Huawei, the Chinese tech powerhouse, has made a significant advancement in the global AI chip race with the introduction of its latest computing system, the CloudMatrix 384 Supernode. Reports indicate that this new system outperforms similar offerings from American leader Nvidia, which could reshape the competitive landscape for AI technology as development continues worldwide amid ongoing US efforts to limit China’s access to advanced tech. The CloudMatrix 384 Supernode has been described as a “nuclear-level product,” achieving a remarkable computing power of 300 petaflops, surpassing Nvidia’s NVL72 system that provides 180 petaflops.
It was designed to alleviate computing bottlenecks that arise as AI models grow increasingly complex. First unveiled in September 2024, the CloudMatrix infrastructure caters specifically to the burgeoning demand within China’s domestic market. This new variant not only demonstrates impressive performance but is also built using Chinese-made components.
It reportedly matches the capabilities of Nvidia’s H100 chips, achieving a throughput of 1,920 tokens per second while maintaining high accuracy levels. Despite facing stringent sanctions that limit its access to advanced semiconductor technology and design tools, Huawei has managed to achieve this technological breakthrough. The new system utilizes an innovative alternative to Nvidia’s NVLink, allowing multiple GPUs to communicate efficiently.
This development is part of a broader context in which Chinese tech companies are making concerted efforts to create a robust domestic AI computing infrastructure. For instance, Alibaba recently announced a colossal investment in AI resources, signaling a larger shift toward technological self-sufficiency in China. Ultimately, the emergence of Huawei’s CloudMatrix 384 Supernode highlights a crucial moment in the global AI landscape, potentially leading to increased competition, better options for developers, and a significant step towards technological independence for China despite existing sanctions.