedded systems. But Huawei needs computing platforms to use for its other devices, so the HiSilicon Hi3861 is just what the doctor ordered at this time.

When it comes to high-volume products made by Huawei, a non-Arm-based chip makes exceptional sense and also gives the company experience working with an open-source architecture. Only time will tell if the company uses its RISC-V experience to develop other, more powerful devices. 

Earlier this year Taiwan Semiconductor Manufacturing Co. outlined plans to increase its CapEx budget to $25 billion ~ $28 billion in 2021, many industry observers attributed the increase to building out 3-nm capacities in preparation to produce CPUs for Intel and other large customers. This is not quite the case, according to a forecast by China Renaissance Securities. 

At present, TSMC's N5 capacity is around 55,000 ~ 60,000 wafer starts per month (WSPM), according to a forecast by China Renaissance Securities, which reiterated it earlier this week. Given the fact that TSMC can no longer serve Huawei's HiSilicon, the foundry's main and biggest N5 customer is Apple, which uses this technology to produce its A14 Bionic as well as M1 system-on-chips for its latest products. Apple is one of TSMC's key customers that has early access to the latest process technologies and is among the first to adopt the newest nodes. 

Later this year other TSMC's customers, such as AMD and Qualcomm, are projected to start using the manufacturer's N5 fabrication process, which is when demand for this technology increases significantly. In a bid to satisfy demand for its N5 node in 2021 and in the following years, TSMC will spend the lion's share of its CapEx on expansion of its N5 capacity, the analysts say.  

China Renaissance believes that TSMC will increase its N5 capacity by two times compared to today's levels, or to 110,000 ~ 120,000 WSPM, which essentially means that a year from now TSMC will have a GigaFab dedicated solely to N5 family, which includes N5, N5P, and N4 technologies. All of these processes are compatible on design rules, IPs, and SPICE levels, they will be used for years to come. 

Another long node for TSMC will be its N3 (3 nm) technology. When compared to N5, the N3 process is expected to provide up to 15% performance improvement (at the same power and transistor count), or up to 30% power reduction (at the same speed and transistor count), along with up to 20% higher SRAM density and up to 70% higher logic density. TSMC expects N3 to enter risk production later this year, whereas volume production is expected in the second half of 2022. TSMC says that the interest towards N3 is higher than the interest towards N5 and N7 at the same stage of development (i.e., five to six quarters before formal launch). 

Keeping in mind that N3 will not be needed for high volume manufacturing (HVM) for another year, it is logical for TSMC to spend big on N5 in calendar 2021 rather than prep for N3 in 2022 just now. The company is still expected to build a pilot N3 line with a capacity of around 10,000 ~ 15,000 WSPM this year.  

Both N3 and N5 nodes extensively rely on extreme ultraviolet (EUV) lithography. To expand N5 capacities and equip an N3 fab, TSMC will need to procure loads of ASML's Twinscan NXE scanners and deliver them. EUV tools take around six months to calibrate, so TSMC's expansion plans have to be very precise in a bid get the equipment ready exactly when it needs them. 

The main uncertainty for TSMC (and to a large degree the semiconductor industry) today is Intel's outsourcing plan. So far, the company has disclosed plans to outsource certain GPUs and SoCs to TSMC and it is believed that these chips will be made using TSMC's N5 or N7 nodes. Yet, it is unclear whether the chip giant intends to outsource production of its mainstream and performance CPUs to TSMC.  

From transistor density point of view, Intel's 10nm SuperFin fabrication process (~100 MT/mm2) is comparable to TSMC's N7+ technology (~115 MT/mm2), but Intel's own technology might still be a better fit for the company just now as its CPUs have been tailored for this node. Meanwhile, TSMC's N5 has a significant advantage in terms of transistor density (~170 MT/mm2), which is something hard to ignore particularly when it comes to GPUs. When Intel is ready with its 7 nm technology (which is projected to double its existing transistor density) sometimes in 2023, it will still be behind TSMC's N3 by two or three quarters, which means that it will make sense for Intel to outsource some of its products to TSMC, exactly what the new chief executive Patrick Gelsinger told several weeks ago. Meanwhile, from CapEx allocation point of view it makes more sense for Intel to outsource a substantial portion of its products to TSMC and save on fabs.  

"In the interest of engineering resource/CapEx allocation, Intel would be unlikely to run a hybrid internal-external foundry production for the same process, in our view; thus we expect any outsourcing decision will be a binary outcome (i.e., “winner takes all”, others take nothing)," wrote Szeho Ng, an analyst with China Renaissance Securities, in a note to clients. "Given that TSMC has clearly been well ahead of Intel in terms of EUV tech and capacity readiness, we expect CPU outsourcing for Intel to be the best path forward considering its current stance." 

If Intel proceeds with the plan to outsource a significant portion of its production to TSMC in a bid to catch up with AMD in 2022, then it remains to be seen whether the foundry will have enough capacity to serve Intel and its existing customers, especially keeping in mind the fact that demand for N3 node may be higher when compared to demand for current nodes due to AI, HPC, and edge computing megatrends. 

While it makes a great sense from financial point of view for Intel to CapEx save money and use TSMC's manufacturing capacities to build its products while concentrating on architectural innovations, it may not be exactly a good plan for Intel's long-term future. TSMC already has more EUV experience than Intel and it also has a certain standard-setting capability in the EUV ecosystem. If Intel does not catch up, then it will follow for many years to come and will not be able to use its superior fabrication processes as a competitive advantage over rivals. 

It is not a secret that it is extremely hard to get AMD's latest Radeon RX 6800/6800 XT graphics cards and Ryzen 5000-series processors (see our article on how and where to find Radeon RX 6800 XT and how and where to find Ryzen 5000). Officially, the company says that demand for these parts is so high that it greatly exceeds supply. Meanwhile, a media report says that AMD cannot get additional chips from Taiwan Semiconductor Manufacturing Co. because the foundry is busy making chips for Apple's latest devices as well as new consoles from Microsoft and Sony.  

Leading-Edge Capacities Fully Booked

TSMC can process a relatively limited number of wafers using its leading-edge 5 nm and 7 nm process technologies. Therefore, all of TSMC's clients that need these nodes sign wafer purchase agreements to allocate a certain number of contract wafers per quarter. Such agreements are signed many quarters before (and in case of alpha customers, who get access to leading-edge processes ahead of everyone else, years before) actual production starts and cannot be changed overnight. Large companies tend to commit for huge volumes, whereas smaller players usually allocate a lower number of wafers. 

For Q4 2020, all of TSMC's 5-nm and 7-nm capacities have been booked, according to a report from Commercial Times, which did not disclose its sources. TSMC's board of directors recently approved a plan to expand its leading-edge capacities, but it will take months before new equipment is installed.  

Apple's Need for Chips

In the recent years Apple and Huawei's HiSilicon were TSMC's alpha customers for leading-edge nodes and could access these technologies before others. Now that TSMC cannot supply chips to Huawei, Apple booked all of the foundry's 5-nm capacity for its own A14 system-on-chips for iPhone 14 and iPad Air 4 as well as M1 SoCs for the latest Macs, the report claims. The story says that TSMC is contracted to ship 150,000 wafers containing A14 SoCs and 18,000 wafers with M1 processors in Q4 2020. These numbers cannot be verified as contractual obligations are confidential.  

Since Apple's A14 does not have a built-in modem, it is paired with Qualcomm's X55 modem that is made by TSMC using one of its 7 nm manufacturing technologies. Since Apple sells boatloads of iPhones and Qualcomm itself ships tons of its own 7-nm mobile SoCs to various customers, the company became TSMC's largest 7-nm client in Q4 2020. Qualcomm's shipment of X55 modems to Apple alone is projected to total 80,000 wafers this quarter, Commercial Times reports. 

The New Consoles

AMD is considerably smaller than Apple or Qualcomm in terms of volumes, so it never allocates (or rather cannot allocate) the number of wafers that SoC companies do. Meanwhile, in Q3 and Q4 2020, the company has to supply certain pre-ordered 7 nm SoCs for Microsoft's Xbox Series X/S as well as Sony's PlayStation 5 game consoles. Since AMD has supply agreements with its console partners, it has to prioritize production of their SoCs over its own CPUs and GPUs in a bid to enable Microsoft and Sony to ramp up production of their new game machines. 

Commercial Times claims that AMD is contracted to ship around 120,000 wafers containing SoCs for the latest game consoles (40,000 to Microsoft and 80,000 to Sony) in Q4 2020. Meanwhile, AMD's contracted 7 nm allocation for the fourth quarter is about 150,000, the report says, which essentially means that 80% of AMD's 7 nm wafers will be used for consoles. 

Can't Get Enough?

Demand for PCs, computer hardware, game consoles, and other digital devices is indeed very high these days as many people now work from home and entertain themselves at home. Such demand could not be predicted several quarters ago, so it is not particularly surprising that there are shortages of the latest hardware and game consoles.  

If the numbers reported by Commercial Times are correct, then most of the 7 nm chips produced for AMD this quarter will go to Microsoft and Sony as per contracts, whereas some previous-generation 7 nm products will be shipped to PC OEMs in accordance with obligations. As a consequence, only some wafers will be left for AMD's new Radeon RX 6000 as well as Ryzen 5000-series products. Keeping in mind that TSMC's 5 nm and 7 nm capacities have been fully booked for Q4, AMD cannot order any additional wafers just now.

A Huge Grain of Salt

While the claim that AMD cannot get enough Big Navi and Vermeer chips because TSMC's production capacities are busy making chips for Apple (the company still consumes 7 nm SoCs), Qualcomm, Microsoft, and Sony sounds reasonable, it should be noted that the number of wafers shipped by TSMC to its clients cannot be verified. Furthermore, actual allocation of wafers to a particular company is a trade secret and nobody will ever confirm it. 

