Chipmakers are tightlipped on China activities following a U.S. crackdown on hardware exports to the country. But Nvidia remains unfazed, and is doubling down on China being an important country for its computing hardware and software.
The company will ship some of its latest hardware, including a custom Hopper GPU, to customers in the country. The company also announced it developed software tools for Mandarin Chinese speakers.
“The most important thing is whatever the export restrictions are, whatever the laws and regulations in any country are, we comply to it in word and in spirit,” said Jensen Huang, CEO of Nvidia, during a press conference.
“That is probably as simple as that and … we know that regulation. We know the specification. It is very, very clear, and it’s not ambiguous,” Huang said.
China does compute that uses a lot of power, and Nvidia wants to help the country achieve carbon neutrality with its hardware products.
“China has some extraordinary tech companies, and they’re tech companies that are building products and services that have great social good,” Huang said.
Nvidia is shipping a Hopper-based GPU, called the H800, to China, according to a report by Reuters. The H800 is a modified version of the H100, with an interconnect speed of 300Gbps, which is roughly half the speed of the 600Gbps on the H100 version for PCIe slots.
Nvidia also said it will ship BlueField-3 data processing unit, which is a key product to boost the speed of computing infrastructures. It is emerging as an important piece of hardware for AI computing in high-performance computing environments. For example, the DPU is used to offload AI tasks to accelerators like its Hopper GPUs.
“BlueField-3 DPUs are available in China, subject to the same restrictions and controls that apply to previous generation Nvidia DPUs,” a spokeswoman from the company said in an email.
The DPU, which is now in full production, has a 400Gbps network connectivity, and 16 Arm-based Cortex-A78 cores. It is being used by Chinese companies that include Baidu, JD.com and Tencent. Bluefield-3 is an advanced chip with 22 billion transistors and was made on the 7nm process.
“There are also now licensing requirements to export a wide array of products, including networking products, destined for certain end users and for certain end uses in China,” Nvidia said in its annual report last month.
The export controls established by the U.S. are designed to choke China’s access to the latest semiconductor and artificial intelligence technologies. The export controls also block access to interconnects that enable supercomputing, but the DPU does not fall under that list.
Nvidia can’t export A100 or H100 GPUs – which are used in AI and supercomputing – to China. But Nvidia came up with a workaround, creating the A800 GPU for the Chinese markets that meets the U.S. export control requirements. The A800 was derived from the Ampere-based A100 GPU.
Intel in a recent press conference a few months back declined to comment on whether it would ship its Sapphire Rapids server chips to China. An Intel spokeswoman didn’t say anything beyond that the company was working with customers in the China market.
Nvidia is also recompiling some of its latest software stacks to work on older GPUs. A new computer manufacturing software stack, called cuLitho, could be used by Chinese customers on older GPUs to potentially design advanced chips. The U.S. is also limiting exports of advanced chip manufacturing tools from companies like ASML to China.
Nvidia’s CuLitho library is a set of new algorithms that speeds up underlying calculations to etch out finer features on pieces of silicon. The algorithms are supported by TSMC, toolmaker ASML and EDA leader Synopsys
“As transistors continued to become smaller, the blurring became worse. The calculations became more complicated and it requires more computers,” said Vivek Singh, vice president at Nvidia’s advanced technology group, during a briefing with a press ahead of GTC.
“Five hundred Hopper systems can now do the work of 40,000 CPU systems, while using only 1/8 the space and only 1/9 the power. A chip design that used to take two weeks to process can now be processed overnight,” Singh said.
CuLitho has been recompiled to work on GPUs as old as Volta, meaning customers in China could use the algorithms. When asked whether the toolkit would be usable in China, Singh said “It will work on architectures from Volta onwards… it shines the most on Hopper, so if somebody were to start anew that’d be the one to get.”
Nvidia also touted a big breakthrough in its speech-to-face AI interface, called Audio2Face, which can create an animated talking head to sync up lip movement. The voice can be fed to the AI, which compiles the talking head. The new audio-to-face feature includes new capabilities for Chinese Mandarin, which is spoken in China and Taiwan.
The company showed a video of an animation speaking Chinese Mandarin based on the audio input, with the lip moving in sync with the audio. View demo on YouTube.
“[This is] really a milestone in the advancements of computer graphics,” said Richard Kerris, vice president of Omniverse platform at Nvidia, during a call with the press.