China is not significantly behind the U.S. when developing AI hardware and software, said Jensen Huang, chief executive of Nvidia, at the Hill and Valley Forum. While Chinese companies have made remarkable progress with AI services and software, Huang’s comments mainly refer to Chinese AI hardware. Coincidentally, Huawei has started to ship its latest AI system, which is largely deemed competitive with Nvidia’s GB200 NVL72, reports Financial Times. However, there is a catch.
“China is not behind anybody, China is right behind us, we are very, very close,” said Huang at the sidelines of the Hill and Valley Forum, where business leaders and lawmakers met to discuss technology and national security, according to Bloomberg Podcasts.
Regarding hardware, China has numerous companies developing AI chips that are more or less competitive against Nvidia’s AI hardware. Huawei is ahead of everyone else as its AI strategy spans from its Ascend 900-series AI accelerators to servers to rack-scale solutions for cloud data centers. Recently, the company announced CloudMatrix 384, which packs 384 dual-chiplet HiSilicon Ascend 910C interconnected using a fully optical mesh network. So far, Huawei has sold over 10 CloudMatrix 384 systems to Chinese customers, which signals that some would like to try domestic alternatives to restricted Nvidia hardware.
“There is no question that Huawei is one of the most formidable technology companies in the world, and they are incredible in computing, they are incredible in networking technology, and in software capabilities, all of the essential capabilities to advance AI. And they have they’ve made enormous progress in the last several years,” said Huang.
This Huawei CloudMatrix 384 system spans 16 racks: 12 compute racks each housing 32 Ascend 910C accelerators, plus four dedicated racks managing optical connectivity via 6,912 optical transceivers at 800G LPO speeds. CloudMatrix 384 achieves roughly 300 PFLOPs dense BF16 compute, nearly double Nvidia’s GB200 NVL72, with 2.1 times higher memory bandwidth and over 3.6 times greater HBM capacity using HBM2E. CM384’s optical interconnects provide substantial bandwidth advantages, including 2.1 times better scale-up and 5.3 times better scale-out capabilities. However, it falls short of Nvidia’s GB200 NVL72 efficiency, consuming 2.3 times more power per FLOP, 1.8 times more per TB/s memory bandwidth, and 1.1 times more per TB of HBM memory.
Shipping 10 CM384 systems is hardly a big deal per se, especially considering that Chinese technology firms like Alibaba, ByteDance, and Tencent reportedly invested $16 billion in Nvidia’s H20 HGX GPUs in the first quarter, anticipating U.S. export restrictions called the AI Diffusion Rule set to begin in mid-May.
Nonetheless, 10 systems may be the start. Huawei reportedly has about a million HiSilicon Ascend 910C, enough for 2,600 CloudMatrix 384 systems. Then again, 2,600 CM384 clusters can deliver about 780 ExaFLOPS of dense BF16 compute. One needs around 4,300 of Nvidia’s GB200 NVL72 machines (or around 312,000 B200 AI GPUs) to get similar performance. Analyst Ming-Chi Kuo estimated that shipments of GB200 NVL72 racks alone in 2025 could reach roughly 25,000 – 35,000 racks, but in addition, the company will also ship GB300 NVL72 racks as well as B100, B200, and B300 AI GPUs.
This means that Nvidia alone will supply orders of magnitude more AI compute performance than Huawei this year. However, Huawei has technologies that may rival those of the American company.
Follow Tom’s Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.
Leave a Reply