A group of Chinese scientists published (PDF) Paper titled “Pushing the Boundaries of Machine Design: Automated CPU Design with AI”. The paper details the researcher’s efforts to design a new industrial-scale RISC-V CPU within his five hours. The AI-automated feat is said to have been about 1000 times faster than a human team could complete a comparable CPU design. However, some might tease that the AI-designed CPU’s performance is nearly on par with the i486.
The Chinese research team’s goal was to answer the question of whether machines can design chips in the same way humans do. The team believes that early designs created by AI were relatively small or limited in scope. So the researchers wanted to let AI automatically design his RISC-V CPU to test the limits of AI design.
Such projects usually start with a period of machine learning. Training consisted of observing a series of CPU inputs and outputs. Scientists generated a binary speculation diagram (BSD) from this I/O and leveraged Monte Carlo-based extensions and Boolean function principles to refine the accuracy and efficiency of AI-based CPU designs. Therefore, the scientists explain that the CPU design was formed “only from external input/output observations, not from formal program code.” It also boasted an incredible accuracy of 99.99999999999%.
Using the process above, an automated AI design for the CPU was created. The taped-out RISC-V32IA instruction set CPU was manufactured at 65nm and could run up to 300 MHz. We verified its functionality by running a Linux (kernel 5.15) operating system and SPEC CINT 2000 on an AI-generated CPU. In the Drystone benchmark, the AI-generated CPU performed on par with the i486. Interestingly, it seems to be slightly faster than the Acorn Archimedes A3010 in the same tests.
While some may not be impressed with the performance of AI-generated CPUs, scientists also believe that the BSDs they generate are ” von neumann architecture from scratch. ”
Building new RISC-V CPUs from scratch with AI is not only of academic interest. It could also potentially be used to create a new CPU from scratch. Researchers say AI has the potential to significantly shorten the existing semiconductor industry’s design and optimization cycle. Furthermore, in their conclusions, the scientists also consider whether this work can be developed further to form the basis for self-evolving machines.
This is not our first story of AI being used to advance the design of computer processors. Back in March, we reported on Nvidia using AI to optimize chip design, especially floorplanning work. Also in May, we reported that Synopsys boasted that its DSO.ai software was used in over 200 of his customer chip designs.
