Editor’s take: The Trump administration tried to curb China’s capability to advance its computing capabilities by tariffs and different commerce restrictions. Nevertheless, this unprecedented commerce battle could have as a substitute pushed Beijing to speed up its efforts in creating highly effective new AI techniques.
As newly appointed US tech czar David Sacks predicted only a month in the past, Trump’s tariffs look like backfiring in spectacular style. Chinese language tech large Huawei is reportedly creating a strong AI system that may already compete with Nvidia’s most superior infrastructure within the race for the world’s quickest AI platforms.
Huawei publicly unveiled its new system, the CloudMatrix 384, on the latest World Synthetic Intelligence Convention in Shanghai. The three-day occasion was full of firms showcasing their newest AI improvements, and in response to Reuters, Huawei’s sales space was among the many most crowded and talked-about on the present.
Many attendees have been desirous to study extra in regards to the CloudMatrix 384, however Huawei officers declined to offer additional particulars. Whereas the system could have been on show primarily to generate buzz, sufficient data has emerged to present us a clearer sense of the place Huawei – and China’s broader AI business – may be heading.
In response to a latest research by SemiAnalysis, Huawei’s CloudMatrix 384 is the corporate’s reply to Nvidia’s GB200 NVL72 – a flagship AI supersystem designed to run trillion-parameter fashions in actual time. Nvidia’s resolution contains 36 Grace CPUs and 72 Blackwell GPUs, working collectively in a unified rack to operate as a single large GPU that dramatically accelerates giant language mannequin inference.
Whereas Huawei lags behind Nvidia and different Western companies in superior silicon design, its engineers selected to compensate with scale and innovation moderately than uncooked pace. The CloudMatrix 384 incorporates 384 Ascend 910C chips, interconnected in an all-to-all topology to maximise efficiency. The Ascend 910C, designed by Huawei’s personal fabless semiconductor arm HiSilicon, successfully combines two Ascend 910B processors to ship efficiency on par with Nvidia’s H100 GPU.
A totally configured CloudMatrix 384 can attain 300 petaFLOPs of dense BF16 compute, almost double the compute capability of Nvidia’s GB200 NVL72. It additionally boasts 3.6× higher complete reminiscence capability and a couple of.1× extra reminiscence bandwidth.
SemiAnalysis famous in April that with additional yield enhancements, Huawei could quickly outpace Nvidia – and by extension, america – within the AI race. The foremost caveat? Energy consumption. The CloudMatrix 384 requires greater than 4 instances the power wanted to run a GB200 NVL72 at full capability.
Nevertheless, not like many Western nations, China has been aggressively increasing its energy technology infrastructure utilizing coal, photo voltaic, hydro, wind, and different sources.
