◆ introduction to the current situation and development trend of AI chip
Series of lectures on innovation and Entrepreneurship of electronic information in Nanjing University
On the evening of April 24, 2020, a series of lectures on innovation and Entrepreneurship of electronic information of Nanjing University continued. Lin Jun, an associate professor of the school of Electronic Science and engineering of Nanjing University, was invited to give a lecture entitled "Introduction to the current situation and development trend of AI chips". Affected by the epidemic situation, this lecture took the form of online classroom. The lecture attracted more than 100 people from different departments and grades.
At the beginning of the lecture, Mr. Lin introduced the concept of AI - artificial intelligence. At present, in the field of artificial intelligence, there are many achievements, which are called "deep learning", or deep neural network. The basis of this kind of model is to imitate the "nerve pulse" in biological system: the signal is received by dendrites, sent by axons, and the signal strength is regulated by synaptic weight, which can be changed.
In order to adapt to the deep neural network, many manufacturers have launched their own artificial intelligence chips, such as IBM's truenorth chip, Intel's Loihi chip, etc. when a specific neural network model runs on these chips, its efficiency is many times higher than the usual x86 chip.
In order to let students have a more comprehensive understanding of deep neural network, Professor Lin Jun next introduced the basic concept of neural network in detail. In terms of model, neural network can be divided into three parts: input layer, hidden layer and output layer, while deep neural network generally refers to the high number of hidden layers. The operation of neural network consists of two steps, namely "reasoning": getting output based on input and given weight, and "training": mainly including back propagation algorithm and gradient calculation.
On the basis of deep neural network, Professor Lin also introduced the concept of so-called deep convolution neural network. Its hidden layer mainly includes convolution layer and pooling layer, and the mathematical principle of convolution is vividly displayed in the way of combining graphics and text.
After introducing the basic concepts of artificial intelligence, Professor Lin Jun mainly introduces the so-called "deep learning special chip" and the necessity of using deep learning special chip. At present, in the field of deep learning, it mainly relies on GPU as the basic hardware platform. Although it has the advantages of high parallelism, strong computing power and large capacity, the high power consumption of GPU results in the limitation of application scenarios.
And through the exclusive deep learning chip, in the case of ensuring accuracy, it can greatly save power consumption, which is the advantage of this kind of chip. However, the design of deep learning chip also faces many challenges: the realization of core computing, how to realize fast matrix multiplication with special circuit, how to adjust the balance of computing power, cache size and storage read-write speed, and how to complete reasoning (such as TPU) with lower precision numerical expression.
At the end of the lecture, Professor Lin made a prospect and summary for the future of AI chip: there are still many unsolved problems in the research of AI chip at present, new materials, devices and mechanisms in the future may bring new chip structures, and different chip architectures and collaborative development of algorithms and hardware are also needed for different scenarios.
Through this lecture, Professor Lin taught students a science popularization course about "artificial intelligence" from shallow to deep. As one of the most cutting-edge fields at present, students are also full of great interest in this topic. Through the detailed introduction of Mr. Lin, we can see how the electronic information specialty can play its professional advantages in the wave of artificial intelligence. I believe that in the future, AI technology can further benefit human society!
Editor: Zhang Zhijian
Electronic experiment center of Nantah University
Editor in chief: Zhang Zhijian
END