Huawei Shows Off 384-Chip AI Computing System That Rivals Nvidia's Top Product

Long-time Slashdot reader hackingbear writes: China's Huawei Technologies showed off an AI computing system on Saturday that can rival Nvidia's most advanced offering, even though the company faces U.S. export restrictions. The CloudMatrix 384 system made its first public debut at the World Artificial Intelligence Conference (WAIC), a three-day event in Shanghai where companies showcase their latest AI innovations, drawing a large crowd to the company's booth. The CloudMatrix 384 incorporates 384 of Huawei's latest 910C chips, optically connected through an all-to-all topology, and outperforms Nvidia's GB200 NVL72 on some metrics, which uses 72 B200 chips, according to SemiAnalysis. A full CloudMatrix system can now deliver 300 PFLOPs of dense BF16 compute, almost double that of the GB200 NVL72. With more than 3.6x aggregate memory capacity and 2.1x more memory bandwidth, Huawei and China 'now have AI system capabilities that can beat Nvidia's,' according to a report by SemiAnalysis.

The trade-off is that it takes 4.1x the power of a GB200 NVL72, with 2.5x worse power per FLOP, 1.9x worse power per TB/s memory bandwidth, and 1.2x worse power per TB HBM memory capacity, but SemiAnalysis noted that China has no power constraints only chip constraints. Nvidia had announced DGX H100 NVL256 'Ranger' Platform [with 256 GPUs], SemiAnalysis writes, but 'decided to not bring it to production due to it being prohibitively expensive, power hungry, and unreliable due to all the optical transceivers required and the two tiers of network.
The CloudMatrix Pod requires an incredible 6,912 400G LPO transceivers for networking, the vast majority of which are for the scaleup network.'

Also at this event, Chinese e-commerce giant Alibaba released a new flagship open-source reasoning model Qwen3-235B-A22B-Thinking-2507 which has 'already topped key industry benchmarks, outperforming powerful proprietary systems from rivals like Google and OpenAI,' according to industry reports. On the AIME25 benchmark, a test designed to evaluate sophisticated, multi-step problem-solving skills, Qwen3-Thinking-2507 achieved a remarkable score of 92.3. This places it ahead of some of the most powerful proprietary models, notably surpassing Google's Gemini-2.5 Pro, while Qwen3-Thinking secured a top score of 74.1 at LiveCodeBench, comfortably ahead of both Gemini-2.5 Pro and OpenAI's o4-mini, demonstrating its practical utility for developers and engineering teams.

Read more of this story at Slashdot.