Alibaba Cloud has introduced the Aegaeon system, which it says reduces NVIDIA H20 GPU usage by up to 82%. In a three-month test, Aegaeon required only 213 H20 GPUs compared to the previous 1,192 GPUs. to process an artificial intelligence (AI) model with 72 billion parameters.
This GPU efficiency is achieved by combining GPU resources so that one GPU can support multiple models simultaneously. Previously, one GPU could only be used to process one model. This technique was developed by Alibaba in collaboration with researchers at Peking University.
Aegaeon was developed not only to improve the efficiency of AI data centers but also to solve the issue of purchasing NVIDIA GPUs imposed on China by the United States. Therefore, for more than a month, the Chinese government has been advising local companies not to buy NVIDIA chips and to switch to locally made AI chips such as those produced by Huawei.
