Huawei's Ascend SuperNode enables massive, fast AI model training

Huawei has launched Ascend SuperNode technology, achieving the industry's largest 384-card high-speed bus interconnection, significantly boosting AI model performance. This new technology, unveiled at the Ascend AI Developer Summit, replaces traditional Ethernet with high-speed bus connections, increasing communication bandwidth by 15 times and reducing latency tenfold. This innovation addresses bottlenecks in large-scale AI model training. The Ascend 384 SuperNode, composed of 12 computing cabinets and 4 bus cabinets, allows for faster training of complex models like LLaMA 3 and Qwen, and can scale to support even larger AI clusters.