If the numbers published by Commercial Times are inaccurate, then the whole story either contains inaccuracies, or is completely wrong. 

But What About Nvidia?

AMD is not the only company that cannot meet demand for its latest products these days. Nvidia's latest GeForce RTX 30-series 'Ampere' graphics cards are in very tight supply (see where and how to buy an RTX 3080, 3070 or 3090). Since it is so hard to buy the new cards from official channels, the scalping industry flourished in the recent months. 

Nvidia produces its GA102 and GA104 graphics processors at Samsung Foundry using 8N process technology. It is unknown how much capacity Samsung has for its N8 node and how many wafers for Nvidia it can process. Meanwhile, Nvidia said that demand for its latest GeForce RTX 3070/3080/3090 products greatly exceeded its ability to supply, which is why shortages would continue into 2021. 

Bottom Line

It is evident that demand for high-tech products is high these days as people spend more time at home. Some products are in short supply because of insufficient production volumes and others are rare because of scalpers (yet again, scalpers can only make profit if a product is in tight supply). But the sad fact is that numerous companies that cannot meet demand for their products now are also not going to meet it for quite some time.

According to a report from Asian Nikkei Review, TSMC stated yesterday that it has stopped taking orders from Huawei since May 15. The foundry also confirmed that it will not ship wafers to Huawei after September 14.

New U.S. regulations stipulate that non-U.S. companies that utilize American technology or tools must apply for an export license to sell to Huawei. TSMC Chairman Mark Liu didn't comment on whether TSMC will apply for a license to continue doing business with the Chinese tech giant. 

Despite Huawei's exit and the ongoing COVID-19 pandemic, TSMC is optimistic about its future. The foundry estimated revenues between $11.2 billion to $11.5 billion for the third quarter of this year. With Huawei out of the picture, TSMC has freed up a lot of production capacity. 

Liu didn't mention any names, but he affirmed that TSMC is working in close collaboration with its other clients to fill the gap Huawei left behind.

Apple could benefit most from all this. The tech company is already one of TSMC's biggest clients, since the foundry produces the mobile chips inside the iPhones. However, a recent Digitimes report claims that Apple will tap TSMC to produce the its custom Arm-based Apple Silicon, which could further bolster the existing relationship between the two companies. 

Then there are other star clients, such as AMD, Broadcom, Nvidia and Qualcomm. We think TSMC will cope just fine without Huawei.

Huawei had already started obtaining its chips from other sources, with one of them being Semiconductor Manufacturing International (SMIC). However, the orders were for 14nm silicon. 

Notably, SMIC or any other Chinese fab simply can't compete with TSMC because its technology is light years in front of the competition. It looks like a rough road ahead for Huawei, and it wouldn't be surprising if the company eventually has to pull the plug on HiSilicon, its chip division.

Huawei, through its HiSilicon subsidiary, has a line of promising 7nm ARM v8-based Kunpeng processors that stretch up to 64 cores for the data center and support leading tech, like PCIe 4.0. Now at least one model of the chip is being used for desktop systems, too. Chinese YouTube channel 二斤自制 purchased and tested a Huawei-powered desktop PC that features both the company's eight-core eight-thread 7nm Kunpeng 920 ARM v8 processor and the Huawei D920S10 desktop motherboard in a third-party system, giving us the first glimpse of the new products enabled by Huawei's recent entrance into the market as a supplier to OEMs that produce desktop PCs. 

The development could help further China's targeted strategy to reduce its reliance on western semiconductor technology. Still, in many ways, the system highlights the difficulties the country has encountered, particularly in terms of software support. In fact, that's the primary focus of the video. The video doesn't give us much in the way of broadly-comparable benchmarks (though there are a few tidbits), but we do learn some specs that we'll cover below.

The narrator spends much of the video covering the problems she encountered with running meaningful software applications. Due to Kunpeng's ARM architecture, the system is limited to running the China-produced 64-bit UOS operating system that is largely a modified flavor of Linux. The narrator commented that the UOS operating system runs smooth and has an intuitive interface, and it even supports a 4K resolution at 60Hz via a Yeston RX550 graphics card. Still, she had to pay an extra 800 Yuan (~$115) to gain access to the app store. Moreover, the store had a woeful selection of applications, lacking such staples as Adobe and other apps. That's exacerbated by the system's lack of support for 32-bit software.

The system did run a Blender BMW test render, but it completed in 11 minutes and 47 seconds, which is woefully slower than most modern chips. The system did well streaming 4K video, but 'choked' during local playback due to poor encoding performance. The narrator says the system is obviously best for light office work only. 

The channel purchased the system for 7,500 Yuan (roughly $1,060 USD), and it comes with an eight-core eight-thread 2.6 GHz Kunpeng 920 2249K processor soldered to the motherboard. We can't find specs for this processor online, but the video lists it with 128K of L1 memory (64K I$, 64K D$), 512K of L2, and 32MB of L3. 

The Huawei D920S10 motherboard has four DIMM slots, but the system only has 16GB of Kingston DDR4-2666 memory spread across two DIMMs. Despite the chip's seeming support for PCIe 4.0, there are only three PCIe 3.0 slots available (x16, x4, x1), and the motherboard's connectivity options are also pretty mundane (6 SATA III ports, two M.2 slots, two USB 2.0 and USB 3.0 ports, and a VGA connection). The board also has a Gigabit ethernet port, or you have the option to use an optical connection of an unspecified speed. 

A 256GB SATA hard drive, 200W power supply, and a Yeston RX550 graphics card round out the other accommodations. The system also comes with an optical drive.

The difficulties the channel encountered with software availability, and performance, highlights that even the best chips in the world aren't very effective without a robust software and developer ecosystem. That's yet another facet of the challenge that China faces as it looks to reduce the amount of silicon it procures from external vendors, and the Huawei-powered systems could be designed to help foster a developer ecosystem for the ARM architecture and the UOS operating system. 

According to IC Insights, China-native vendors only produce 6.1% of the country's total silicon consumption, and a recent report indicates the country will fall far short of its 2025 goals for 70% semiconductor self-sufficiency, instead only hitting one-third of its original target. 

The country has invested heavily in multiple native semiconductor producers and their projects, with its multi-pronged efforts including chips like the x86 Zhaoxin KaiXian processor we recently tested. The EPYC-based Hygon Dryhana x86 processors were also developed under a joint venture with AMD that was later scuttled by the US government, and we've also seen signs that Huawei is testing a desktop PC design with AMD's Ryzen 4000 series processors. The latter shows that Huawei might also be pursuing an x86 route into the desktop PC market, albeit with non-native processors.

HiSilicon, Huawei's elite chip division, has placed 14nm orders with Semiconductor Manufacturing International (SMIC), China's top chipmaker, according to a DigiTimes report today citing unnamed industry sources. 

HiSilicon's Kirin 710 smartphone processor, which is based on TSMC's 12nm FinFET node, has been out since mid-2018. There is a rumor that HiSilicon is planning to release a gimped variant of the Kirin 710, the Kirin 710A. Kirin 710A is expected to leverage the 14nm FinFET process and, thus, HiSilicon needs a fab that's capable of pumping out the chips.

If DigiTimes' report is true, SMIC has scored a big victory over TSMC, as the Taiwanese foundry has been a core provider of chips for HiSilicon. The recent change would indicate that SMIC's 14nm FinFET manufacturing process has improved to a point where it can go head-to-head against TSMC's process node. Furthermore, the U.S. may block TSMC from selling to Huawei, so that could be another reason why HiSilicon decided to switch from TSMC.

• The best (non-gaming) performance CPUs out now
• The best gaming CPUs out now

SMIC's first-generation 14nm FinFET process has been up and running since the fourth quarter of 2019. The Chinese fab's financial data reveals that the node contributed to about 1% of the company's total wafer revenue in Q4; however, SMIC plans to ramp up production progressively this year.

Although SMIC might seem like an underdog compared to TSMC, the Chinese chipmaker shouldn't be underestimated. SMIC is looking to skip the 10nm node completely and transition straight to the 7nm node. The company expects to run risk production of its 7nm process by the end of 2020. TSMC's EUV process is proceeding smoothly. HiSilicon is allegedly on TSMC's client list to leverage the EUV node; however, that'll depend on the outcome of the proposed U.S. block on Huawei.

China continues to push hard to phase out foreign tech and achieve self-sufficiency. SMIC is one of the pieces on the board that China can use to accomplish its goals.

Despite the downturn of events around the world, TSMC is witnessing a significant increase in demand for its Chip-on-Wafer-on-Substrate (CoWoS) packaging, according to DigiTimes' unnamed industry sources. The Taiwanese silicon manufacturer is purportedly running its CoWoS production lines at full capacity.

CoWoS as  is a 2.5D method of packaging multiple individual dies side-by-side on a single silicon interposer. The benefits are the ability to increase the density in small devices as you run into the limits of how big individual dies can be produced, better interconnectivity between dies and lower power consumption.

According to DigiTimes, AMD, Nvidia, HiSilicon, Xilinx and Broadcom have placed orders for the tech, with demand for high-performance computing chips, high bandwidth memory (HBM)-powered AI accelerators and ASICs during the past two weeks.

Examples of CoWoS packaged silicon are AMD's Vega VII graphics cards, as well as Nvidia's V100 cards, which have HBM on the same silicon interposer where the GPU is. With the GPU and memory so close together, memory bandwidth is significantly higher on these chips compared to those using GDDR6 memory located elsewhere on the graphics card's PCB. Additionally, the PCB becomes much smaller.

