Alibaba's Zhenwu M890 SuperNode features 64 AI accelerators with 800G interconnect, supporting FP8/FP4, and is used for ADAS and embodied AI training, achieving a 3x gain over the Zhenwu-810E. Each M890 SuperNode can perform inference for a 10T MoE model and is installed in Ulanqab, with Moonshot advising Kimi K3 runs on at least 64 accelerators.
Alibaba's SuperNode specifications reveal the scale of compute infrastructure required for next-gen AI models, driving demand for high-density GPU clusters and advanced networking.