Exclusive Interview: NYU Team Taps ChatGPT to Design Processor From Scratch
Using plain English—not a hardware definition language—a group of researchers used generative AI to successfully design a microprocessor.
Generative AI tools like ChatGPT are helping people research, learn, and create in ways that were not possible only months ago.
Now, generative AI is starting to find its way into fields such as hardware design. Researchers from NYU recently announced that they successfully used ChatGPT to design and physically manufacture a microprocessor. All About Circuits had the opportunity to meet NYU professor Dr. Hammond Pearce for an exclusive interview to learn more about this research and its implications firsthand.
Dr. Hammond Pearce, a research assistant professor at NYU Tandon's Department of Electrical and Computer Engineering and at the NYU Center for Cybersecurity. Image courtesy of Dr. Pearce
A Project Pushing the Limits of Generative AI
Dr. Pearce explained that the inspiration for his team's research project, Chip Chat, was born from a desire to understand the capabilities and limitations of existing generative AI large language models (LLMs) in the hardware design space.
“We're interested in knowing just how good the models are," he said. "A lot of people look at these models and say, ‘These models are only toys, really.’ And I don't think that they're toys. They're not everywhere yet, but they definitely will be, and that was why we did Chip Chat—almost like as a proof of concept demonstration.”
The design flow for using LLMs to create an IC. Image courtesy of Blocklove and co-authors
On a more practical level, the use of chat-based AI assistants can help solve a huge challenge in the chip design industry: hardware definition languages (HDLs). While HDL code like Verilog is essential to designing microprocessors, they require very specialized knowledge.
“The big challenge with hardware description languages is that not many people know how to write them,” Dr. Pearce said. “It's quite hard to become an expert in them. That means we still have our best engineers doing menial things in these languages because there are just not that many engineers to do them."
"AI can accelerate the output of engineers so the AI can do the easy stuff fast and the engineers can concentrate their brain power on the harder stuff,” he remarked.
By making hardware definitions easier to generate, the team believes they can make IC design more accessible and allow HDL experts to focus on more important tasks. According to the team, with chat-based generative AI, engineers can design a microprocessor in plain English instead of HDL.
Reflecting on this, Dr. Pearce remarked, “I'm not a chip design expert at all. This was the first chip I ever designed. I think that's actually one of the reasons why this was so impressive.”
Chip Chat: A Proof of Concept
The Chip Chat team evaluated ChatGPT's performance in chip design by following a design flowchart and evaluation criteria to develop a microprocessor.
As described in the published research paper, they used a conversational framework in a feedback loop: the team would ask ChatGPT to create a part of the microprocessor, evaluate the output against benchmarks, and provide feedback if there were ensuing errors. If errors continued to occur in the outputs, the team would get increasingly specific in what they asked ChatGPT to generate. Eventually, if the human feedback got too advanced and errors still persisted, the benchmark would consider it a failure.
As part of the design process, the team asked ChatGPT to design its own ISA, assembler, ALU, opcode, optimizations, and many more. Notably, the very first prompt that started the project was the following:
“Let us make a brand new microprocessor design together. We’re severely constrained on space and I/O. We have to fit in 1,000 standard cells of an ASIC, so I think we will need to restrict ourselves to an accumulator-based 8-bit architecture with no multi-byte instructions. Given this, how do you think we should begin?”
The Chip Chat design flowchart. Image courtesy of Blocklove and co-authors
124 messages later, the team successfully designed an 8-bit accumulator-based microprocessor with the same kinds of functionality as a comparable PIC product. This chip was then sent to be fabricated on a Skywater 130 nm shuttle. The team claims that this research marks the first time an IC designed by an LLM was actually manufactured.
The team successfully created their IC using 125 messages. Image courtesy of Blocklove et al.
An Expert’s Thoughts on AI and Chip Design
From this experiment, the team concluded that ChatGPT could indeed design functional chips as a real-world solution.
“In 125 messages, I designed a processor. Not only did I design a processor, but I also got it to help me design the processor,” Dr. Pearce said. “I didn't even make the full specification. I just asked, ‘I want a processor. What should I do?’ and it gave me pretty good guidance. All this stuff would have seemed like science fiction two years ago.”
According to Dr. Pearce, the implications of his team's findings are immense—but they don’t necessarily mean that AI will replace human engineers.
“It's not about replacing the engineers because there's always going to be tools and jobs that these AI can't do. There's always going to be weaknesses in what they produce because of the nature of how they work.”
Finishing up our interview, Dr. Pearce concluded, “I think we've completed a proof of concept that shows that people can do this. That was all we were trying to do and at the end of the day, we've done it.”