China is the world’s second largest economy, it has the world’s largest population, and it is only a matter of time before has a world-class technology ecosystem spanning the smallest transistors to the largest hyperscale and HPC systems. It simply has enough money, and enough time, and enough smart people, to get there.
But that does not mean that the United States wants to make that transition for the Middle Kingdom an easy one. And so, every now and then, export controls and licenses are put into effect by the US Department of Commerce to restrict the sale of hardware, software, and other vital pieces of technology to China. And even before it was waging war against Ukraine, sales of technology to Russia were also restricted.
Nvidia and AMD were just served such notices restricting sales of their top-end GPU engines by the US government last week, and we are certain that Intel will come under the same restrictions for its future GPUs.
And that is only going to spur on Chinese tech companies to create their own CPUs, GPUs, and AI accelerators. Much as restrictions on the sale of the “Knights Landing” many-core CPUs to certain Chinese HPC centers hastened the development of the Matrix 2000 DSP-based math accelerator for the Tianhe-2A supercomputer, delivered in 2015 far faster than anyone expected, we think the restrictions on the sale of high-end GPU accelerators is going to raise the profile of Biren Technology’s BR100 GPU as well as potentially those of other chip designers we have written about, such a Jingjia with its JM9 series and Tianshu Zhixin with its “Big Island” GPU. (The latter code-name, we think, might be referring to the island of Taiwan. . . . It is probably not referring to Hawaii in the United States, the big island of the Hawaiian volcanic chain, or Honshu in Japan, which is the largest island in the world and the bulk of that nation’s land mass. It would be ominous if it referred to all three, right?)
The lesson from this development of chips and restriction of sales is that you cannot stop progress when there is a will and a government with money.
The Department of Commerce was not explicit about the threat to national security it saw from Nvidia and AMD selling their GPUs into China. Nvidia filed an 8-K statement with the US Securities and Exchange Commission to notify shareholders that the US government’s license requirement “will address the risk that the covered products may be used in, or diverted to, a ‘military end use’ or ‘military end user’ in China and Russia.” Nvidia told Wall Street that it did not sell products into Russia, but it did warn Wall Street that it had reckoned it would sell $400 million of technology restricted by the license in its third quarter of fiscal ending in October. We don’t know if that $400 million is for the weeks not already completed in the quarter up until August 26 or for the full thirteen weeks of the fiscal third quarter. So we don’t know what impact the restrictions will have on sales at Nvidia’s datacenter division. By out math and the guidance that Nvidia gave last month, we were expecting datacenter sales to be around $3.9 billion, give or take.
Nvidia was served notice on August 26 that sales of the current A100 and A100X GPUs (the latter has the GPU embedded on a ConnectX-5 network interface card) and the future H100 GPU as well as any DGX systems that use these components have “a new license requirement,” which is Govspeak for can’t be sold into China or Russia without permission from Uncle Sam. This license requirement is meant to put the peak performance and chip-to-chip I/O performance of the A100 accelerator as a barrier, so anything faster, such as the H100 and we also presume (though it was not mentioned) the H100 CNX (which is a converged GPU-NIC card like the A100X) is also restricted although it is not explicitly forbidden in the statement put out by Nvidia.
The license is not just for sales of products, but also development of products in either Russia or China, and interestingly, Nvidia said in its 8-K filing that the development of the new H100 and the support of existing A100 customers might be interrupted by the geographical provisions of the license requirement set by the Department of Commerce. On September 1, it filed an amendment to the 8-K filing saying that the US government authorized exports, re-exports, and in-country transfers to finish the development of the H100 GPUs, and allows Nvidia to export parts needed to fulfil the A100 needs of US customers through March 1 of next year. The Commerce Department is also allowing A100 and H100 order fulfillment and logistics at Nvidia’s facility in Hong Kong through September 1, 2023.
As for AMD, the company did not file an 8-K form with the SEC because its sales in Russia and China for its high-end Instinct MI250 GPUs are non-existent. But an AMD spokesperson had this to say to The Next Platform: “AMD has received new licensing requirements from the US Department of Commerce that prevent us from shipping MI250 integrated circuits to China and Russia. At this time, we do not believe that shipments of MI100 integrated circuits are impacted by the new requirements. We do not currently believe it is a material impact on our business.”
We reached out for comment from Intel, but everyone we know in the server and GPU groups is on sabbatical this week. But what seems clear to us, based on the expected 52 teraflops FP64 and FP32 performance of the impending “Ponte Vecchio” GPU and the much better performance of the “Rialto Bridge” GPU, Commerce will be restricting sales of these devices in China and Russia, too.
What we can’t understand is why the Commerce Department is not restricting sales of even Nvidia V100 GPUs and AMD Instinct MI100 GPUs. It is not like China cares if it is using old technology or not in a pinch – it built the “OceanLight” exascale-class supercomputer with a homegrown mathed-up CPU using the 14 nanometer processes that is the most advanced decently yielding process at state-controlled chip foundry Semiconductor Manufacturing International Corporation (SMIC) .
What we don’t understand is why the restrictions stop at the top-end GPU engines. Why not all GPUs? While it is difficult, both China and Russia could build working exascale supercomputers using Nvidia P100 and V100 GPUs or even AMD Instinct MI50 or MI100 GPUs. The resulting machines might burn a tremendous amount of energy, but when you are designing a fighter jet or a nuclear bomb, this might not matter as much as having a machine with such capacities. Heck, you could do it with fleets of Tesla K80s bought off of eBay if you really wanted to do it.
The other thing we wonder about is if the license requirements affect the distribution and sale of Nvidia’s CUDA environment, AMD’s ROCm environment, and Intel’s oneAPI environment. If the GPUs are dangerous, so is the software.
Whatever the national security interest, it seems clear that the United States wants to slow China down. But there is no way to stop China from inventing analogous technologies for anything America or the European Union can create. The only edge any country has is to keep fostering more innovation.
Sign up to our Newsletter
Featuring highlights, analysis, and stories from the week directly from us to your inbox with nothing in between.
” And that is only going to spur on Chinese tech companies”
So are the restrictions counterproductive?
How about software?
It is known – I guess I read it in TNP – that software is a big hurdle for Chinese HPC. What is going to happen now is that Chinese will have to turn to Jingjia, Biren et al. Presumably there is even less optimized software developed for them?
Biren has a stack of its own, although I doubt it has the library breadth and depth that CUDA has. They have a start, they have money, and they have smart people. So it is really just a matter of time.
Someone please explain how these high end GPUs don’t end up in China via multiple levels of distribution. . . Okay, probably they can’t get the most performant systems like a DGX, but bottom line this restriciton only equates to delayed revenue for Nvidia and AMD. Even the software is attainable through partners. Too much technology flows through SE Asia so that this is much of an issue, China can basically have anything they want with time and patience.