Nvidia streets ahead of China in AI chip race – Asia Times

Nvidia streets ahead of China in AI chip race - Asia Times

On March 18, CEO Jensen Huang’s keynote presentation at Nvidia’s GPU Technology Conference ( GTC 2024 ) provided an outstanding presentation of the opposition Huawei and another Chinese high- tech firms face alongside the punitive issue of US- led restrictions.

In a two- afternoon excursion to the frontiers of artificial intelligence, Huang introduced the agency’s fresh Blackwell AI computer, which he said” will be the most effective product launch in our story”.

Named for US mathematician David Blackwell, the Blackwell B200 graphics processing unit ( GPU) has 208 billion transistors compared with 80 billion for its Hopper H100 predecessor and is five times faster. The H100 has several times more computing power than the H20 that the US Commerce Department still allows Nvidia to sell in China.

The result of two rounds of enforced dumbing down, the H20 is being rejected by Alibaba, Baidu, Tencent and other Chinese cloud computing companies in favor of Huawei’s Ascend and other Chinese- designed AI processors, which are as good, or almost as good, without the risk of being subject to future sanctions.

But despite their impressive indigenous efforts, the Chinese are falling behind in the race to develop more powerful AI processors. Blackwell GPUs are scheduled to ship later this year to Amazon, Dell, Google, Meta, Microsoft, OpenAI, Tesla and other customers outside China.

They are made by TSMC using 4nm process technology based on ASML’s EUV lithography, which is not available in China. The same day as Huang’s presentation, TSMC and Synopsys announced they are going into production with Nvidia’s cuLitho computational lithography platform, which uses GPU- accelerated computing and generative AI to accelerate IC fabrication.

The world’s leading IC foundry and the world’s largest electronic design automation ( EDA ) company have integrated cuLitho into their software and manufacturing processes, resulting in “great leaps in performance, dramatic throughput improvement, shortened cycle time and reduced power requirements”, according to TSMC CEO C C Wei.

Nvidia says that 350 NVIDIA H100- based cuLitho systems can replace 40, 000 CPU- based systems, accelerating production while reducing costs, floor space and power consumption. Additional gains should be forthcoming in the not- too- distant future with the introduction of Blackwell architecture GPUs.

Nvidia CEO Jen- Hsun” Jensen” Huang has an AI vision. Image: Twitter

Sassine Ghazi, president and CEO of Synopsys, says” With the move to advanced nodes, computational lithography has dramatically increased in complexity and compute cost. Our collaboration with TSMC and NVIDIA is critical to enabling angstrom- level scaling … to reduce turnaround time by orders of magnitude through the power of accelerated computing”.

While SMIC, China’s leading foundry, embarrassed the US Commerce Department by using DUV lithography and multiple patterning to fabricate 7nm chips and develop 5nm capability, TSMC and Intel are advancing toward process nodes measured in angstroms, first equivalent to 2nm ( 20A in Intel terminology ) and 1.8nm ( 18A ) but aiming for 1nm around the end of the decade and sub- 10A after that.

Nvidia calls Blackwell” the engine of the new industrial revolution”. Talking about the next generation of industrial software at the GTC event, Siemens president and CEO Roland Busch said that the introduction of Nvidia- accelerated computing and generative AI will lead to the release of new cloud- based product lifecycle management software later this year that” will provide engineering teams with the ability to create an ultra- intuitive, photorealistic, real- time and physics- based digital twin that eliminates workflow waste and errors”.

Busch demonstrated this with a presentation of a workflow visualization developed for South Korean shipbuilder HD Hyundai’s ammonia- and hydrogen- powered ships, which have more than seven million parts. Workflow optimization that previously took days can be done in hours, according to Siemens and Nvidia,” with engineering data contextualized as it would appear in the real world”. They call this the “industrial metaverse”.

Nvidia also announced that several electric vehicle ( EV ) manufacturers and fleet operators have adopted DRIVE Thor, a new in- vehicle assisted and autonomous driving platform that is expected to go into mass production beginning next year. The successor to the widely- used DRIVE Orin, the DRIVE Thor centralized auto computer will be based on the Blackwell architecture, supporting large language models and generative AI.

BYD – China’s and now the world’s largest electric vehicle maker – will not only adopt DRIVE Thor but also use Nvidia’s AI architecture for factory planning. While BYD sells more EVs than Tesla many observers wonder if can it match Tesla’s autonomous driving technology. With DRIVE Thor, it probably can.

Hyper, a premium brand owned by China’s GAC AION, plans to use DRIVE Thor in EVs with Level 4 highly automated driving capability that is scheduled to go into production in 2025. Hyper currently uses DRIVE Orin in Level 2 partially automated vehicles.

Chinese EV makers XPENG, Li Auto and ZEEKR have also adopted DRIVE Thor, as has Nuro, a US developer of Level 4 autonomous driving technology founded by engineers who previously worked at Google, Plus, a Silicon Valley provider of Level 2 automated driving software for commercial vehicles that will use DRIVE Thor for Level 4 autonomous driving, Waabi, a Canadian company that is developing autonomous driving software for long- haul trucks, and WeRide, the Chinese robotaxi company founded in Silicon Valley that now operates in California, China, Singapore and the Middle East.

More than a dozen Chinese EV makers use the older DRIVE Orin processors, including BYD, Li auto, Nio, Xpeng and ZEEKR. In the first half of 2023, Nvidia had 53 % of China’s autonomous driving chip market, according to Gaogang Industry Research Institute. Nvidia was followed by Horizon Robotics with 31 %, Texas Instruments with 9 % and Huawei and Mobile Eye with 4 % each.

China’s BYD is leaving rivals behind in both China’s and global EV markets. Image: BYD

The risk for Nvidia and TI is that, as in the case of Apple, anti- American sentiment and hedging against the possibility of future sanctions will drive Chinese EV makers to shift to Chinese vendors.

US President Joe Biden, Commerce Secretary Gina Raimondo and vocal members of Congress have all called Chinese EVs a threat to US national security. In response, the Chinese government has asked Chinese EV makers to use domestic autonomous driving chips.

For the time being, they seem to prefer Nvidia but if the US government decides to interfere in this market next as part of its tech war with China, it might create an opportunity for Horizon Robotics and Huawei.

Follow this writer on&nbsp, X: @ScottFo83517667