Ascend Cluster Leading in China's Semiconductor Field
TapTechNews June 19th news, according to a report by Titanium Media on June 18th, at a sub-forum of the 2024 World Semiconductor Conference and Nanjing International Semiconductor Expo held recently, Wang Tao, the chief operating officer of Jiangsu Kunpeng · Ascend Ecosystem Innovation Center, said that Ascend cluster is the only technical route in China that has completed training a large model with hundreds of billions of parameters and its performance has exceeded that of Nvidia A100.
Wang Tao also disclosed that among China's large models, nearly 50% have chosen the Ascend technology route. The Ascend AI chip can reach up to 1.1 times the training efficiency of Nvidia. In training models such as MetaLlama and BloomGPT, the computing power training efficiency of Ascend AI is'much better than' that of Nvidia A100 and '10 times ahead of' other domestic competitors.
Wang Tao said that currently Ascend 'has become the first choice for domestic large models'. After testing, iFlytek has also adopted Huawei's ten-thousand-card cluster, which can achieve 0.8 to 1.2 times the efficiency compared to Nvidia A100. 'There is indeed still a certain gap with Nvidia A100 (referring to 0.8 times), but there is no obvious gap with the Nvidia A100 chip in this regard of large model training. Especially in the ten-thousand-card computing power cluster, including Kunpeng Cloud Brain and iFlytek, this has all passed the market test.'
According to TapTechNews' previous report, during the 2023 Huawei Full-Connection Conference held last September, Wang Tao, the standing director of Huawei, director of the ICT Infrastructure Business Management Committee, and president of the Enterprise BG, officially released the brand-new architecture of the Ascend AI computing cluster - Atlas 900 SuperCluster, which can support the training of large models with ultra-trillions of parameters. The new cluster adopts the brand-new Huawei Xinghe AI computing switch CloudEngine XH16800. With its high-density 800GE port capacity, a two-layer switching network can realize a super-large-scale and non-converging cluster networking of 2250 nodes (equivalent to 18000 cards).
