“As the scale of parallel processing grows, cross-machine bandwidth in traditional server architectures has become a critical bottleneck for training,” said Zhang Dixuan, president of Huawei’s Ascend computing business, in a speech at the Shenzhen-based tech giant’s Kunpeng Ascend Developer Conference held last Friday.
Zhang pointed out that addressing future training requirements will demand innovative computing architectures like Supernode 384.
Privately held Huawei’s supernode architecture forms the basis of its previously announced CloudMatrix 384 system, which is a cluster of 384 Ascend AI processors – spread across 12 computing cabinets and four bus cabinets – that delivers 300 petaflops of computing power and 48 terabytes of high-bandwidth memory. A petaflop is 1,000 trillion calculations per second.
Huawei has deployed CloudMatrix 384, which it has touted as the industry’s largest AI training platform, at its data centres in eastern Anhui province, Inner Mongolia and southwestern Guizhou province.
