Recently, the U.S. ban on sales to ZTE has aroused much attention to the chip industry in China. Today, Alibaba DAMO Academy (DAMO stands for Discovery, Adventure, Momentum, and Outlook) claims that it is developing a neural network chip called Ali-NPU, which will be used in AI reasoning calculations such as video image analysis and machine learning. DAMO Academy says that according to its design, the chip’s cost performance will be 40 times that of existing products.
Jiao Yang, a researcher from DAMO Academy, explained that as general-purpose computing chips, CPU and GPU are designed to deal with linear logic and graphics, but they have high power consumption and low cost performance when dealing with AI computing. Thus specially designed chips are needed for AI computing.
At present, AI chips developed by Alibaba are mainly used to solve AI-related computing problems in commercial scenarios including image and video recognition as well as cloud computing, and to improve computing efficiency and reduce costs.
According to Jiao Yang, Alibaba‘s Ali-NPU is developed independently based on the large number of AI algorithm models accumulated by Alibaba A.I. Labs and other teams. It can achieve AI model algorithmic computing of the maximum efficiency at minimum cost according to the microstructure and instruction set based on AI algorithm models.
The Ali-NPU chip is designed to have a special architecture with a performance that is 10 times better than mainstream AI chips based on CPU or GPU architecture, while the manufacturing cost and power consumption are only half of theirs. Its cost performance is over 40 times of existing options. In the future, Ali-NPU can not only meet the demand for video and image processing, but could also for computing output through Ali Cloud, empowering all industries.
In fact, Alibaba has been building a team to develop their own AI chips for a long time.
The Alibaba-developed AI chip is a part of Alibaba‘s “China Chip” strategy. At present, the R&D team of DOMA Academy has hired dozens of talent in the United States and Shanghai, and will have 100 people by the end of this year. Prior to this, Alibaba has invested in a number of chip companies including Cambricon, Barefoot Networks, Deephi, Kneron, ASR, and C-Sky.