Last month, TSMC announced its new CoWoS tech that packs a whopping 1700mm-squared of silicon onto a single interposer. With this change, the interconnect bandwidth was also increased to a staggering 2.7 TBps, which is a 2.7 times increase over TSMC's 2016 technology.

Keep in mind that this tech doesn't come cheap and remains a high-end solution for product manufacturing. It is unlikely to come to best graphics cards targeting consumers market anytime soon but does offer a glimpse of where the technology is going.

Researchers have uncovered a backdoor in firmware made by China-based Hangzhou Xiongmai Technology and used in digital video recorder (DVR) and network video recorder (NVR) cameras often used for surveillance. Detailed this week, the issue affects many brands that have licensed the firmware from Xiaongmai. You can find a full list of affected devices here. 

In 2016, Xiaongmai was also hit with a recall order by the U.S. government, due to the Mirai botnet enabling backdoors in its products.

Earlier versions of the firmware relied on access to telnet, a network protocol for remote connections, to be enabled by default. The researchers said it did this through use of "a static root password which can be recovered from firmware image with (relatively) little computation effort." A Russian security researcher found this hardcoded backdoor for the first time in 2013. He also discovered multiple remote code execution bugs in the built-in server.

More recent firmware versions did come with disabled telnet access but also had open port 9530/tcp "listening for special commands." Xiaongmai “upgraded” the remote root access connection by requiring cryptographic authentication. But while these upgrades may have made it more difficult for bot makers or other attackers to gain full access, it doesn’t change the fact that Xiaongmai has maintained its (now slightly more hidden) remote root access to all security camera systems powered by its firmware.

The company still kept a short list of static passwords it could use to login remotely and take over anyone's security camera system, just as any other third-party could, too, once they'd discover the backdoor. The good news is that this sort of poor security hygiene and use of default passwords will soon be illegal at least in the European Union and United Kingdom, as both have recently said that they will ban default passwords on Internet of Things (IoT) devices.

Even if the backdoors were not maliciously created by Xiaongmai, these sort of security issues leave the door wide open for other malicious parties to come right in and take over millions of IoT devices at once via bots. These bots can then go on and wreck havoc against online services, while also making the backdoored products dysfunctional.

Huawei’s HiSilicon chip design unit has launched 4G communication chips on the open market. It is the first time HiSilicon is supplying its chips to external companies, marking a change in strategy, for which it has formed the Shanghai HiSilicon wholly-owned susidiary.

HiSilicon, fully owned by Huawei, is reputed to be the largest Chinese designer of integrated circuits, but Huawei has kept is an internal unit, meaning that it would design and supply chips only for Huawei. For instance, the company has the Kirin series for smartphone SoCs, Balong modems and Ascend AI chips. This now appears to be changing.

HiSilicon would no longer be Huawei’s captive chipmaker, as it has launched 4G communication chips on the open market at a Shenzhen electronics show that was held in October, according to EETimes.

The launch was done by Shanghai HiSilicon, a wholly-owned subsidiary of HiSilicon that was formed in April 2019. Shanghai HiSilicon has been established as the chip design company for the open market, while Shenzhen HiSilicon will remain responsible for Huawei and HiSilicon’s internal chip activities.

This change of strategy has come with the tagline ‘Everything Connected’ as the company is broadening its scope. To that end, HiSilicon has changed the name of some of its product lines. Its security camera product line has been renamed to Smart Vision and wants to expand to consumer and automotive. Its set-top and TV lines are merging into a Smart Media product line. And thirdly, it wants to expand its connectivity business and pursue AI/IoT, shortened as AIoT.

MyDrivers reports that China's homegrown Zhaoxin processors are now operational with its self-developed Unity Operating System (UOS). It's no surprise that China wants to stop depending on U.S. technology, and the latest development is an important step towards the country's autonomy. The news come on the tail of new Chinese government restrictions that force Chinese institutions to replace all foreign-powered PC hardware and software with domestically-produced products over a three-year span. 

Tongxin Software, the developer behind UOS, has been able to get the KaiXian KX-6000 and KaiSheng KH-30000 processors to work on the desktop and server versions of UOS, respectively. The aforementioned chips are Zhaoxin's latest 16nm parts that sport eight cores with base clock speeds up to 3 GHz. The Chinese CPU maker has already laid out its ambitious plans for the company's next generation of processors.

Thanks to the new integrated display driver in UOS, the KaiXian KX-6000 processors offer improved video and audio performance. The chips also support 3D graphics and ultra-high-definition video decoding, among other features. On the server side, the company says the KaiSheng KH-30000 parts have exhibited excellent stability, and the chips should work fine inside file and database servers.

Tongxin Software will continue to work hand-in-hand with Zhaoxin to optimize and improve application performance with the chipmaker's processors. Tongxin Software is also committed to getting other Chinese-made chips, such as Loongson, FeiTeng, ShenWei, Kunpeng and Hygon, to work with UOS. In due time, the country hopes UOS will have matured enough to replace Microsoft's Windows operating system completely.

The China Times reports that TSMC's upcoming 5nm process has already hit yields of 50%, meaning 50% of all processors yielded on a 5nm wafer are fully functional. The report also specifies that AMD's Zen 4 architecture will be fabbed on the 5nm process.

Although the yield rates have reportedly reached 50%, the other 50% of these dies are not totally broken but can be salvaged as processors with fewer cores enabled, for example. TSMC's 5nm is also entering risk production, which includes the very first wave of products that are expected to help iron out any kinks so TSMC can bring all of its 5nm capacity fully online.

The new 5nm node reportedly brings a 1.8x greater density over 7nm and 15% higher clock speeds at the same power, or reduce power consumption by 30% (meaning processor designers will have to choose one or a blend of both). This is not quite as good as the jump from 16nm to 7nm was, but as many are aware, silicon is reaching its limits rather soon, and we should expect density and power efficiency gains to slow down a little.

Two new processors, Apple's A14 and Hisilicon's new Kirin processor, have reportedly already been taped out on 5nm. The chips will supposedly be mass-produced in July. The China Times also points out AMD's Zen 4 CPUs will be based on 5nm. The node almost doubles density, so perhaps we might see a doubling in core count, though a 50% increase is more likely because 1.8x isn't quite 2x.

Zen 4 is, of course, quite a ways off, expected in 2021 at the earliest, and possibly even 2022. Since it is on a brand new node, it's easy to expect we will see some packaging difference between it and Zen 2/3. Perhaps AMD will increase the number of cores per die, or increase the number of dies and keep them all at eight cores each.

Meanwhile, Samsung and Intel (TSMC's only real competition), have been struggling to get out new nodes. Intel suffers the most and is still on its 2014 14nm process for the vast majority of its CPUs because its 10nm development was extremely troubled. Intel's 10nm process is finally live but the company only uses it for a handful of laptop CPUs, likely due to supply constraints.

Samsung is apparently going to manufacture Nvidia's upcoming 7nm GPUs, but that remains to be seen. It's not clear how well Samsung's 7nm process compares to TSMC's. TSMC certainly seems like the obvious winner today, but Intel isn't out of the race quite yet, as its highly anticipated 7nm process could bring the company back into the forefront of CPU manufacturing.

We can likely expect any processor sporting TSMC's upcoming 5nm process to be significantly ahead of any chip on an older process, though the lead won't be quite as pronounced as it used to be between a new node and an old node. Because of this, we can also expect that 5nm capacity will be in high demand just as 7nm is, so hopefully, TSMC has enough supply for the whole world to share.

In yet another sign that Huawei is seeking to enter new markets, the company has posted details about a new desktop PC motherboard that is designed to house the company's Kunpeng ARMv8 processors that it typically uses for server infrastructure. These chips will come in four- and eight-core 7nm flavors for the initial desktop motherboards, but scale up to 64-cores and support PCIe 4.0 for the full-fledged server models. The new desktop PC motherboards and reference system designs, not to mention access to modern core-heavy chips, indicates that Huawei has plenty of options if it makes a serious effort at grabbing a piece of the desktop PC market. 

There's been no shortage of ink spilled on the US government's restrictions levied against Huawei in the name of national security, but those restrictions are increasing the company's resolve in developing its own alternatives to American chips.

The trade restrictions against Huawei initially threatened the company's use of ARM processors for its smartphones and laptops, but recent developments have found ARM insisting that it will continue to license its technology to Huawei, and its silicon-spinning subsidiary HiSilicon, after determining that the technology isn't subject to US trade restrictions. 

And ARM's continuing support of HiSilicon is key for Huawei: the company is reportedly the largest consumer of TSMC's 7nm silicon that is in high demand by a wide range of industry behemoths, including Apple, Nvidia, and AMD, and those chips currently power 70% of Huawei's record shipments of smartphones. 

We recently came across a report from ithome.com detailing Huawei's announcement and display of a new desktop PC motherboard at its Full Connect Conference in September. The Kunpeng Desktop PC motherboard certainly caught our attention, so we sent queries to several US-based Huawei representatives, but we couldn't get any direct answers to our questions, such as pricing or regional availability. We did finally receive contact information for a representative in China, but we still haven't received a response after several queries. The silence is understandable given the current political climate, but it does leave our questions unanswered: Will we see these motherboards shipped en masse into standard desktop PCs, and when and where?

The motherboards, not to mention HiSilicon's Kunpeng chip, are impressive. According to the posted specifications, the Kunpeng Desktop Board D920S10 supports the PCIe 3.0 interface, six SATA 3.0 hard interfaces, and two M.2 SSD slots. Memory consists of a quad-channel DDR4-2400 interface that supports a maximum of 64GB of capacity, and the board supports ECC, too. It also has an embedded Ethernet controller and naturally supports additional networking cards up to 25 GbE. For mundane everyday connectivity, the motherboard also supports four USB 3.0 and four USB 2.0 ports, giving it all of the requirements for a desktop PC. 

