top of page

ChatGPT Accelerates Chip Design


Exciting new research shows that large language models like ChatGPT can be trained to design computer chips faster than humans. This has big implications for the future of chip design and the semiconductor industry.


Researchers at Georgia Tech recently published a paper showing how they used natural language prompting to get ChatGPT to generate verilog code for parts of a RISC-V chip. While ChatGPT can't yet fully design a chip from scratch, this shows the potential of using AI for chip design automation.


The conventional chip design process takes 1-2 years with a team of experienced engineers. But AI and natural language processing could accelerate this dramatically. The researchers gave ChatGPT prompts like "design a RISC-V AI chip in 7nm technology."


Of course, ChatGPT can't do this from just a simple prompt yet. It needs more context and examples to learn from. The problem is ChatGPT doesn't inherently understand hardware concepts or make connections between code and hardware.


The researchers had to train the model specifically on chip design by feeding it code examples and documentation. Interestingly, they used ChatGPT itself to generate documentation for code samples, which was then used to train the model.


While ChatGPT can generate code, it lacks creativity and the ability to reason at a hardware level. The goal is to teach large language models to explore the design space and optimize the architecture and microarchitecture. There's still a long way to go.


Meanwhile, Google DeepMind made progress using neural networks to optimize circuit design. Once code is generated, it must be synthesized into a physical chip layout. DeepMind uses reinforcement learning to find the most efficient circuit architecture.


In a recent contest, DeepMind's AI achieved the most optimized chip designs for 82% of the entries, showing the promise of AI for circuit optimization and lowering power usage.


While actually designing full chips with AI is still far off, tools like ChatGPT and circuit neural networks could one day accelerate and improve parts of the chip design process. This could lead to better and more specialized chip architectures.


The key will be providing enough quality data and documentation to train the AI systems. In the future, prompt engineering and reinforcement learning may allow an AI assistant to generate optimized chip layouts for given parameters and goals.


The semiconductor industry is ripe for advancements from AI and machine learning. While humans will still lead creative chip design for the foreseeable future, AI tools could make chip designers far more productive. This research shows glimpses of that future.



Comentarios


bottom of page