Researchers Build an AI for Chip Design, Which Churns Out a 486-Like RISC-V Chip in Just Five Hours

Researchers from the Chinese Academy of Sciences, the University of Science and Technology of China, and Cambricon Technologies Corporation have taped out what is positioned as the first fully-functional CPU designed by artificial intelligence (AI) — performing on a par with Intel's classic, human-designed i486 chips.

"To explore the boundary of machine design, here we present a new AI approach to automatically design a central processing unit (CPU), the brain of a computer, and one of the world's most intricate devices humanity have ever designed," the research team explains of its work. "This approach generates the circuit logic, which is represented by a graph structure called Binary Speculation Diagram (BSD), of the CPU design from only external input-output observations instead of formal program code."

Traditionally, designing a CPU is an extremely labor-intensive job — even if using an off-the-shelf instruction set architecture. By using an AI investigating a gigantic search space, though, the team was able to spit out a chip built around a 32-bit implementation of the free and open-source RISC-V instruction set architecture (ISA) in just five hours — with, the team claims, a greater than "99.99999999999% accuracy" despite having nothing but external observation of inputs and outputs to go on.

"Compared to human-designed CPUs," the team boasts, "our approach reduces the design cycle by about 1000× because the manual programming and verification process of traditional CPU design is completely eliminated. Thus, our approach changes the traditional CPU design flow and potentially reforms the semiconductor industry."

As anyone who has played with a large language model (LLM), the current generative-AI technology of the day, will know, the output of a generative AI model isn't always elegant, functional, or even factual. The team's chip, though, has been proven in hardware — taped out, produced, and tested running the Linux kernel. Its performance, meanwhile, benchmarks to somewhere in the region of an Intel i486SX, a 32-bit chip released back in 1991.

The work of actually getting an AI model which can generate a usable processor design, however, takes considerably more than five hours — and only time will tell if such an approach can keep up with human designers at the cutting edge, rather than the state of the art from three decades ago.

A preprint detailing the team's work is available on Cornell's arXiv server.