Huawei's website also indicates that it will provide reference guides for chassis, cooling, and power supplies, so it appears this design will be opened up to OEMs and ODMs.

Huawei's documentation mentions a slant towards office applications and support for Linux, but the capabilities of the 7nm HiSilicon Kunpeng 920 processor and its TaiShan v110 microarchitecture are impressive.

The Kunpeng desktop motherboard supports dual- and quad-core 920's, but the Kunpeng product stack extends up to 64 core server chips that run at 2.6 GHz and support up to 1TB of eight-channel memory. These chips come armed with 20 billion transistors spread over three dies in a multi-chip module, indicating there is scalability, and support up to DDR4-3200 and 40 lanes of the PCIe 4.0 interface, not to mention support for the CCIX protocol, meaning these chips rival AMD's market-leading desktop silicon in terms of connectivity. And they feature the same 7nm process as AMD's chips, too, which brings about performance and efficiency advantages. For instance, the full-featured 64-core chips have a relatively tame 180W TDP. Those server-focused processors can also scale up to four sockets. 

Huawei's S920X00 server motherboard also appears to be quite robust, with support for two Kunpeng 920 processors, 16 storage devices in SATA, SAS, or NVMe flavors, up to 32 memory DIMMs spread across eight channels, and PCIe expansion.

This means that Huawei could easily scale its freshman effort at a desktop PC motherboard up with more capable motherboards and processors, and the company certainly has the know-how to develop those products. With years of PCB design for smartphones under its belt, and an impressive server variant of the motherboard already available, it's conceivable that Huawei could fill in the gaps in its product stack to scale from low-power office machines to robust desktop PCs, powerful workstations, and anything else in-between. 

But the specter of x86 compatibility looms large. Enterprises and normal users alike have balked at adopting the ARM architecture for desktop PCs and servers due to a lack of an established mainstream software ecosystem and kludgy x86 emulation, but those efforts have recently begun to reach critical mass as Microsoft reportedly plans to step up its 64-bit app support for Windows 10 on ARM. Amazon also announced its new Graviton 2 processors yesterday, with new EC2 instances being deployed that will help foster the ARM ecosystem further, which is especially important given the heft of the company's backing with AWS.

Huawei has also been developing its own operating system in the background since 2012, so the company has several options to address teething pains. Given the continuing support from ARM Technology, Huawei can also move forward to newer ARM architectures for its chips. 

However, none of this means that Huawei has meaningful plans for making a brute-force entry into the desktop PC market, but the company certainly has the capability to throw its financial and engineering muscle into the project. And it already has the first product. Unfortunately, the company's plans aren't clear, and given the lack of communication, it appears we won't know more any time soon.  

Huawei today announced its Kirin 990 5G system on a chip, which it calls the “world’s first flagship 5G SOC.” It made the announcement here in Berlin at its IFA keynote.

The SOC is produced on a 7nm FF+ EUV process and has 10.3 billion transistors on board. Huaweiclaims to have a board area 36 percent smaller than its competitors, Samsung and Qualcomm.

This will compete with the Samsung Exynos 980, which that company announced on Wednesday in a surprise drop. That chips is aimed at the mid-range, however, while Huawei is considering the Kirin 990 5G a flagship chip. Huawei is also announcing the Kirin 990, a slightly less powerful chip without 5G support.

Huawei suggests that the chip and receivers will use machine learning to ensure high data speeds. When you enter a weak 5G signal area, the modem will split uplink between 4G and 5G. Kirin 9905G will support both 5G non-standalone (NSA) and 5G standalone (SA) networks.

It claims that the chip will offer a downlink up to 2.3Gbps and uplink of up to 1.25Gbps. The chip supports four sub-6 antennae.

As for the image signal processor (ISP), Huawei is using its fifth-generation dual ISP, and it is claiming improvement in throughput, power efficiency, image and video noise reduction. It also uses block-matching and 3D filtering (BM3D) for what it says will result in DSLR-style noise reduction on its phones.

The GPU on the Kirin 990 5G is the 16-core Mali-G76; that’s up from 10 cores on the Kirin 980. Huawei claims that a smart cache will help with gaming, reducing GPU-to-memory bandwidth needs by 15% and reducing DDR power consumption by 12%.

The neural processing unit uses Huawei’s Da Vinci architecture and a new “Tiny Core” in combination, which Huawei claims makes it much more efficient. It said that 90% of CV neural networks are supported.

We haven’t yet heard when these new Kirin will be released, or which of Huawei’s devices they’ll wind up in. CEO Richard Yu suggested on stage there will be more news at the launch event for the Huawei Mate 30 Series in Munich later this month. There’s also the question of course, of whether or not you’ll actually be able to buy any of Huawei’s phones or tablets in the US once they actually arrive with these new SoCs.

TSMC’s 7nm process technology will not be fully utilized in the first half of 2019 due to order cutbacks from several customers, according to an article from Chinese-language publication Commercial Times and reported by DigiTimes reported this week. This could leave some room for AMD and other chip makers to further increase their chip production and shipments in early 2019.

According to this report, TSMC’s 7nm process is expected to be utilized at only 80-90 percent of full capacity during the first six months of 2019. Apple, HiSilicon and Qualcomm have all cut their expected orders for the first half of 2019, according to the publication. TSMC has declined to comment on the reports. 

Recent reports have claimed sales for Apple's latest iPhone have been  much lower than expected, likely due to the company increasing price points even further this year. These events also coincide with Apple announcing that it will no longer disclose the number of units sold to investors, a first since the company introduced the iPhone a decade ago. This gives credence to the idea that Apple’s sales may continue to decline in 2019 if the company is unwilling to go back to previous price points for its devices.

Opportunity for AMD

AMD is expected to launch its next-generation CPUs based on the Zen 2 architecture in the first half of 2019. Because other TSMC customers have cut back on their expected sales, AMD is now free to increase its planned production of Zen 2 chips.

An increase in planned production would still depend on how much demand there is for these chips, too, but going by what we know so far about these chips, AMD could see a significant increase in sales for Zen 2 chips compared to the Zen 1-based chips. The 7nm Zen 2 chips are expected to be much more competitive in performance compared to Intel’s chips.

TSMC is expected to provide an outlook update to its investors in mid-January. The company previously disclosed that it was set to tape out over 50 chip designs with its 7nm process technology and over 100 chip designs in total with both of its 7nm and 7nm EUV nodes by the end of 2019. TSMC expects 7nm wafer revenues to exceed 20 percent of its total revenues in 2019.

Huawei CEO Richard Yu announced the company’s new Kirin 970 SoC (system on a chip) with built-in AI capabilities at IFA, Europe's biggest tech show held each year in Berlin.

The Kirin 970 features an 8-core processor with a 12-core GPU manufactured by TSMC using a 10nm process. The chipset is equipped with four ARM Cortex-A73 cores running at 2.4GHz and four ARM Cortex-A53 cores running at 1.8GHz. This chipset is also the first commercial SoC to use the the Mali-G72 second generation Bifrost-based GPU from ARM.

The chipset features a total of 5.5 billion transistors, and is the first SoC to have a dedicated Neural Processing Unit (NPU). This means the Kirin 970 chipset will offer on-device AI, as opposed to phones that need to be online for AI interaction. All these features are packed into a chipset that measures just 1cm².

Yu said:

As we look to the future of smartphones, we're at the threshold of an exciting new era. Mobile AI = On-Device AI + Cloud AI. HUAWEI is committed to developing smart devices into intelligent devices by building end-to-end capabilities that support coordinated development of chips, devices, and the cloud. The ultimate goal is to provide a significantly better user experience. The Kirin 970 is the first in a series of new advances that will bring powerful AI features to our devices and take them beyond the competition.

The company claimed that, due to the Kirin 970's heterogeneous computing architecture, this new chipset can perform AI computing tasks up to 25x faster with 50x greater efficiency than a quad-core Cortex-A73 CPU alone. To demonstrate the processing power of this chipset, Huawei offered up the results of a benchmark image recognition test showing the Kirin 970 processing 2,000 images per minute.

Huawei said at its IFA keynote  that the Kirin 970's chipset and mobile AI platform will be open to developers and partners around the globe.The company plans to use its flagship processor to power its upcoming Mate 10 series phones, which are set to be unveiled on October 16.

Introduction & Updates

With our increasing mobility and addiction to information, smartphones—our powerful pocket computers that help us create, locate, communicate, and vegetate—have become an indispensable part of modern life. Our reliance on these wireless wonders will only increase as they become increasingly intelligent, filtering our data and predicting precisely when we’ll need it most, even performing actions on our behalf, eventually. This is why it is so important to pick the phone that best serves you—at least until the roles are reversed.

Updates

In the time since our last update, we've been busy evaluating new SoCs and CPU architectures. Our preview of Qualcomm's new Snapdragon 820 SoC discussed the company's focus on heterogeneous computing and how this influenced the design of its first custom 64-bit CPU, Kryo. Our performance tests showed an emphasis on floating-point performance and sequential memory bandwidth. The 820's new Adreno 530 GPU is also a beast, setting new records in nearly every graphics test.

Next, we plunged into the architecture of ARM's Cortex-A72 CPU, which replaces the Cortex-A57 as its flagship 64-bit processor. While an evolution of its previous design, ARM made a number of tweaks to improve performance and, more importantly, reduce power consumption. We got our first look at the A72 as well as ARM's Mali-T880 GPU in HiSilicon's Kirin 950 SoC that makes its debut in the Huawei Mate 8 smartphone. Our initial impression was positive, and it will help make 2016 an exciting year for mobile SoCs.

