Relive the best moments of the Apsara Conference 2019 at https://www.alibabacloud.com/apsara-conference-2019.
Jeff Zhang, Dean of Alibaba DAMO Academy, unveiled Alibaba’s first AI inference chip, Hanguang 800, during Apsara Conference on September 25, 2019. Hanguang 800 is the world’s most powerful AI inference chip. In the Resnet-50 industry test, the peak performance of the new chip reached a whopping 78,563 images per second, which is four times higher than the second best AI chip in the world. The peak efficiency of the chip also reached 500 IPS/W, which is 3.3 times higher than the second best option.
“Alibaba is a new player in the chip industry. The launches of XuanTie 910 and now Hanguang 800 are only the first steps towards a new chip revolution,” Jeff Zhang said in a statement.
Hanguang, roughly translated as Sword of Light, is named after a legendary sword from Chinese mythology, said to be invisible but provide the swordsman wielding it with incredible power. A Hanguang 800 chip can offer the computing power equivalent to 10 traditional GPUs. Something that has been wholly confirmed from initial tests and application in the Hangzhou City Brian project.
Hanguang 800 is the result of a culmination of both software and hardware development. In terms of hardware, it uses an in-house designed chip that takes advantage of such technologies as inference acceleration, which can help to resolve traditional performance bottlenecks. Next, on the software side, the chip is integrated with several algorithms developed at the DAMO Academy, which are specifically optimized for convolutional neural network (CNN) and computer vision algorithms, granting the tiny neural processing unit (NPU) the capacity to complete the computing operations of a large neural network.
The chip has already been applied and tested in some of the core business units of Alibaba Group. At the conference, it was revealed that the new chip was used in the City Brain project in Hangzhou to excellent results. For the project, previously 40 GPUs were needed to process the video feeds generated from the main urban districts of Hangzhou, but the same job could be done with only 4 Hanguang 800 chips. Also, at the same time, with the new chip, latency could be reduced from 300 to 150 milliseconds.
An AI cloud service developed based on Hanguang 800 was also launched on the same day in this year’s Apsara Conference. Powered by the new chip, the service offers a much higher cost-performance ratio, with up to 100% in performance increases, compared to traditional GPUs.
Connected with these announcements, Alibaba’s semiconductor subsidiary Pingtouge over the past six months has also launched the XuanTie 910 processor and Wujian SoC chip platform. The launch of Hanguang 800 marks Alibaba as being well on its way to providing an innovative chip suite for cloud and edge computing scenarios — one that covers chip processors, all-in-one chip development platforms, and, as of now, AI inference chips.