September 25, at the Apsara Computing Conference in Hangzhou, Zhang Jianfeng, the head of Alibaba DAMO Academy, showcases the world’s most advanced AI chip, the Hanguang 800. In the industry-standard Resnet-50 Test, Hanguang 800’s inference performance is 78,563 IPS, 4 times higher than the world’s current best AI chip. The chip’s peak efficiency is 500 IPS/W, 3.3 times that of the runner-up.
In a presentation at the conference, 4 Hanguang 800 chips are used with City Brain, Alibaba’s AI-driven traffic-management system to process videos of Hangzhou’s urban traffic in real time, with a latency of 150ms. The same task would require 40 traditional GPUs with a longer latency of 300ms. Pailitao, Alibaba’s visual search engine needs to process one trillion new images every day. With Hanguang 800, what used to take an hour with traditional GPUs can now be done in around five minutes.