All of this testing and analysis (and CES) has kept us pretty busy lately, but we've been working on product reviews too. Our "long-term evaluation" of the iPhone 6s and 6s Plus is complete, culminating in a new flagship phone recommendation. Apple has vastly improved the iPhone's user experience by increasing RAM to 2GB and adding its new 3D Touch feature.

Before choosing a smartphone, of course, you have to pick an ecosystem to play in, whether it's Apple's, Google's or Microsoft's. This choice is highly personal and depends on what you do with your phone and what criteria are important to you. Our picks focus on hardware and user experience and leave this bigger choice up to you.

Our Best Picks

MORE: How We Test Smartphones
MORE: Apple iPhone 6 & 6 Plus Review
MORE: Google Nexus 6 Review


MORE: All Smartphone Content

There are many factors to consider when choosing a smartphone: size, performance, features, software, price. Our comprehensive reviews give you the knowledge and data to make informed decisions, but if you do not have the time to read our smartphone novellas, or feel overwhelmed by all the charts and numbers, we also give you a shortcut. Every month, we publish our top picks in several different categories and tell you what makes these particular phones stand out.

While we could categorize phones based on ecosystem (Apple, Google, Microsoft) or screen size, we decided to differentiate based on price, with one exception. Our recommendations cover three different price tiers: Low-End (less than $200), Mid-Range ($200 - $500), and Flagship (more than $500). We also include a separate Phablet category for phones that make the best use of their larger screens.

To gain a better understanding of how we evaluate display, audio, camera, system, gaming, and battery performance, please read How We Test Smartphones And Tablets.

About Our Recommendations

• We only recommend smartphones we’ve actually tested.
• We only review phones for sale in the North American market.
• Our recommendations are based on data that we’ve collected and our own subjective experiences. We recognize that our readers are diverse, however, with different needs, preferences, and opinions, so our best picks may not be the best for you.
• The list is based on full retail U.S. prices from online retailers. We do not list carrier subsidized prices, because it’s not fair to phones that are only offered unlocked at full price (Also, why should we have to sign a two-year contract just to buy a computer?).
• The prices are for new phones only, not used or open-box.
• Prices and availability change on a daily basis, but the embedded green links provide real-time pricing.

Best Smartphones

Best Low-End Smartphone

Best Mid-Range Smartphone

Best Flagship Smartphone

Alternate Flagship Pick

Best Phablet

MORE: How We Test Smartphones
MORE: Apple iPhone 6 & 6 Plus Review
MORE: Google Nexus 6 Review


MORE: All Smartphone Content

Matt Humrick is a Staff Editor at Tom's Hardware, covering Smartphones and Tablets. Follow him on Twitter.

Follow us on Facebook, Google+, RSS, Twitter and YouTube.

Introduction

Huawei is not a household name in North America, but if the Chinese company's performance in other markets is any indication, this will soon change. Best known for selling telecommunications equipment, Huawei is quickly making a name for itself in smartphones, too. According to IDC, it is now the third largest smartphone vendor in the world by volume. Its success is centered in China, where its smartphone shipments increased 81 percent year-over-year in Q3 2015, but Huawei is expanding into emerging markets, such as Latin America and Africa, too. The company is also experiencing solid growth in the established European market, where smartphone shipments nearly doubled over the same period, according to Huawei.

Huawei recently pierced the North American market after being selected by Google as the manufacturing partner for the Nexus 6P. Looking to capitalize on this significant design win, Huawei is shipping another flagship phablet, the Mate 8, across the Pacific. We recently spent some time with this new device and were generally impressed (see Huawei Mate 8 Hands On). Its aluminum construction and minimal bezels give it a premium look and help hide the size of the large 6-inch screen. It comes with most of the features we've come to expect from a premium phone, including a fingerprint sensor, 3 to 4GB of RAM, 802.11ac Wi-Fi, NFC, and a decent camera. And while the battery is not removable, it's at least a generous 4000mAh.

In addition to being the first Huawei branded phone available globally, the Mate 8 has a couple of other firsts hidden inside its aluminum chassis, starting with HiSilicon's Kirin 950 system on a chip (SoC). Like MediaTek and Marvell, HiSilicon, a Huawei subsidiary, is an ARM processor licensee, which allows it to build SoCs using stock ARM cores. This approach differs from architectural licensees such as Apple and Qualcomm, who design custom CPU cores compatible with ARM's architecture.

Inside the Kirin 950 is an octa-core CPU with four ARM Cortex-A72 cores at up to 2.3GHz and four Cortex-A53 cores at up to 1.8GHz in a big.LITTLE arrangement. This makes HiSilicon only the second vendor behind MediaTek to ship a product using A72 cores, and the first to implement them on a current process node running close to its target frequency (the quad-core MT8173 is made using TSMC's older 28nm HPm node).

ARM's Mali-T880 GPU also makes its debut appearance in the Kirin 950, offering higher performance and up to a 40 percent improvement in energy efficiency over the Mali-T760 GPU. HiSilicon's implementation differs from other recent SoCs by opting for fewer, higher frequency cores. Specifically, it uses four GPU cores (MP4) that ramp up to a rather high 900MHz max frequency. In contrast, Samsung's Exynos 7420 uses the previous generation Mali-T760 GPU in an eight core configuration clocked at 772MHz. It will be interesting to see if the Mali-T880's architectural enhancements, combined with a higher clock frequency, can make up for the core count deficit.

Unlike previous Huawei devices that relied on external ISPs provided by outside vendors, the Kirin 950 benefits from Huawei's increasing vertical integration by incorporating its brand new, custom 14-bit PrimISP and IVP32 DSP into the SoC. We do not know much about these new blocks other than they enable up to 1080p60 H.264 encode/decode and 4K H.265 decode. Oddly missing is a 4K encode option.

Internally, the Kirin 950 uses ARM's CCI-400 Cache Coherent Interconnect. It's a bit odd that the newer CCI-500 interconnect is not used considering it was announced alongside the A72. This could potentially limit bandwidth between the CPUs and memory, since CCI-400 lacks a snoop filter on the interconnect itself.

Speaking of system memory, the Kirin 950 uses a hybrid LPDDR3/LPDDR4 memory controller. The Mate 8 takes advantage of LPDDR4's higher bandwidth and lower operating power, but other OEMs have the option of using less expensive LPDDR3 RAM. This feature, along with the smaller die from using only four GPU cores, should make the Kirin 950 a good fit for mid-range phones, an intriguing possibility considering the performance it should deliver.

With all of these new components, it's only fitting that the Kirin 950 is manufactured on a current generation process node, specifically TSMC's 16nm FinFET+. We've already seen FinFET's power and performance gains in our testing of Apple's A9 and Samsung's Exynos 7420 SoCs. More than a mere die shrink, FinFET is a new technology that fundamentally differs from the planar processes that came before it. It significantly reduces leakage current and power density, increasing battery life and reducing the amount of thermal throttling.

Performance Testing

This performance preview is a bit different than some of our others, because we're comparing not just the Kirin 950 SoC -- and its A72 CPU and Mali-T880 GPU cores -- to its peers, but also the Mate 8 to some other flagship phones.

We'll be comparing the Mate 8 and Kirin 950 to a range of different phones and SoCs, including the dual-core Apple A9, quad-core Intel Atom Z3580, and big.LITTLE octa-core designs using a combination of A57 and A53 cores. We're also including Qualcomm's latest Snapdragon 820 with its new, custom Kryo CPU. These results were collected from Qualcomm's smartphone MDP, which is its own internal development hardware, so they are still preliminary. Our comparison devices also contain a variety of Mali, PowerVR, and Adreno GPUs. With this range of hardware, spanning from mid-range to high-end, we should get a good feel for where the Mate 8, Kirin 950, and A72 fall.

All of these devices, other than the Qualcomm smartphone MDP, were evaluated using our standard testing procedures. Do note, however, that the Mate 8's software build is not final yet, so there could be some differences once it starts shipping.

MORE: Best Smartphones
MORE: Smartphones in the News
MORE: All Smartphone Content

CPU And System Performance

Now that we have a better understanding of the hardware, it's time to see how it performs. In this section, we'll evaluate system-level performance by running a series of synthetic and real-world workloads, along with some browser-based Web tests. If you're interested in learning more about how these benchmarks work, what versions we use, or our testing methodology, please read about how we test mobile device system performance.

The Huawei Mate 8 scores well overall in Basemark OS II, held back only by its GPU performance. In the OpenGL ES 2.0-based Graphics test, the Galaxy S6's Mali-T760MP8 easily outperforms the Mate 8's Mali-T880MP4 GPU by 54 percent, which is a larger margin than we would expect. This test does perform alpha blending and various texture operations, so it's possible having only half as many ROPs limits throughput compared to the Galaxy S6.

The Mate 8 performs much better in the CPU-centric System and Web tests. Only Apple's iPhone 6s Plus, buoyed by its Twister CPU's higher instructions per cycle (IPC), performs better. The octa-core Kirin 950 does finish ahead of Qualcomm's Snapdragon 820 by a small 7 percent margin, but keep in mind the 820's quad-core CPU is at a disadvantage in the multi-threaded portions of the test, and the 820's Kryo CPUs also run at a lower peak frequency. The fact that Snapdragon 820 finishes so close to the Kirin 950 despite these limitations seems to suggest that the Kryo CPU core has higher IPC than the A72, although we cannot pinpoint by how much from this test.

Compared to the Exynos 7420 in the Galaxy S6, the Mate 8's Kirin 950 scores about 11 percent higher in both the System and Web tests, a lower than expected result that basically mirrors the average difference in CPU clock frequency. Perhaps the A72's architectural improvements will have a bigger impact in our other tests.

