32 Bit RISC-V CPU
In the first half of 2023, the AI wave led by ChatGPT has swept the world, and major manufacturers have begun to deploy their own AI. The high demand for computing power of generative AI has also made the sales of GPU chips soar all the way, and the global computing power resources are in a state of shortage. In order to improve the productivity of the chip, the Institute of Computing Technology of the Chinese Academy of Sciences used AI technology to design a 32-bit RISC-V CPU “Enlightenment No. 1”-this is the world’s first fully automatic CPU chip without human intervention.
It is understood that this CPU uses a 65nm process with a frequency of 300MHz, and the CPU has been taped out in December 2021. The CPU successfully runs the Linux operating system and the SPEC CPU 2000 program after Bring-up, and its performance is equivalent to that of the 80486SX of the Intel 486 series, but the design cycle is shortened to 1/1000 of it.
Traditional CPU Vs. AI CPU Design
According to reports, advanced chip manufacturing must go through more than 1,000 steps. Complex calculations are required at each stage, and each step must be nearly perfect, and the same is true for CPU design and manufacturing. In this challenging job, a team of engineers is usually required to write the code and then complete the circuit design with the aid of EDA tools. In this process, repeated testing and verification optimization are indispensable. It is precisely because the process is extremely complex and tedious, and it is extremely labor-intensive and resource-intensive. Sometimes it takes months or even years to complete a large-scale chip design project. Taking the Intel K486 CPU as an example, it takes 190 days for the verification stage alone.
The researchers of Enlightenment 1, based on the BSD Binary Speculation Diagram (BSD) algorithm, use AI technology to automatically generate CPU designs directly from the input-output (IO) of test cases. In this process, there is no need for engineers to provide any code or natural language description, and it is completed within 5 hours. Although the training process takes less than 5 hours, it can achieve >99.999999999% verification test accuracy, so only test cases are needed to directly generate circuit logic that meets the requirements, which can eliminate the logic design and verification in the traditional design process link.
In fact, this is not the first time that AI has generated a chip. Previously, researchers from the Tandon School of Engineering at New York University had used ChatGPT-4 to generate a microprocessor chip through simple English “dialogues”. Four million logic gates were generated within an hour, which is 4,000 times larger than the current circuit scale that can be designed by GPT-4.
At present, with the development of AI technology, more and more companies have begun to introduce it into chip design and manufacturing. Some industry experts said that although the performance of the Enlightenment No. 1 CPU is still relatively poor and cannot be compared with top-level CPUs, as long as the AI is continuously trained and the technical path is continuously explored, then the CPU automatically generated by AI is expected to be in 5 years. -In 10 years, reach or even exceed the level of human expert design CPU.
However, I think it is doubtful whether AI will really “eliminate” human chip designers in the future. After all, before reaching the stage of strong artificial intelligence, AI may only be used as an auxiliary After all, the continuous innovation of human chip designers is the foundation of the continuous development and progress of chip technology.