China has officially inaugurated its first Ascend 950DT 1024-card intelligent computing cluster in Wuhan. The cluster boasts liquid-cooled super-nodes, each armed with 1024 Ascend NPU accelerator cards and designed with a fully liquid-cooled system. These super-nodes achieve an impressive FP8 computing capacity of up to 1 EFLOPS, house a massive 256TB of globally unified memory, provide TB-level bandwidth for inter-card interconnection, and ensure microsecond-level end-to-end communication latency. Utilizing the Lingqu architecture, the cards within the cluster can seamlessly access one another, providing robust support for the training of large models with trillion-level parameters and intelligent agent inference.
