Editor's note: at the 2018 CCF-GAIR global artificial intelligence and robot summit in Shenzhen, Wei Shaojun, director of the Tsinghua University microelectronics, gave a speech on the architecture of the AI era chip. After the meeting, Wei Shaojun accepted an interview with NetEase intelligence, talking about the development of China's chip industry, architecture innovation and core upsurge.
Produced NetEase intelligence (public number smartman163) to collect Xiao Yi.
As the director of the Microelectronics Institute of Tsinghua University and the top bull in the chip industry, Wei Shaojun asserted that chips are the natural carrier of artificial intelligence.
So, it has also caused the current chip industry is particularly restless and hot, whether it is a giant company or a start-up company, whether it is a traditional manufacturing company or the Internet Co, in the layout of the AI chip industry.
In the era of AI, the architecture innovation of chips is an unavoidable task.
As market demand for chip computing increases, chip manufacturing technology is also improving, with chip manufacturing costs rising. People in the industry say that the cost of designing a 10nm chip is tens of millions of dollars. Even if the current Cambrian, deep learning technology and other enterprises specializing in AI chip, even using the 28nm chip technology, the cost of the earlier period from the input to the film should be more than $4 million. That is to say, the shipments of a single chip can not be recovered without millions of shipments.
Wei Shaojun also told NetEase intelligently that he wants to make AI chips more usable
Based on this, Wei Shaojun listed several basic elements that AI chips should have:
1, programmability, to adapt to the evolution of algorithms and application of diversity;
2, the dynamic variability of the architecture, adapting to different algorithms and achieving efficient computation.
3, efficient architecture transformation ability,
4, high computational efficiency and avoid inefficient architecture such as instruction.
5, high energy efficiency, ~10 TOPS/W;
6, low cost, access to household appliances and consumer electronics;
7, small size, can be loaded on the mobile device;
8. Application development is convenient without knowledge of chip design.
Judging from the above factors, the current approach of CPU SWU CPU GPU CPU FPGA CPU ASIC is not an ideal architecture.
At the same time, Wei Shaojun has divided the current AI chip scheme from the hardware and software programmability dimensions.
The first one is CPU, DSP and other processors. The hardware of the chip is very programmable, the software is programmable, and the chip needs software. Although these chips are versatile and highly flexible, they are expensive, expensive, inefficient in energy efficiency and efficient in computation, and are oligopolistic in the industry.
The second schemes are ASIC, SoC and other specialized integrated circuits. The software and hardware of the chip are very weak. Although the price is cheap, the energy efficiency and the computing efficiency are very high, the special chip does not have the flexibility. Once the hardware is finished, the hardware can not be changed.
The third scheme is FPGA, EPLD and other programmable logic devices. The chip has high programmability and weak programmability. Although hardware is programmable, static programming can not be changed once the configuration is completed, logic cannot be replicated, and the cost is high, the development cycle of the chip is long, energy efficiency and computing efficiency are low.
The fourth schemes are software definition chips (SdC) such as RCP and CGRA. The software and hardware can be programmed. The function of the chip varies with the software, and it has general and high flexibility, and the cost can be reduced in the future.
The software defined chip is proposed. The Thinker chip research group will finance and establish the company.
The software defined chip is a new chip architecture definition proposed by Tsinghua microelectronics, led by Wei Shaojun, which enables the chip to adapt and adjust according to the software. Based on this framework, Tsinghua microelectronics related research group has developed a new Thinker chip, using reconfigurable computing chip architecture. In this architecture, Thinker can configure the array of computing units and execute according to the requirements of the control unit, and can also send the divided tasks to the data channel according to the requirements of the software. Moreover, unlike the Von Neumann equivalent architecture, the reconfigurable computing mode is a functional hardware architecture.
At present, Thinker series chips have already introduced the Thinker 1 generation for general neural network computing, Thinker 2 for extremely low power neural network computing, and Thinker S for extremely low power voice applications. Before the media also reported that the power consumption of the Thinker chip is very low, only need 7 AA battery can run for a year.
Wei Shaojun also told NetEase intelligence that the Thinker chip R & D team is currently financing to set up a new company to operate in the market, but Wei Shaojun is inconvenient to disclose the amount of financing.
Suggestions for domestic core enterprises: money alone can not be broken, and must be opened.
For the current layout of Chinese enterprises chip, Wei Shaojun expressed support, but made some suggestions. He said,
Talking about the current AI chip development, Wei Shaojun said that the current AI chip market is overhyped.
Can AI chips exist independently? I don't know yet, but don't deceive yourself by making voice chips.
Speaking of many domestic startups starting to make voice chips, Wei Shaojun commented, "what exactly is the voice chip going to do?" What are the functions to be implemented? First of all, we must make it clear.
In Wei Shaojun's view, the speech chip must look at the application scene, at present many scenarios do not need the artificial intelligence technology or the specialized speech chip.
Finally, Wei Shaojun said that although AI technology is constantly improving, at present it is still mainly focused on training and reasoning for individual affairs, so that AI can really make multiple judgments and decisions at the same time as human beings. A greater breakthrough is needed in the research of algorithms.