While the Memory test purports to measure the speed of the internal NAND storage, it turns into more of a memory test on high-end devices due to how the OS uses a RAM cache to buffer storage access. Because this test does not work as intended, we cannot draw any definitive conclusions here.

The Mate 8 achieves the best overall score we've seen in AndEBench, narrowly defeating the Galaxy S6 and outpacing the Moto X Pure Edition by 30 percent. Its advantage comes primarily from the CoreMark-HPC CPU performance test, where its Kirin 950 outperforms the Exynos 7420 by 29 percent and Snapdragon 820 by 34 percent in a mixture of single- and multi-threaded integer and floating-point workloads. The newer revision of the Snapdragon 810 falls to the bottom of the chart, below the Snapdragon 808 that has two fewer A57 cores, because of thermal throttling. This exemplifies the A57's problems at 20nm: poor sustained performance. Unable to keep its four A57's, let alone both the A53 and A57 islands, running concurrently, the OnePlus 2 shuts down the A57's and relies only on the lower-power, lower-performing A53 cores. The combination of the relatively power-hungry A57 core and the 20nm planar process node results in a chip with a high power density and poor thermal performance. Being able to use FinFET and the power-optimized A72 CPU helps the Kirin 950 avoid this problem.

In both the streaming memory bandwidth and the memory latency tests, the Mate 8 performs similar to the Galaxy S6, not unexpected since they both use LPDDR4 RAM. The latest Snapdragon SoCs use a memory controller optimized for serial access patterns, which gives the 820 a big advantage in the memory bandwidth test (the 808 uses lower-bandwidth LPDDR3 RAM and the 810 cannot fully utilize the bandwidth of LPDDR4). This works against them, however, in the memory latency test that measures the time to complete a series of random memory operations. The memory controllers in the Kirin 950 and Exynos 7420 are more equal opportunity, favoring neither serial nor random patterns.

The Platform test mimics real-world workloads, testing CPU, memory, and storage performance. This test seems to use more random memory access patterns, which is why we see the BLU Pure XL and Asus ZenFone 2 perform quite well. The Mate 8 performs better than both of these devices and 17 percent better than the Galaxy S6. Storage performance for the Mate 8 is typical for an eMMC solution, trailing the S6's UFS 2.0 NAND, so its advantage in this test comes primarily from the A72 CPU.

Turning to the synthetic Geekbench suite, we get a clearer picture of the IPC differences between the A72, A57, and Qualcomm's Kryo CPU cores. Looking at the single-core overall scores first, the A72 in the Mate 8 performs 16 percent better in the integer tests than the A57 in the Galaxy S6, which turns into a mild 6 percent increase after accounting for the A72's 9.5 percent clock speed advantage. The A72 increases its lead to 25 percent (15 percent normalized) in floating-point workloads.

The small improvements over the A57 in these tests are not enough to match Kryo's performance, which outperforms the A72 by 12 percent (20 percent) in the integer tests and 32 percent (41 percent) in the floating-point tests, where the numbers in parentheses are the normalized differences after accounting for the A72's 7 percent clock speed advantage.

Geekbench 3 Pro Integer Results *

* Values in parentheses are the percent advantage for Kirin 950

Taking a closer look at the individual Geekbench integer workloads reveals several instances where the A72 sees no performance gain relative to the A57 (the values in the table are not adjusted for the 9.5 percent clock speed difference), including BZip2 Decompress, Sobel, and the various JPEG operations. This is not a complete surprise, because the integer execution units are similar between the A72 and A57, with a new Radix-16 integer divider (doubling bandwidth over A57) and a 1-cycle CRC unit the primary improvements. Both the Sobel and JPEG workloads rely heavily on multiplication and addition, so they see no improvement when running on the A72. The impact of the A72's improved branch predictor is also minimized when running benchmarks like Geekbench that primarily run math operations in a tight loop.

There are a few workloads, however, where the A72's architectural improvements make a noticeable difference: Lua, Dijkstra, AES, and SHA1. These tests rely heavily on lookup tables and data structures and seem to benefit from the A72's higher cache bandwidth and expanded zero-cycle forwarding.

Qualcomm's Kryo CPU clearly holds an IPC advantage over the A72 in integer operations. In our Snapdragon 820 Performance Preview, we determined that Kryo has a single integer multiply/divide unit just like the A57 and A72; however, we estimate it has only a 3-cycle latency compared to the longer 4-cycle latency of the ARM cores.

While single-core IPC is still the best indicator of smartphone application performance, multi-core throughput is becoming increasingly important, especially on Android. Comparing the multi-core integer performance of Kirin 950 to Exynos 7420 reveals an interesting pattern: the multi-core results are the exact opposite of the single-core results. Namely, tests that see a small gain when running on a single core see a larger gain when running on multiple cores and vice versa. Since the tests that see the biggest single-core gains take advantage of the A72's improved cache bandwidth, it's possible that the Kirin 950's reliance on the CCI-400 interconnect, instead of the newer CCI-500 that ARM recommends for the A72, is limiting the A72's performance. We'll need to see an example of a big.LITTLE processor using CCI-500 to know for sure, though.

Geekbench 3 Pro Floating-Point Results *

* Values in parentheses are the percent advantage for Kirin 950

In our detailed discussion of the A72 architecture we highlighted the A72's new Floating-Point/Advanced SIMD units, whose shorter pipeline lengths reduce execution latencies by up to 40 percent. There's also a new Radix-16 FP divider, which does more work per cycle, doubling bandwidth. These changes give the A72 a significant boost in floating-point performance over the A57 as shown in the table above. Other than two outliers (DGEMM and SFFT), the A72 performs anywhere from 4 percent (SGEMM) to 57 percent (Mandelbrot) better than the A57 after accounting for the difference in clock speed. 

For its Kryo CPU, Qualcomm made floating-point performance a priority, increasing IPC with a wider architecture that looks similar to Apple's Typhoon core in last year's A8 SoC. Even with latency minimized, the A72's narrower architecture cannot keep pace. It manages to almost pull even in Mandelbrot, gets less than half of Kryo's throughput executing Blur Filter, and finishes 41 percent behind overall after removing the A72's frequency advantage.

ARM's big.LITTLE philosophy of using multiple, narrower cores allows the Kirin 950 to pull ahead of the Snapdragon 820 in most of the multi-core tests. 

Unlike Geekbench, PCMark tests overall device performance by stressing the CPU, memory, and storage systems. Its lighter and more varied workloads are also affected by the device's CPU governor, just like other common apps, so it's indicative of real-world performance. 

The Mate 8 and its Kirin 950 SoC post the highest PCMark score of any device we've tested, outperforming the Moto X Pure Edition by 12 percent and the Galaxy S6 by 24 percent. Looking at the individual test scores, there's no obvious weak points for the Mate 8, which scores the highest in the Writing, Web Browsing, and Video Playback tests. It's only in the Photo Editing test, where image processing is supposed to occur on the GPU using the android.media.effect API, that it falls behind the Asus ZenFone 2 and Snapdragon 820 MDP.

The Mate 8 also does well in our JavaScript benchmarks, averaging 27 percent better than the Galaxy S6 in these three tests. Its performance is also very similar to the Snapdragon 820.

MORE: Best Smartphones
MORE: Smartphones in the News
MORE: All Smartphone Content

GPU And Gaming Performance

This section explores GPU performance with several synthetic and real-world game engine tests. To learn more about how these benchmarks work, what versions we use, or our testing methodology, please read our article about how we test mobile device GPU performance.

HiSilicon's Kirin 950 uses an ARM Mali-T880MP4 GPU. While the architectural differences between the T880 and the model it replaces, the T760, are not completely known, we do know that the T880 includes three ALU units per core versus two ALUs per core for the T760.

The table above summarizes the relevant current and previous generation GPUs from ARM and Imagination Technologies. Differences in architecture and nomenclature make it difficult to directly compare GPU hardware, so it focuses more on end performance. Qualcomm's Adreno GPUs are not included, because the company does not disclose details about its architecture.

Mali's Midgard differs from other GPU architectures in several ways. For starters, Midgard's vector SIMDs (single instruction multiple data) rely exclusively on instruction level parallelism (ILP) to keep its ALUs full of instructions. Competing GPU architectures are not vector based, and use a combination of ILP and thread level parallelism (TLP). There's pros and cons for both methods, making it difficult to say which offers better performance, but it ultimately depends on the workload.

The table highlights another difference between the Mali Midgard, PowerVR Rogue, and Qualcomm's Adreno architectures: IPC. Midgard performs fewer operations per cycle than its peers, but ramps up to a higher max frequency. The Kirin 950 in the Mate 8, for example, runs at up to 900MHz, resulting in 108 GFLOPS of FP32 throughput. This trails the A8's 115 GFLOPS (6 percent), Exynos 7420's 124 GFLOPS (15 percent), and the A9's more than 173 GFLOPS (exact GPU frequency unknown).

Here we see the Mate 8's Kirin 950 performing very much like a mid-range device rather than a flagship, its quad-core GPU besting only the PowerVR G6200 in the MediaTek Helio X10. The Mate 8 actually does pretty well in the first graphics test that focuses on vertex operations (with minimal pixel processing), essentially tying with the Galaxy S6 and performing 12 percent better than the Moto X Pure Edition's Adreno 418 GPU. It's in the second graphics test, which focuses heavily on pixel operations, where it falls behind.

The Mate 8's Physics score is surprisingly low. Since this tests CPU and memory performance, the Mate 8 should score the same as or a little better than the Galaxy S6. Instead, the S6, with its A57 CPU cores, scores 31 percent better than the Mate 8's higher clocked A72 cores. Unfortunately, we had to return our review unit before we could investigate this further. Without a closer look (and a device running final software), we do not want to make too much of this result, but it warrants further investigation.

