It’s been less than a month since the Biden administration effectively barred the export of most American-designed AI accelerators to China, yet Nvidia has already found a way to weave around those rules and get high-ish-end silicon out to the Middle Kingdom.
Multiple reports indicate the California chipmaker has three new GPUs in the works which slip under, or otherwise avoid, the performance limits on AI accelerators sold in China. According to Chinastarmarket, which cites supply chain sources, the GPUs include the H20, L20, and L2, and could be announced as early as November 16 – after the US export restrictions go into full force.
While that report doesn’t offer any clues to the processors’ performance, the industry watchers at SemiAnalysis said they managed to get their hands on the speeds and feeds for the chips. H20 will allegedly offer 96GB of memory, 4Tb/s of memory bandwidth, and 296 teraFLOPS of FP8 performance. That’s less than a sixth that of Nvidia’s flagship H100, though Jensen’s crew thinks they should sell.
Alongside the H20 SXM, is the 48GB L20 and 24GB L2, which will apparently deliver 239 teraFLOPS and 193 teraFLOPS of FP8 performance, respectively.
While those chips would crank out fewer FLOPS than Nvidia’s latest crop of datacenter GPUs, one of the export-friendly components will reportedly be 20 percent faster than the H100 when it comes to inferencing — the process of putting an AI model to work — on large language models. These performance targets appear to have been engineered to sidestep the Biden administration’s export controls.
As a refresher, the sale of GPUs and accelerators exceeding 600GB/s of interconnect bandwidth in China has been restricted since last fall. The rules introduced last month have added an extra set of caps on performance density, which we explored in detail here.
Nvidia was among the hardest hit by these rules, so much so that the Silicon Valley giant swiftly warned investors it was unlikely it would be able to sell a wide swathe of its datacenter GPU portfolio in China in the near future. But thanks to ravenous demand for GPUs everywhere else, it predicted the issue probably wouldn’t hurt their financials.
Among the GPUs affected were its L40, L40S and China-market A800 and H800. We’ll note that the investor note also mentioned A100 and H100, however these have been restricted in China since last fall.
The restrictions were so strict that even Nvidia’s flagship gaming GPU, the RTX 4090, will require a special license to continue selling them in the Middle Kingdom. And while Nv initially believed that it would have 30 days to get Chinese orders for these parts out the door, less than a week after they were announced the US government stepped in and ordered the corporation to cease shipments immediately.
Nvidia is hardly the only US chip house developing sanctions-compliant chips for the Chinese market. Earlier this year Intel announced a nerfed version of its Habana Gaudi2 accelerator for the region, while AMD has said it is also working on new accelerators that comply with Chinese export restrictions.
Nv declined to comment for this story. ®
- SEO Powered Content & PR Distribution. Get Amplified Today.
- PlatoData.Network Vertical Generative Ai. Empower Yourself. Access Here.
- PlatoAiStream. Web3 Intelligence. Knowledge Amplified. Access Here.
- PlatoESG. Carbon, CleanTech, Energy, Environment, Solar, Waste Management. Access Here.
- PlatoHealth. Biotech and Clinical Trials Intelligence. Access Here.
- Source: https://go.theregister.com/feed/www.theregister.com/2023/11/09/nvidia_china_gpu/
