At the 2025 World Artificial Intelligence Conference in Shanghai, Huawei unveiled its most ambitious AI computing platform to date: the CloudMatrix 384. Designed as a direct answer to NVIDIA’s top-tier NVL72 system, the CloudMatrix brings together 384 Ascend 910C processors using Huawei’s proprietary optical interconnect. It boasts around 300 petaflops of BF16 performance, outpacing NVIDIA’s cluster in raw throughput by nearly 70%.
While the performance numbers look impressive on paper, they come with substantial energy costs. CloudMatrix draws nearly four times more power than its NVIDIA rival, making it less efficient and costlier to run. This trade-off reflects Huawei’s strategy: scale out rather than match NVIDIA’s chip-level efficiency. Its system features more memory capacity and higher bandwidth, ideal for model training at enormous scales, but with significant overhead in cooling and infrastructure.
Despite the muscle, Huawei faces challenges in the software department. The Ascend ecosystem, still in its early maturity, can’t yet rival NVIDIA’s CUDA environment. That means deploying and maintaining AI workloads on CloudMatrix requires more specialized engineering—up to five times more, by some estimates. Yet China’s domestic workforce and lower energy costs give Huawei an edge in pushing forward regardless of these hurdles.
Ren Zhengfei, Huawei’s founder, admitted their chips lag behind NVIDIA’s, but emphasized that China’s strength lies in its system-building and manpower. CloudMatrix 384 is a clear manifestation of this vision—a bold step toward AI autonomy. It may not be the most elegant or efficient machine on the planet, but it’s a powerhouse engineered under constraint and determination, signaling a shift in the global AI compute race.
#Huawei #AI #CloudMatrix384 #NVIDIA #Ascend910C #WAIC2025 #ChinaTech #Supercomputing #AIInfrastructure #TechNews