ChatGPT: How NYU Researchers Used AI to Design a CPU
26-06-2023 | By Robin Mitchell
In a feat to demonstrate the capabilities of advanced AI systems, researchers recently published their findings on a CPU entirely designed using ChatGPT, avoiding the need for HDL. This groundbreaking research was conducted by a team at the NYU Tandon School of Engineering, who fabricated a microprocessing chip using plain English 'conversations' with an AI model, marking a first-of-its-kind achievement that could lead to more democratized and faster chip development [1].
What exactly did the researchers do, how does this demonstrate the capabilities of AI, and what does this mean for the future of hardware development?
Researchers create CPU using ChatGPT
Recently, a team of researchers from New York State University demonstrated the power of ChatGPT by using it to design a CPU entirely through prompts. While the CPU designed by ChatGPT was only simple in nature (being an 8-bit accumulator system), the resulting HDL code was turned into a physical device that was benchmarked and demonstrated to work successfully.
In their study posted to the arXiv pre-print repository, the research team presents how two hardware engineers 'talked' in standard English with ChatGPT-4 to design a new type of microprocessor architecture. The designs were then sent for manufacturing [1].
Dr. Hammond Pearce, one of the researchers involved in the project, explained in the study, 'Our work demonstrates the potential of using AI models to have conversations that can aid non-experts in designing hardware, thereby making the field more accessible' [2].
Dr. Pearce further elaborated on this in his interview, stating, 'We wanted to see if we could leverage the power of these AI models to help in areas where there's a high barrier to entry, like hardware design. The results were promising and showed that AI could indeed be a valuable tool in this field' [3].
During the design of complex logic devices, engineers need to take a set of requirements and abstract plans (such as the general structure of a CPU) and convert this into a Hardware Description Language (HDL) model. From there, automatic translation processes can then turn the HDL into a physical layout, with minimal interference by engineers, and this is crucial in modern designs containing billions of transistors.
Reflecting on the potential implications of their research, Dr. Pearce said, 'Imagine a future where anyone with a great idea can design and create their own microprocessor. This could revolutionize the tech industry and lead to a surge of innovation' [3].
However, the development of HDL is extremely complex and time-consuming. Furthermore, HDL can be complex to understand, meaning that in the design of a CPU, only those involved with the HDL are able to understand it (thereby limiting the contributions of others on the team). In the case of the CPU designed by ChatGPT, all the HDL code was automatically generated from plain English parameters provided by the researchers, and while some minor errors had to be fixed by hand, the vast majority of the HDL was perfectly usable.
How does this demonstrate the capabilities of AI?
While ChatGPT has been frequently used to generate code in all major computer languages, the ability to design hardware indicates a significant step in AI technological development. If AI can be used to create new hardware, it is possible for an AI system to self-improve, with each iteration utilising better hardware and software. Thus, in theory, it is also possible for AI to reach a singularity of ability whereby it quickly becomes more powerful than anything ever developed by mankind.
In an exclusive interview with All About Circuits, Dr. Hammond Pearce, a research assistant professor at NYU Tandon's Department of Electrical and Computer Engineering, explained that the inspiration for the 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 [3].
Of course, current AI systems have their limitations, and the hardware that they design is entirely dependent on human-driven industries such as semiconductor fabrication and circuit manufacturing. Furthermore, the hardware developed by ChatGPT was based on HDL, which is a logical representation of a design, as opposed to an analogue representation which would consider real-like physical limitation.
For example, circuits in real life need to use decoupling capacitors to filter noise, use negative feedback to provide stability and choose components based on the environment that the design will be used in. By contrast, the CPU designed by ChatGPT is closer to a software problem than a hardware one, as the resulting HDL doesn’t consider the impacts of real-life, such as signal interference, environmental conditions, or technology node.
Regardless, the fact that a working CPU was designed by ChatGPT demonstrates just how the abilities of AI continue to progress rapidly. If taken further, it is highly likely that the researchers could have also designed external memory chips and peripherals for the CPU, turning it into a complete SoC.
What does this mean for the future of hardware design?
Despite the fact that AI is still in its infancy, it has already been deployed in numerous engineering tools, including antenna design and chip layout. Its ability to self-improve makes it a highly effective tool that only gets better with time, and by reducing the number of man-hours used, the costs of a project can be reduced while accelerating its time to market. But would the mass introduction of AI hardware designs eliminate the human factor completely?
While this is unlikely to be the case in the short term, it is perfectly possible for AI to replace humans largely in the design phase of a product, especially for mass-produced devices utilising older technology. Considering that AI systems are very good at identifying patterns and learning from massive datasets, it is not far from the realm of science fiction for an AI to study product enclosure design, recognise patterns in product design trends, and create new product enclosures that conform to current design and fashion trends.
According to the researchers, if implemented in real-world settings, using LLM conversations in chip fabrication could reduce human error in the HDL translation process, contribute to productivity gains, shorten design time and time to market, and allow for more creative designs [1].
From there, an AI could then consider the basic building blocks needed for a design (such as basic functions)```html and then design a piece of digital hardware translated from those building blocks. In fact, if future hardware trends towards fully programmable hardware platforms, such as FPGAs, then it is possible for AI to not only design the hardware but even identify issues and provide updates on-the-fly. For example, if a new hardware bug is identified by cybersecurity experts, an AI could find solutions to fix the bug in hardware, verify the design, and push out updates to all vulnerable devices.
Engineers will always be needed, and there is no doubt that this will be the case for decades to come. Instead of taking over engineering jobs, AI will empower engineers with new tools and abilities to spend more time focusing on important aspects of a design instead of tedious tasks such as layout and routing. But that isn’t to say that AI won’t take over engineering jobs, and what the researchers demonstrated with their AI-developed CPU does show that AI is certainly capable of far more than we could have ever imagined.
In conclusion
The research conducted by the NYU Tandon team has demonstrated the potential of AI in hardware design. The successful use of ChatGPT in designing a CPU not only showcases the capabilities of AI but also opens up new possibilities for the future of hardware development. As Dr. Pearce remarked in his interview, 'All this stuff would have seemed like science fiction two years ago' [3].
References:
