Latest news with #DylanPatel
Asia Times
5 days ago
- Business
- Asia Times
Huawei chief hasn't a chip worry in the world
In a recent interview with China's state-run People's Daily, Huawei founder and CEO Ren Zhengfei provided an assessment of the Chinese semiconductor industry that many might find surprising. An English version of the interview was published by the Communist Party-run Global Times. 'There's actually no need to worry about the chip issue,' Huawei's chief said. 'By leveraging methods such as superposition and clustering, computational results can match the most advanced global standards. In terms of software, thousands upon thousands of open-source software will meet the needs of the entire society in the future.' This optimism comes from objective analysis supported by Huawei's own experience, with some self-deprecation: 'There are many companies in China making chips, and many are doing well; Huawei is just one of them. The US has exaggerated Huawei's achievements – the company isn't that powerful yet. We need to work hard to live up to their evaluation. 'Our single chips still lag behind the US by a generation. We use mathematics to compensate for physics, non-Moore's Law approaches to complement Moore's Law, and group computing to make up for single-chip limitations, which can also achieve practical results.' This squares with the conclusion of Dylan Patel and his colleague at SemiAnalysis, who found that Huawei's Ascend 910C AI processor is more impressive when used in the company's CloudMatrix 394 rack-scale AI data center solution, which is a complete system consisting of 384 Ascend 910C processors, servers, networking, storage, power management and cooling. In their estimation, the CloudMatrix 394 'competes directly' with Nvidia's top-end GB200 Grace Blackwell Superchip. 'The engineering advantage,' they write, 'is at the system level not just at the chip level, with innovation at the accelerator, networking, optics, and software layers… Huawei is a generation behind in chips, but its scale-up solution is arguably a generation ahead of Nvidia and AMD's current products on the market.' With regard to the ongoing effort to develop the basic semiconductor devices needed to support the country's consumer electronics, automotive and other industries, Ren said, 'China has opportunities in low- and mid-range chips, with dozens or even hundreds of chip companies working hard. The opportunities are even greater for compound semiconductors.' One prominent example is China's rapid advance in silicon carbide (SiC) power semiconductors, which have become standard in electric vehicles (EVs). Approximately two-thirds of the world's electric vehicles (EVs) are manufactured in China, making this both an obvious opportunity and a strategic necessity. Compared with ordinary silicon, SiC-based power devices are more energy-efficient and reliable. They improve the performance of not only electric vehicles and battery chargers, but also industrial machinery, solar and wind power and data centers. In March, BYD announced a new high-speed EV charging system, which enables 400 kilometers of driving in five minutes – about twice the performance of Tesla's supercharger. According to DigiTimes, 'Silicon carbide (SiC) semiconductors played an instrumental role in this technological advancement, as key advantages of the wide bandgap material, including high voltage and temperature resistance and low energy loss, help enhance the efficiency and reliability of electric drive systems to support high-voltage charging.' Nomad Semi wrote that, 'This achievement was made possible by BYD Semiconductor's breakthrough in high-power 1,500V SiC chips. It marks the first large-scale application of 1500V SiC chips in the global automotive industry.' BYD is also starting to make its own SiC wafers, which should give it a complete internal SiC supply chain from substrates to chips and modules. Established in 2002, BYD Semiconductor also makes other types of discrete power semiconductors, power management ICs, microcontroller units (MCUs), sensors and optoelectronic devices used in new energy vehicles (NEVs, which include both battery-powered and hybrid vehicles). BYD appears to be well on its way to self-sufficiency in automotive semiconductors. Ren also emphasized the importance of theoretical scientific research. 'We must understand and support those doing theoretical work,' he said. 'We need to appreciate their vision; their great, quiet dedication… those engaged in theoretical research are the hope for our country's future.' Huawei is doing its part: 'We invest 180 billion yuan (US$25 billion) in research and development each year, with approximately 60 billion yuan allocated to basic theoretical research, which is not subject to performance evaluation. About 120 billion yuan is invested in product research and development, which is subject to evaluation. Without theoretical support, there can be no breakthroughs, and we will not be able to catch up with the US.' For example, more than 20 years of research into hybrid stochastic number systems has led to the development of a Hybrid Stochastic Computing SoC (System-on-Chip) for high-performance computing at the School of Electronic and Information Engineering of the Beijing University of Aeronautic and Astronautics (BUAA). Led by Professor Li Hongge, the research and development team combined binary (0 – 1) and stochastic (probability-based) values, in-memory computing, and heterogenous SoC design (multiple specialized processing units) using open-source RISC-V architecture, which is beyond the reach of US government sanctions. As reported by the Guangming Daily, the hybrid chip features higher fault tolerance, stronger resistance to interference, and much greater energy efficiency than conventional binary digital chips. As translated by TrendForce, 'Professor Li explains that stochastic computing expresses values through the probability of a CMOS logic signal remaining 'high' during a given time period. In other words, the frequency of high-level pulses represents the numerical probability.' BUAA is already applying the technology to touch recognition, instrument display panels, and flight control. Beyond that, the research team is working on more complex functions such as voice and image processing and AI model acceleration. The chips themselves are fabricated by the Chinese IC foundry SMIC. Similar R&D programs are underway in the US, Japan and Europe, but for the time being, China leads the world in the practical application of hybrid stochastic computing. The negative implications for the US policy of technology containment should be obvious. 'For the US semiconductor industry, China is gone,' electronics industry analyst Handel Jones told The New York Times. Jones is the founder and CEO of California consulting firm International Business Strategies, Inc. 'He projects that Chinese companies will have a majority share of chips in every major category in China by 2030.' Follow this writer on X: @ScottFo83517667
CNBC
12-06-2025
- Business
- CNBC
China's racing to build its AI ecosystem as U.S. tech curbs bite. Here's how its supply chain stacks up
With the U.S. restricting China from buying advanced semiconductors used in artificial intelligence development, Beijing is placing hopes on domestic alternatives such as Huawei. The task has been made more challenging by the fact that U.S. curbs not only inhibit China's access to the world's most advanced chips, but also restrict availing technology vital for creating an AI chip ecosystem. Those constraints span the entire semiconductor value chain, ranging from design and manufacturing equipment used to produce AI chips to supporting elements such as memory chips. Beijing has mobilized tens of billions of dollars to try to fill those gaps, but while it has been able to "brute force" its way into some breakthroughs, it still has a long way to go, according to experts. "U.S. export controls on advanced Nvidia AI chips have incentivized China's industry to develop alternatives, while also making it more difficult for domestic firms to do so," said Paul Triolo, partner and senior vice president for China at advisory firm DGA-Albright Stonebridge Group. Here's how China stacks up against the rest of the world in four key segments needed to build AI chips. Nvidia is regarded as the world's leading AI chip company, but it's important to understand that it doesn't actually manufacture the physical chips that are used for AI training and computing. Rather, the company designs AI chips, or more precisely, graphics processing units. Orders of the company's patented GPU designs are then sent to chip foundries — manufacturers that specialize in the mass production of other companies' semiconductor products. While American competitors such as AMD and Broadcom offer varying alternatives, GPU designs from Nvidia are widely recognized as the industry standard. The demand for Nvidia chips is so strong that Chinese customers have continued to buy any of the company's chips they can get their hands on. But Nvidia is grappling with Washington's tightening restrictions. The company revealed in April that additional curbs had prevented it from selling its H20 processor to Chinese clients. Nvidia's H20 was a less sophisticated version of its H100 processor, designed specifically to skirt previous export controls. Nevertheless, experts say, it was still more advanced than anything available domestically. But China hopes to change that. In response to restrictions, more Chinese semiconductor players have been entering the AI processor arena. They've included a wide array of upstarts, such as Enflame Technology and Biren Technology, seeking to soak up billions of dollars in GPU demand left by Nvidia. But no Chinese firm appears closer to providing a true alternative to Nvidia than Huawei's chip design arm, HiSilicon. Huawei's most advanced GPU in mass production is its Ascend 910B. The next-generation Ascend 910C was reportedly expected to begin mass shipments as early as May, though no updates have emerged. Dylan Patel, founder, CEO and chief analyst at SemiAnalysis, told CNBC that while the Ascend chips remain behind Nvidia, they show that Huawei has been making significant progress. "Compared to Nvidia's export-restricted chips, the performance gap between Huawei and the H20 is less than a full generation. Huawei is not far behind the products Nvidia is permitted to sell into China," Patel said. He added that the 910B was two years behind Nvidia as of last year, while the Ascend 910C is only a year behind. But while that suggests China's GPU design capabilities have made great strides, design is just one aspect that stands in the way of creating a competitive AI chip ecosystem. To manufacture its GPUs, Nvidia relies on TSMC, the world's largest contract chip foundry, which produces most of the world's advanced chips. TSMC complies with U.S. chip controls and is also barred from taking any chip orders from companies on the U.S. trade blacklist. Huawei was placed on the list in 2019. That has led to Chinese chip designers like Huawei to enlist local chip foundries, the largest of which is SMIC. SMIC is far behind TSMC — it's officially known to be able to produce 7-nanometer chips, requiring less advance tech than TSMC's 3-nanometer production. Smaller nanometer sizes lead to greater chip processing power and efficiency. There are signs that SMIC has made progress. The company is suspected to have been behind a 5-nanometer 5G chip for Huawei's Mate 60 Pro, which had rocked confidence in U.S. chip controls in 2023. The company, however, has a long way to go before it can mass-produce advanced GPUs in a cost-efficient manner. According to independent chip and technology analyst Ray Wang, SMIC's known operation capacity is dwarfed by TSMC's. "Huawei is a very good chip design company, but they are still without good domestic chipmakers," Wang said, noting that Huawei is reportedly working on its own fabrication capabilities. But the lack of key manufacturing equipment stands in the way of both companies. SMIC's ability to fulfill Huawei's GPU requirements is limited by the familiar problem of export controls, but in this case, from the Netherlands. While Netherlands may not have any prominent semiconductor designers or manufacturers, it's home to ASML, the world's leading supplier of advanced chipmaking equipment — machines that use light or electron beams to transfer complex patterns onto silicon wafers, forming the basis of microchips. In accordance with U.S. export controls, the country has agreed to block the sale of ASML's most advanced ultraviolet (EUV) lithography machines. The tools are critical to making advanced GPUs at scale and cost-effectively. EUV is the most significant barrier for Chinese advanced chip production, according to Jeff Koch, an analyst at SemiAnalysis. "They have most of the other tooling available, but lithography is limiting their ability to scale towards 3nm and below process nodes," he told CNBC. SMIC has found methods to work around lithography restrictions using ASML's less advanced deep ultraviolet lithography systems, which have seen comparatively fewer restrictions. Through this "brute forcing," producing chips at 7 nm is doable, but the yields are not good, and the strategy is likely reaching its limit, Koch said, adding that "at current yields it appears SMIC cannot produce enough domestic accelerators to meet demand." SiCarrier Technologies, a Chinese company working on lithography technology, has reportedly been linked to Huawei. But imitating existing lithography tools could take years, if not decades, to achieve, Koch said. Instead, China is likely to pursue other technologies and different lithography techniques to push innovation rather than imitation, he added. While GPUs are often identified as the most critical components in AI computing, they're far from the only ones. In order to operate AI training and computing, GPUs must work alongside memory chips, which are able to store data within a broader "chipset." In AI applications, a specific type of memory known as HBM has become the industry standard. South Korea's SK Hynix has taken the industry lead in HBM. Other companies in the field include Samsung and U.S.-based Micron. "High bandwidth memory at this stage of AI progression has become essential for training and running AI models," said analyst Wang. As with the Netherlands, South Korea is cooperating with U.S.-led chip restrictions and began complying with fresh curbs on the sale of certain HBM memory chips to China in December. In response, Chinese memory chip maker ChangXin Memory Technologies, or CXMT, in partnership with chip-packaging and testing company Tongfu Microelectronics, is in the early stages of producing HBM, according to a report by Reuters. According to Wang, CXMT is expected to be three to four years behind global leaders in HBM development, though it faces major roadblocks, including export controls on chipmaking equipment. SemiAnalysis estimated in April that CXMT remained a year away from ramping any reasonable volume. Chinese foundry Wuhan Xinxin Semiconductor Manufacturing is reportedly building a factory to produce HBM wafers. A report from SCMP said that Huawei Technologies had partnered with the firm in producing HBM chips, although the companies did not confirm the partnership. Huawei has leaned on HBM stockpiles from suppliers like Samsung for use in their Ascend 910C AI processor, SemiAnalysis said in an April report, noting that while the chip was designed domestically, it still relies on foreign products obtained prior to or despite restrictions. "Whether it be HBM from Samsung, wafers from TSMC, or equipment from America, Netherlands, and Japan, there is a big reliance on foreign industry," SemiAnalysis said.
Asia Times
25-04-2025
- Business
- Asia Times
Trump pushing China toward AI processor self-sufficiency
China's stockpile of Nvidia H20 artificial intelligence (AI) processors is likely to run out in about a year now that sales to Chinese customers have been banned by the Trump administration, according to the assessment of several market analysts. Chinese tech giant Huawei thus must ramp up production of its new Ascend 910C alternative as quickly as possible while other Chinese AI chip designers step up efforts to avoid a chip shortfall in the years ahead. Alibaba, ByteDance, Tencent and other Chinese companies ordered between US$12 billion and $16 billion worth of H20 processors, or perhaps even more, in the first quarter of this year, according to various sources. At least one million of the chips were reportedly delivered before shipments were cut off. The new US sanction 'didn't come as a surprise because it was widely anticipated across the industry,' an unidentified Chinese corporate executive told Japan's Nikkei Asia newspaper. 'Every major Chinese tech company had been stockpiling H20 in advance. After all, it wasn't banned at the time, and given its strong performance, why not?' The H20 sanction marked the third time that the US has put a ceiling on the performance of AI processors that may be exported to China and then, after new dumbed-down versions turned out to be best-sellers, lowered the ceiling again. Huawei's Ascend series of AI processors had its origins in the company's Da Vinci design architecture, which was introduced in 2018 to serve as a platform for AI computing and replace Nvidia in data center, cloud, edge device and other applications. The Ascend 910 processor was launched the following year. The Ascend 910B was introduced in 2022 after the US had forced Taiwan's TSMC to cut off chip foundry services to Huawei. Fabricated in China by SMIC using non-sanctioned 7nm DUV lithography, its performance is assessed as close to that of the Nvidia A100, which was launched two years earlier, but about 20% below the H100. Exports of those two Nvidia chips to China were blocked in 2022. The 910B was reportedly adopted by China's Alibaba, Baidu, Tencent, speech recognition company iFLYTEK, AI software company SenseTime, local universities and national laboratories, and Huawei itself. Mass shipments of the Ascend 910C, the development of which was revealed last August, are expected to begin within the next several weeks. Already used by China's AI revelation DeepSeek, it is currently the most advanced Chinese alternative to Nvidia. The 910C consists of two 910B chips connected together in a single package, similar to Nvidia's leading edge Blackwell processor's structure. On a stand-alone basis, the 910C's performance is close to the H100 and exceeds the H20. But the 910C is more impressive when used in Huawei's new CloudMatrix 394 rack-scale AI data center solution – a complete system consisting of 384 Ascend 910C processors, servers, networking, storage, power management and cooling. In the estimation of analyst Dylan Patel and his colleagues at SemiAnalysis, the CloudMatrix 394 'competes directly' with Nvidia's top-end GB200 Grace Blackwell Superchip. 'The engineering advantage,' they write, 'is at the system level not just at the chip level, with innovation at the accelerator, networking, optics, and software layers… Huawei is a generation behind in chips, but its scale-up solution is arguably a generation ahead of Nvidia and AMD's current products on the market.' Huawei's solution uses considerably more electricity than Nvidia's, but 'the deficiencies in power are a relevant but not a limiting factor in China.' While this is the best that China can do now under current US sanctions, that probably won't always be the case. SemiAnalysis notes that Huawei still depends on imported high-bandwidth memory (HBM) from South Korea's Samsung and that its AI processors are fabricated using imported equipment. But those two vulnerabilities are also being addressed. Korean stockbroker Hyundai Motor Securities writes that Chinese DRAM maker CXMT is investing heavily in HBM, 'targeting deployment in Huawei's Ascend AI chips within two to three years.' Furthermore, China's semiconductor equipment and materials industry has advanced to the point where SMIC and other foundries, as well as CXMT and other Chinese memory chip makers, can upgrade their process technology and increase capacity in spite of US sanctions. China's leading maker of semiconductor production equipment Naura, for example, is now rated by some in the global Top 10. Ascend 910C yields are still low, according to industry sources, but improving at a rate that suggests SMIC's production capacity could reach 400,000 chips per month later this year. And Huawei has already announced its successor, the Ascend 920, fabricated using SMIC's 6nm process, will be 30% to 40% more efficient than the 910C. The 910C production estimate and Ascend 920 specs speak to the failure of US sanctions to block China's access to AI processors. As for Nvidia, if those sanctions are not dropped, it will likely see its once overwhelming share of the China market drop to zero. Follow this writer on X: @ScottFo83517667
