BEIJING, Dec 9 (Reuters) – U.S. President Donald Trump said the United States would allow Nvidia’s H200 processors to be exported to China, though it remains unclear whether Beijing will give the green light to Chinese companies to purchase them.
China now has a number of AI chipmakers, as the country looks to wean itself off reliance on foreign technology. Here is how Chinese rivals compare with Nvidia’s H200.
DO CHINESE CHIPS MATCH THE H200’S POWER?
No. The most advanced chip from Chinese suppliers is Huawei’s Ascend 910C, which lags significantly behind the H200 in computing power and memory bandwidth, according to a recent report from the Institute for Progress.
The 910C delivers total processing performance (TPP) of 12,032, compared with the H200’s 15,840, and has memory bandwidth of 3.2 terabytes per second versus the H200’s 4.8 TB/s.
Other domestic players offer less competitive products. Cambricon’s top chip, the Siyuan 590, and Hygon’s BW1000 both trail Huawei’s 910C in performance.
ARE THERE ANY CHINESE CHIPS THAT CAN MATCH NVIDIA’s H20?
Chinese chips can match Nvidia’s H20, a downgraded processor Nvidia designed for the Chinese market but stopped shipping earlier this year after Washington banned it.
Huawei’s 910B delivers total processing performance of 5,120, surpassing the H20’s 2,368, according to a July Bernstein report. Cambricon’s Siyuan 590, with TPP of 4,493, also outperforms the H20.
WHY IS IT HARD FOR CHINESE FIRMS TO DISPLACE NVIDIA?
Nvidia’s dominance stems from its CUDA software platform, which developers have long relied on to build AI models.
Switching to domestic chips would require developers to rewrite code and retrain on new platforms, a costly and time-consuming process.
Despite domestic chips outperforming the H20 in raw computing power, Chinese internet companies still prefer Nvidia’s offering due to the lack of a mature domestic alternative to CUDA.
WHAT IS IN HUAWEI’S PIPELINE?
Huawei unveiled its AI chip roadmap in September, announcing three new products over the next three years.
The Ascend 950PR will launch in the first quarter of 2026, followed by a higher-memory variant, the 950DT, in the fourth quarter of 2026. The Ascend 960 is slated for the fourth quarter of 2027 and the Ascend 970 for the fourth quarter of 2028.
The 960’s computing power roughly matches the H200, according to a September Bernstein report. However, it features significantly higher interconnect bandwidth at 2,200 gigabytes per second compared with the H200’s 900 GB/s.
Higher interconnect bandwidth enables faster communication between chips in multi-chip systems, which is critical for training large AI models. This suggests Huawei is prioritizing networking speed over raw computing power.
HOW DOES NVIDIA’S H200 COMPARE WITH ITS LATEST CHIPS?
The H200 is an older generation of Nvidia’s AI chips, built on its Hopper architecture launched in 2022, which preceded the latest Blackwell architecture revealed last year.
Nvidia’s latest AI server, which packs 72 Blackwell chips into a single computer, improves the performance of some AI models by 10 times, compared with H200 servers, according to data published by the company earlier this month.
According to a report released on Sunday by the non-partisan think tank, the Institute for Progress (IFP), the Blackwell chip now in use by U.S. AI firms is about 1.5 times faster than H200 chips for training AI systems and five times faster for inferencing work, where AI models are put to use.
However, the H200 would be almost six times as powerful as the H20, the most advanced AI semiconductor that can legally be exported to China, according to IFP’s report.
(Reporting by Liam Mo, Che Pan and Brenda Goh; Additional reporting by Arsheeya Bajwa; Editing by Anil D’Silva)
Comments