The logo of Huawei. /VCG
Chinese tech giant Huawei announced its latest high-performance AI chip clusters at the annual Huawei Connect conference on Thursday in Shanghai, calling them the “world’s most powerful” for AI computing.
“Computing power is – and will continue to be – key to AI. This is especially true in China,” said Eric Xu, Huawei’s deputy chairman of the Board and the current rotating chairman.
The company is developing a “supernode + supercluster” computing solution, leveraging China’s available semiconductor manufacturing processes to sustainably meet long-term computing demand, Xu said.
Nodes are like a rack system that contains numerous chips, which are then grouped into clusters.
Huawei’s new supernode products, the Atlas 950 SuperPoD and Atlas 960 SuperPoD, can support 8,192 and 15,488 Ascend chips, respectively. They excel in key performance metrics, including computing power, memory capacity, and interconnect bandwidth, and are currently the “most powerful” of their kind worldwide, Xu said, adding that “they will remain so for years to come.”
These supernodes are successors to the Atlas 900. Xu said more than 300 of Atlas 900 supernodes has been deployed to over 20 customers across a variety of industries, including internet, telecommunications, and manufacturing.
The Atlas 950 will be launched in the fourth quarter of 2026 followed by the Atlas 960 in the fourth quarter of 2027, with a 970 version expected in 2028.
Built on the supernodes, the company also announced its superclusters – the Atlas 950 SuperCluster and the Atlas 960 SuperCluster, which will utilize 500,000 and 1 million chips, respectively.
Xu said the Atlas 950 supercluster would have 2.5 times the scale of Elon Musk’s xAI Colossus supercomputer and 1.3 times its computing power.
“These too are poised to outperform all other computing clusters on the market,” he said.
In addition, Xu outlined Huawei’s chip roadmap for the next three years, including the upcoming release of the Ascend 950PR, 950DT, 960 and 970 models.
No comments:
Post a Comment