In Basemark X, which runs on the Unity 4.2.2 game engine and uses OpenGL ES 2.0, the Galaxy S6 is 39 percent faster overall than the Mate 8. Most of its advantage comes from Dunes, a test that uses a lot of triangles (more than any of our other benchmarks), where it's 73 percent faster. HiSilicon's decision to use a quad-core GPU is definitely holding it back in this test; however, like all benchmarks, Basemark X puts more sever load on the GPU than actual games. When testing the LG G4 and its Snapdragon 808 SoC, for example, we found it played a number of modern games just fine. Considering that the Mate 8 actually performs better than the LG G4 in both offscreen and onscreen tests, and typical games do not use anywhere near as many triangles as Basemark X, the Mate 8's lower triangle performance should not be a severe limitation.

The Mate 8's 1080p display also works in its favor. By avoiding the greater rendering overhead of QHD, its onscreen Hangar results are actually better than the Galaxy S6's.

The Mali-T880MP4 GPU in the Kirin 950 really starts to struggle in the high quality test. With twice as many cores (giving it twice as much triangle and texturing throughput), the Mali-T760MP8 in the Exynos 7420 extends its lead to 70 percent overall (65 percent Dunes and 75 percent Hangar) when running offscreen.

GFXBench Manhattan runs on an OpenGL ES 3.0 based game engine that uses deferred rendering for its lighting effects. The PowerVR GT7600 in Apple's A9 and Adreno 530 in Qualcomm's Snapdragon 820 flex their ALU muscles, powering through the test's pixel operations. The Exynos 7420 in the Galaxy S6 also outperforms the Mate 8 by 35 percent. The Kirin 950 continues to outshine the Snapdragon 808 devices, though.

In the onscreen test, the Mate 8 pulls ahead of the Galaxy S6, since it only needs to render about half as many pixels.

Despite using an older OpenGL ES 2.0 based engine, we see essentially the same results in T-Rex. The Galaxy S6 is only 27 percent faster in the offscreen test than the Mate 8, and the Kirin 950 still holds a slim lead over the Snapdragon 808.

The limited number of ROPs in its quad-core GPU really holds the Kirin 950 and Mate 8 back in the Alpha Blending test, allowing the Snapdragon 808 to jump ahead of it. There's something amiss with the Galaxy S6's Alpha Blending results: The offscreen values are lower than the onscreen values, which is the opposite of what we should see. This appears to be a driver issue specific to the Galaxy S6, so we'll just ignore its results in this test.

The table at the top of this page documents the comparatively low FP32 ALU throughput of the Mali-T880MP4 GPU in Kirin 950, so it's no surprise to see it land near the bottom of the ALU performance chart. The Galaxy S6 scores 30 percent better than the Mate 8. Even the more budget friendly ZenFone 2 boasts better performance here. Qualcomm has worked steadily over the past several generations to boost the ALU performance of its Adreno GPUs, which is evident in this test. In the ALU onscreen test, nearly all of the devices are capped at the 60fps vsync limit.

While the T880 benefits from one additional ALU per core versus the T760, the Mali Midgard architecture makes no similar provision for texture units; the T880 still only has one per core. With twice as many texture units, the Galaxy S6 is 59 percent faster than the Mate 8 in the Fill test.

The Driver Overhead test measures CPU performance and driver efficiency by making a large number of draw calls. It's curious then to see the Mate 8 fall to the bottom of the chart, considering its strong CPU performance. The Mate 8 we're testing is still running prerelease software, so perhaps this is a driver issue that's left to be fixed.

MORE: Best Smartphones
MORE: Smartphones in the News
MORE: All Smartphone Content

Battery Life And Thermal Throttling

Battery life may be the most important performance metric for a mobile device. After all, it does not matter how quickly a phone or tablet can load webpages or how many frames per second the GPU can crank through once the battery runs down and the device shuts off. To learn more about how we test this critical facet of mobile computing, please read our battery testing methodology article.

The PCMark system test estimates battery life using real-world workloads and reflects how long a phone can last while continuously working on common tasks. Kirin 950's improved power efficiency, combined with the Mate 8's big 4000mAh battery, helps it last 9 hours and 23 minutes, the longest of any device we've tested, edging out Samsung's Galaxy S6 edge+ at 523 minutes and Motorola's Moto G (3rd gen) at 507 minutes. This feat is made even more impressive by the Mate 8's chart-topping performance score. Our colleagues at AnandTech examined Kirin 950's power efficiency in greater detail, which helps explain the Mate 8's outstanding results in PCMark.

The Mate 8 also does well in the GFXBench 3.0 battery test, which focuses on the GPU and is an indicator of battery life during intense gaming, lasting 3 hours and 45 minutes. We've seen a few devices last longer, but only because thermal throttling forced them to lower GPU frequency.

The Mate 8 exhibits good performance stability while gaming, staying within 94 percent of its peak performance over the first 30 minutes of T-Rex. After this we see the Mate 8 throttle back a bit, briefly reducing performance to 85 percent, but overall it's able to maintain near peak performance.

Looking at these results adds new context to HiSilicon's decision to use a quad-core GPU in the Kirin 950. By focusing on sustained performance rather than peak performance, a somewhat risky decision from a marketing standpoint, the Kirin 950 actually provides similar, if not better, performance after a few minutes of gameplay as compared to some other high-end SoCs, such as the Exynos 7420 or Snapdragon 810, at least when paired with a 1080p display like in the Mate 8.

Final Thoughts

ARM's Cortex-A72 CPU is a natural progression of the A57. At a high level, the two processors look similar, but ARM has made a number of power and performance optimizations to every stage in the pipeline. Most integer workloads see no appreciable performance gain, but there's a few specific cases, encryption in particular, that benefit from zero-cycle forwarding and the new Radix-16 integer divider. The A72's lower-latency floating-point units make a larger impact, increasing single-core Geekbench performance by about 15 percent overall relative to the A57 at the same clock frequency, with most of the individual workloads showing 30 to 60 percent gains.

Despite the A72's improvements, it's still a narrower architecture than Apple's Twister CPU or Qualcomm's new Kryo core, which limits IPC. After adjusting for differences in clock frequency, Kryo holds a 20 percent advantage in Geekbench integer and a 41 percent advantage in Geekbench floating-point performance. The A72 can, however, reach higher frequencies (we should see A72 cores running at 2.5GHz), which helps mitigate, and in some cases overcome, Kryo's greater IPC.

HiSilicon is the first to deliver the A72 on TSMC's 16nm FinFET+ process, and the first to use ARM's latest high-end GPU -- the Mali-T880 -- in its Kirin 950 SoC. This combination gives it better system performance and power efficiency than A57-based SoCs, including Qualcomm's Snapdragon 810 and Samsung's Exynos 7420. Kirin 950 also looks like it will be competitive with Snapdragon 820, at least on non-GPU related tasks, with each SoC having an edge in certain workloads.

At first glance, HiSilicon's decision to use the Mali-T880 in a quad-core configuration looks puzzling. While the T880's one additional ALU and higher max frequency helps keep it within 15 to 30 percent of the Exynos 7420's octa-core T760 GPU in shader-heavy games, having only half as many ROPs, texture units, and triangle units hurts peak performance over a wide range of gaming workloads, allowing the Exynos 7420 to extend its lead to around 60 percent. In all of our gaming benchmarks, the Kirin 950 performed more like a mid-range SoC.

A closer look, however, reveals HiSilicon's logic. Sure, the Kirin 950 is not going to wow anyone with its peak performance, but our tests show that its gaming stability is excellent, able to sustain near max performance over long periods of time. When paired with the Mate 8's 1080p display, the Kirin 950 actually performs better than the Exynos 7420 in the Galaxy S6 after a short period of time, because high temperature forces the S6 to throttle back its GPU frequency. This means the Kirin 950 and Mate 8 should not have any issues playing real-world games. We do not think the Kirin 950's Mali-T880MP4 GPU is powerful enough to drive QHD displays, however, limiting its use in some high-end flagships. The smaller GPU core and hybrid LPDDR3/LPDDR4 memory controller does make the Kirin 950 a high-performing mid-range SoC option from a cost standpoint, though.

The Kirin 950 seems a good fit for Huawei's Mate 8, which set new performance and battery life records in PCMark, our best benchmark for predicting real-world behavior. These tests corroborate our own first-hand experience: the Mate 8's UI was very fluid, Web pages loaded and scrolled quickly, and it just felt very fast overall. There's still several things we have not looked at yet, such as the display and cameras, but the Mate 8 seems like the real deal when it comes to performance and battery life.

MORE: Best Smartphones
MORE: Smartphones in the News
MORE: All Smartphone Content

Update, 1/19/16, 6:50am PT: Clarified Huawei's growth statistics in first paragraph.

Matt Humrick is a Staff Editor at Tom's Hardware, covering Smartphones and Tablets. Follow him on Twitter.

Follow us on Facebook, Google+, RSS, Twitter and YouTube.

Huawei’s recent success in the North American market is an unlikely of a success story as can be. It's a brand with no consumer cachet here, blessed with a name that many still have trouble pronouncing (it’s “wah-way”). However, it is one of the biggest smartphone OEMs globally, hugely successful in its Chinese home market, and its presence in European markets has been growing. Its biggest achievement so far, though, was the vote of confidence Google gave Huawei by picking it to be the manufacturer of one of its two Nexus phones this year.

Back in November, Huawei announced the latest phone in its Mate series of phablets, the Mate 8. These phones are usually only sold in China, and a handful of other Asian and European markets, but flush with the reception that the 6P received, it looks like Huawei thinks that the U.S. market is ready for a Huawei-branded premium device. Today’s CES announcement marks the global launch of the Mate 8 (which Huawei is optimistically dubbing the “ultimate high-end smartphone”), including North America.

