Huawei Cloudmatrix 384 is the new Rack scale cluster for artificial intelligence systems That the Chinese giant has presented at the Waic conference held at the World Exhibition Center for Shanghai. And it is extremely interesting because it confirms that Nvidia has competition in AI accelerators.
Huawei has become the most advanced Chinese hardware supplier in the development of chips for IA. Despite the export restrictions of technology imposed by the United States, Huawei has advanced very quickly And CEO himself of Nvidia, Jensen Huang, acknowledged his competition at the same time that Trump requested a relaxation to those rules not to be out of the gigantic Chinese market. This same month, the US raised some chips export restrictions to China.
Huawei Cloudmatrix 384, at NVIDIA?
Developments such as Deepseek confirmed that China is taking significant steps to shorten the distance that separates them from the United States in artificial intelligence. And they will not stop … It is estimated that more than 40% of all the solutions that are developed today from China and there is no week that new projects are not announced, mostly open and free. All this has not gone unnoticed as we saw last week in Trump’s’ action plan, which has the unique objective of win the battle for AI.
All these models need high -level hardware and the Huawei Cloudmatrix 384 is intended to satisfy an insatiable resource. The rack includes 384 Ascend accelerators of Huawei, a direct competitor of the NVIDIA H20 chip. Baptized as Atlas 900 A3 Superpod, the cluster exceeds Nvidia development in some metricsaccording to Semiianalysis technicians.
The highest global cloudmatrix 384 (300 computing petaflops BF16, almost double that the GB200 NVL72) is due to the design capabilities of the Huawei system, which compensate for the lowest speed of individual chips by using a greater number of accelerators, system level innovations and a “supernod” architecture that allows the chips to be interconnected very quick speeds.
Huawei still has the development to match Nvidia’s capabilities, especially in watt yield and speed of individual chips, but is already showing what you can get in the future in competition with the Grace Blackwell of Nvidia. You will also have to improve costs. A single cluster of the Cloudmatrix 384 is 8 million dollars, almost three times higher than the NVL7200 NVIDIA configuration. Therefore, the main objective of the Huawei product is not to offer profitable performance, but a product Manufactured with own resources and capable of competing with Western alternatives.
China advances and Huawei is a direct part of the equation to provide the necessary hardware to Asian servers and perhaps in the future to those of other countries.
