A team of researchers from the State University of New York (NYU) They successfully designed a semiconductor chip without using a hardware definition language. use plain English only – and definitions and examples that can define and explain semiconductor processors in it – the team explores what human ingenuity, curiosity and basic knowledge can do with the help of ChatGPT’s AI capabilities introduced.
Surprisingly enough, the chip wasn’t just designed. it was manufactured. I ran some benchmarks and it worked. The use of plain English by two hardware engineers shows just how valuable and powerful ChatGPT is (even after the numerous awes that have already taken place, we still as if in doubt).
The chip designed by the research team and ChatGPT was not a complete processor. Nothing beats processors from Intel and AMD like processors on our best CPU list. But this is the whole CPU element, the logic that creates the new 8-bit accumulator-based microprocessor architecture. accumulator It is essentially a register (memory) where the results of intermediate computations are stored until the main computation is completed. But they are essential for CPU operation. Perhaps other necessary bits can be designed as well.
Teams typically work in several stages to bring a chip into design and manufacturing. One such stage is transforming “plain English” describing the chip and its functions into a chosen hardware descriptor language ( HDL) (eg Verilog). The chip needed for the etching itself.
ChatGPT is a pattern-recognition machine (just like humans, but we’re a little more than that) that believes in all kinds of languages: spoken languages, written languages, and here especially hardware-based languages. It’s incredibly helpful. ChatGPT allowed the engineer to skip her HDL stage. This is great, but it should make HDL engineering professionals a little nervous. In particular, researchers expect to reduce human error in the HDL translation process, contributing to increased productivity, shortening design time and time to market, and enabling more creative designs. because it says it does.
“We are interested in knowing how good the models are,” he said. “A lot of people look at these models and say, ‘These models are actually just toys,’ and I think they’re not toys. They’re not everywhere yet, but It definitely will, and that’s why we did a tip chat – almost like a proof-of-concept demonstration.”
Dr. Hammond Pierce, New York University Tandon
One thing that is a little more concerning (or at least debatable) is the desire among chip designers to eliminate the need for HDL fluency. It is a relatively rare skill that is very difficult to master as it is a highly specialized and complex field.
“The big challenge with hardware description languages is that not many people know how to describe them,” said Dr. Pierce. “It’s very difficult to become an expert in them. That means there are still good engineers doing simple tasks in these languages because there are not that many engineers who can do these languages.” because there isn’t.”
Of course, automating part of this process is definitely beneficial. New experts can be developed and trained while existing experts can respond quickly, potentially reducing human bottlenecks. But there is a risk in making this skill completely dependent on a software-based machine that relies on electricity (server connectivity in ChatGPT’s case) to operate.
There is also the issue of trusting something that is inherently trustworthy. A black box of arcane software and its output. Having seen what happens with prompt injection, LLM is not immune to vulnerabilities either. Since these are not just software, they are the result of training, so we can also think of them as having extended vulnerabilities. And chip-based LLMs “Clever as the devil” hardware-based backdoor … is connected somewhere. This may sound exaggerated, but indeed it is at the absolute lower end of the scale of possibilities. But with mutating malware and other nasty surprises also coming from today’s versions of large language models, what is there to say about what will erupt from them tomorrow?
The researchers worked on eight example hardware designs using commercially available and publicly available Large Language Models (LLMs), translating plain text into Verilog ( HDL) equivalent text.
“As a result of this work, we believe it is the first fully AI-generated HDL submitted for fabrication into a physical chip,” said Assistant Research Professor and member of the research team at New York University’s Tandon College. said Dr. Hammond Pierce. “Some AI models, such as OpenAI’s ChatGPT and Google’s Bard, can generate software code in a variety of programming languages, but their application to hardware design is still poorly studied.” , shows that AI can also be beneficial for hardware manufacturing, especially when used interactively, where you can kind of interact to complete a design.”
Several electronic design automation (EDA) tools already exist, and AI has shown impressive results in chip layout and other factors. But ChatGPT is not special software. Apparently, he can write poetry and make EDA cameos. The path to becoming an EDA designer has a much lower knowledge barrier. Perhaps someday enough of the CPUs will be published that anyone with enough determination (and valuable backing) for ChatGPT will be able to design their own CPU architecture at home.
Yes, there will be many questions about what that means. But is it possible?