Along with a U.S. release, the other big story of the Mate 8 is its SoC. Huawei is like the other giant on the Android OEM phone block in that it also makes its own SoCs, and the Mate 8 is powered by the new 16nm FinFET+ HiSilicon Kirin 950 chip. This is one of the first SoCs that uses ARM’s powerful new Cortex-A72 microarchitecture, which is expected to be one of the best performing mobile CPUs of 2016.

Huawei was kind enough to let us get our hands on a Mate 8 before today’s announcement, and in the short time we’ve had it, we’ve been able to run our full suite of performance tests on it. Right now, we are giving you a first look at the Mate 8 and also a teaser as to its performance, which (spoiler alert) is impressive to say the least.

Specifications

The specification charts we were provided show two models of the Mate 8, and it’s not clear in which markets each model will be available. The model we’ve been playing with is the dual-SIM NXT-L29 with 3 GB of RAM and 32 GB of storage. This model is optionally available with 4 GB of RAM and 64 GB of storage, and it's the model on sale in China as far as we know.

Beautiful Premium All-Metal Design

Design-wise, the Mate 8 uses the unibody aluminum construction that Huawei has basically perfected. Its design language should be familiar to anyone who’s seen a Nexus 6P, complete with the circular fingerprint reader. The Mate 8’s screen is a lot bigger, though, at 6 inches, which is pushing the upper limits of what’s considered acceptable to North American consumers. However, the Mate 8 also has super-thin bezels and an incredible screen-to-body ratio of 83%. This makes the Mate 8 shorter that the 5.7-inch 6P. When using it, it never feels like you’re using a phone with such a big display.

In contrast to most current flagship smartphones, the Mate 8’s screen is only 1080p, although in the end we couldn’t see much difference between this display and that of all the QHD phones we’ve been using. The Mate 8’s 368ppi clearly demonstrates that once you get about 350ppi, you don’t really notice the pixels.

Huawei Continues Its Photographic Excellence

Before we get into a performance preview of the Mate 8 and Kirin, the other significant component of the Mate 8 is its camera. It uses Sony’s brand-new Exmor RS IMX298 sensor (and is the only phone that does, so far), a chip that is philosophically quite different than the Google-mandated 12MP IMX377 sensor used in the Nexus 6. The 377 favors a lower MP count with a big sensor to create large (1.55 μm) pixels than capture more light. The size of these pixels is large enough that Google dispensed with OIS on the 6P.

The IMX298 is more of a traditional smartphone camera sensor, smaller in size with smaller pixels, so OIS is a necessity, and found on the Mate 8, along with fast-phase detect autofocus. Although we haven’t been able to put the camera through its paces properly, our first impressions are positive. It's also important to note that the physical camera hardware is only half the story when it comes to final picture quality. The ISP (image sensor processor) also plays a big role, and Huawei apparently assembled a team of 300 engineers and spent $98 million to ensure that the ISP in the Kirin 950 was up to snuff. At first glance, it looks like that was money well spent.

How Does It Perform?

Very well, actually. The first result above is PCMark’s Work Performance that measures overall system performance by running a series of realistic workloads. In this test, the Mate 8 and Kirin 950 outperformed every device we’ve tested so far, including the pre-release Snapdragon 820 in the Qualcomm mobile development platform we tested.

When you look at the individual PCMark test scores, the Mate also does very well.

The Mate 8 has a huge 4,000 mAH battery, and Huawei claims that it “delivers industry-leading power efficiency offering over two days of normal usage.” We believe them. Looking at the results above, the Mate 8 clearly outlasts every phone we’ve tested.

There is one negative aspect we want to point out now: Its Mali-T880 MP4 GPU isn’t particularly powerful, especially for a premium flagship phone. This isn’t due to its architecture, because like the A-72 CPUs, it’s brand new. It is the result of Huawei deciding to go with only a 4-core GPU in the 950. That means that despite its incredible overall system performance, the Mate 8 isn’t exactly a mobile gaming powerhouse, and this does explain the choice to stick with a 1080p display.

It Comes With Marshmallow but…

Let’s get the good news out of the way first. The Mate 8 does run Android 6.0 Marshmallow. Unfortunately, it isn’t stock Android, like on the 6P. It uses a new version of Huawei’s standard E Emotion UI (EMUI), and it’s clear that this UI is more designed for its home market. Like many Chinese OEM OSes, it blends elements of iOS and Android, so there is no application drawer, all your program icons reside on your homescreen, and it uses its own notification system, instead of the standard Android one introduced with Lollipop.

This is the one aspect of the Mate 8 that I can’t get behind, and while EMUI is by no means an unattractive UI, and is packed with some useful and intelligent features that go beyond the functionality of standard Android, I still found myself installing a third-party launcher to make it usable for me. Because we were testing a unit running pre-release software, we have encountered some bugs and issues, but it would be unfair to ding Huawei for this. Beta software isn’t supposed to be perfect, so we can’t pass final judgment on the Mate 8’s software until we’ve played with the final version.

Color Us Impressed

Our initial thoughts about the Huawei Mate 8 are very positive. This phone shows that the Nexus 6P wasn’t a fluke and that Huawei is capable of producing a world-class smartphone without having to partner with Google. The Mate 8 is blisteringly fast thanks to its Kirin 950 SoC, and it has excellent battery life and a great camera. Our only reservations about it so far are to do with its EMUI, but seeing as we were testing out a Chinese market model running unfinished software, we’re sure that Huawei will address many of the issues we have with it. Don’t expect EMUI to come with an app drawer anytime soon.

Presently, we don't know the Mate 8’s pricing, nor when it will be coming to North America. Hopefully, we’ll know more soon on that front.

Update, 1/05/16, 12:15 pm PT: At today’s CES press conference Huawei announced that the Mate 8 will be available “globally” and will be €599 for the 3 GB/32 GB model and €699 for the 4 GB/64 GB model. That works out to approximately $650 and $750, respectively.

Unfortunately, it looks like for now it isn’t officially coming to North America, and there it will certainly not be sold by any carriers. However, it is fully compatible with all North American cellular bands, and we were told that you can still buy it online to use it in the U.S. and Canada. We aren’t sure if this means if it will be sold directly by Huawei or if you’ll have to go through an online retailer.

Huawei recently announced a new SoC called the Kirin 930. The chip is built on TSMC's 16 nm FinFET process node, which puts it ahead of many other chip companies, including Qualcomm, in terms of adopting a more cutting edge process.

Being on a next-generation process node means that the chip can be made either more powerful, more efficient, or a combination of both. A chip company can even use the process advantage to hide the lower architectural performance of its CPU if it can raise the clock speed enough to reach the same performance as a competitor's chip, while still maintaining a certain power consumption level.

Apparently, Huawei chose to take advantage of the new process node by raising the speed of Cortex-A53 cores to 2 GHz and then calling them Cortex-A53 enhanced, or Cortex-A53e (likely just a brand name for a high clock speed Cortex-A53).

The Kirin 930 SoC comes with eight CPU cores, of which four are Cortex-A53 and clocked at 1.2 GHz, and the other four are the Cortex-A53e cores. The latter are supposed to replace the higher-performance Cortex-A57 cores we have seen in octa-core chips so far.

According to Huawei's benchmarks, a Cortex-A57 core is "only" 56 percent faster than a Cortex-A53 core, but at the same time, the Cortex-A57 core uses 3.5 times more power per GHz than the Cortex-A53 core (535mW/GHz vs 150mW/GHz).

In Huawei's opinion the trade-off is not worth it, and the company would rather use Cortex-A53 cores that have the same clock speed but use 3.5 times less energy, even if a Cortex-A57 core is 56 percent faster.

With the move to lower-power Cortex-A53 cores for high-end performance, Huawei manages to shoot two birds with one stone, because Cortex-A53 cores also cost less to make, which allows the company to beat competitors on price and offer better value for the money.

Huawei also seems to value making very slim phones, and using smaller and low-power chips that don't heat up too fast seems like a good idea for this purpose.

One of the first devices to use the new Kirin 930 SoC will be the Huawei P8 smartphone, for which the company is planning on holding an event on April 15, 2015 in London.

Follow us @tomshardware, on Facebook and on Google+.

Huawei has announced another smartphone in the form of the quad-core Ascend G615 handset.

The Ascend G615 houses a 4.5-inch IPS display with a 720p resolution and is powered by a quad-core 1.4GHz Hisilicon K3V2 processor with 1 GB of RAM and 8 GB of internal storage, which is expandable through a microSD slot.

As well as running on Android 4.0 Ice Cream Sandwich (it'll be upgraded to 4.1 Jelly Bean soon after launch), the device also sports an 8-megapixel camera, accompanied by a 1.3-megapixel front-facing snapper. Huawei promised 11 hours of talk time with a 2,150mAh battery.

The smartphone, which will retail for €299 ($397), will initially launch in Germany next month, followed by a global release soon after.

During CES, meanwhile, Huawei announced the world's most powerful smartphone, the Ascend D2 (read our hands-on impressions here), and a device with the world's largest screen found on a phone in the form of the 6.1-inch Ascend Mate -- head on over to our CES 2013 preview of the device. The Chinese smartphone manufacturer is also said to be working on the quad-core Ascend P2.

The company itself has been considerably vocal about its investment in the U.S. market. As well as blaming the death of Apple co-founder Steve Jobs on the supposed lack of innovation in the technology industry, the firm claims that it offers the best products rather than Samsung or Apple.

_{Contact Us for News Tips, Corrections and Feedback}