Book Review AI'S "Brain"

By Zheng Lei

In the next 30 years, big data artificial intelligence will be an important focus area of science and technology competition among countries. At the same time, it is also the key driving industry of digital economy. Computing power is one of the core elements of AI development. AI chip is the "brain" of AI computing hardware. From the initial use of graphics processor as a deep learning acceleration chip, to the special chip customized for AI, AI chip has rapidly developed into an emerging industry in just a few years.

AI research results and patents are showing an explosive growth, and the AI field has now become a vast "forest". From the development history of AI, AI chip: cutting edge technology and Innovation future, which is illustrated and illustrated, covers the algorithm, architecture and circuit of the most popular deep learning acceleration chip and brain like chip based on neural morphology computing. It also focuses on various computing paradigms for AI chip design with innovative thinking. One book leads readers directly to the frontier field of artificial intelligence.

The rapid development of AI has been limited by computing power, while the manufacturing process of silicon-based chips based on CMOS has been refined to 3 nm. Not only will the improvement of AI computing power encounter bottlenecks, but also due to the characteristics of AI computing, some computing functions are less used, so a considerable part of the computing power that has been realized by ordinary CPU is not available to AI. This is the objective requirement of AI chip. For example, deep learning needs matrix multiplication (tensor multiplication), and the efficiency of GPU is significantly higher than that of CPU with higher computing speed.

Today, graphics processor (GPU) chips are mainly used for deep learning, which can achieve the operation speed of more than 100 trillion floating-point operations per second, which is more than 6000 times the operation speed of super computer cray-3. But that's a lot worse than the human brain. There are about 100 billion neurons in the human brain, and more than 100 trillion synapses are involved in neuronal signal transduction. The brain can recognize patterns, remember facts, and process other tasks in parallel at lightning speed, consuming less than 20 watts.

With the advent of the "post Moore era", the computing speed of silicon-based chips has been approaching the peak. At the same time, another problem has arisen, that is, this computing structure is more suitable for structured data, but for processing unstructured data such as big data, the method of improving chip computing density is invalid. According to the data given by the author, after the processing technology is higher than 10 nm, although the calculation speed of the chip is still improving rapidly, the relative efficiency of processing big data is getting lower and lower. According to the current technology level, if multiple GPUs are used in parallel, the power can easily exceed 1000 watts. In 2016, when alphago played Li Shishi, a nine stage go master, the power consumption of the server running the AI program reached 1 megawatt, nearly 50000 times the power consumption of the human brain. People must look for other chip materials and adopt computing architecture, models and algorithms suitable for big data artificial intelligence. In this regard, Chinese scientists have made great contributions. In 1971, Professor Cai Shaotang discovered and proved the existence of a new basic circuit element memristor based on circuit theory. In 2008, Stanley Williams of Hewlett Packard first supported the first prototype based on titanium dioxide film in the laboratory. This new device can significantly reduce the computational power consumption, and is a potential hardware solution for deep learning accelerator and brain like chip.

Although the mainstream of AI chip development is based on deep learning algorithm, the main technical route is to integrate special AI chip and multi-core CPU processor into the same chip. In the field of semiconductor chip, the development of FPGA and ASIC is an important direction. The more advanced research is to design "evolvable" chips, which are basically close to improving the performance of chips through "self-learning".

The ultimate goal of AI chip is to be able to "self learn", that is, the chip can learn "how to learn" by itself; Another important goal is to achieve mutual learning and coordination between intelligent machines (equivalent to between AI chips), so that intelligent machines can get more knowledge. The performance of this "self-learning" is likely to increase exponentially over time, and eventually lead to the intelligent level of intelligent machines surpassing that of human beings. This kind of chip is often close to the biological characteristics of the brain in the design process, so it is called brain like chip or neural morphology chip. Based on the new chip architecture, the key components of the chip include pulse neurons, low precision synapses and scalable communication network.

China has listed chips as a technology that must be controlled by itself, and in big data, where the gap between China and Europe and the United States is very small, artificial intelligence has become the next technological highland to be occupied. At present, most of the basic patents in the field of silicon-based semiconductor chips are occupied by Europe and the United States, so it is difficult to overtake along this route. AI chips have brought us a rare new opportunity. The research and development of brain like chips and fields based on new neural network algorithms, quantum inspired algorithms, natural bionic computing, in memory computing and new memory, as well as quantum computing and quantum machine learning, are still in the stage of laboratory samples or a small number of trials, which can be used as the goal of overtaking Chinese chip technology.

?