Ascend AI Chip Surpasses Nvidia A100 in Large Model Training, Becoming the Preferred Choice in China, said by Want Tao from Huawei

Recently, at a sub-forum during the 2024 World Semiconductor Conference and Nanjing International Semiconductor Expo, Wang Tao, the Chief Operating Officer of the Jiangsu Kunpeng & Ascend Ecosystem Innovation Center, delivered a significant speech. He stated that the Ascend AI chip has surpassed the Nvidia A100 in terms of performance within the technical route for training large models with billions of parameters in China, becoming the preferred technical route for large model training domestically.

Wang revealed that nearly 50% of China’s large models have chosen the Ascend technical route. The training efficiency of the Ascend AI chip can reach up to 1.1 times that of Nvidia. In training models such as Meta Llama and BloomGPT, the computing power training efficiency of Ascend AI is significantly superior to that of Nvidia A100, leading by 10 times compared to other domestic competitors. This achievement has been validated by companies like iFlytek, which adopted Huawei’s ten-thousand-card cluster, achieving an efficiency of 0.8 to 1.2 times that of Nvidia A100.

Wang Tao also emphasized that Ascend has become the first choice for domestic large model development. After testing, compared to the Nvidia A100 chip, Ascend does not fall short in training efficiency within a ten-thousand-card computing power cluster, and even surpasses Nvidia A100 in certain aspects. This achievement is of great significance for the development of China’s domestic chips.

The outstanding performance of the Ascend AI chip in training domestic large models has made it the preferred technical route for such training in China. This achievement not only promotes the development of the domestic semiconductor industry but also provides strong technical support for research and application in the field of artificial intelligence.