Huawei touts low-latency AI architecture

Huawei is promoting its Supernode 384 architecture as a next-gen AI compute platform to rival Nvidia’s NVL72. Built on 384 Ascend chips, the system reportedly delivers up to 750 TPS per card on complex models and cuts latency tenfold using high-speed bus links. Already deployed in multiple data centers, it targets scalable support for dense and MoE AI workloads.