Originally designed with 384 Ascend 910C neural processing units and 192 Kunpeng server central processing units, which are interconnected through a unified bus providing ultra-high bandwidth and low latency, the system can support up to tens of thousands of interconnected chips to train and run AI models, according to Huawei executives at the company’s sixth Innovation and IP Forum in Beijing on Tuesday.
The top invention integrated innovations in hardware, software and architecture, according to Sun Hongwei, Huawei’s chief computing software architect.
Sun said his team tackled plenty of challenges to maintain reliability, power supply and data storage for large computing systems.
