Huawei unveiled the Atlas 960E SuperPoD at HUAWEI CONNECT 2026 in Shanghai on September 17. This computing cluster houses 4,096 Ascend Neural Processing Units delivering 8 EFLOPS for FP8 and 16 EFLOPS for FP4 computations. The system reduces power consumption by 550 kilowatts per pod while delivering 2.3 to 4 times better performance than its predecessor. With 99.8% system availability, this architecture directly targets Nvidia’s dominance in AI infrastructure for models approaching 10 trillion parameters.
Source: Read the original article
