The New York State University (NYU) have achieved a remarkable feat by designing a semiconductor chip using plain English descriptions and the assistance of ChatGPT. This chip, which represents a component of an 8-bit accumulator-based microprocessor architecture, was not only designed but also manufactured and successfully benchmarked. Typically, chip design involves multiple stages, including the translation of plain English descriptions into Hardware Descriptor Language (HDL) for the manufacturing process. However, the researchers were able to bypass this stage by leveraging ChatGPT’s language understanding capabilities.
The elimination of the need for HDL fluency among chip designers raises both excitement and concerns. While automating parts of the chip design process can enhance productivity and shorten design time, it also raises questions about the reliance on software-based machines and their potential vulnerabilities. Trusting the outputs of an inscrutable software black box, like a large language model (LLM), carries some risks, as we have witnessed with prompt injection and other vulnerabilities. There is a remote possibility that a chip-based LLM could be compromised during training, potentially introducing a hardware-based backdoor. While this may be on the lower end of the possibility scale, it underscores the importance of considering the potential risks associated with the increasing capabilities of LLMs.
The researchers used LLMs, including OpenAI’s ChatGPT and Google’s Bard, to convert plain English descriptions into Verilog (HDL) code through interactive conversations between engineers and the models. This study marks the first instance of an AI-generated HDL being sent for fabrication into a physical chip. While there are already impressive AI tools in Electronic Design Automation (EDA) for chip layout, ChatGPT’s ability to assist in hardware design demonstrates its versatility beyond specialized software applications.
This advancement in chip design lowers the knowledge barrier for entry into the field of EDA design. It opens up the possibility that, with determination and the help of models like ChatGPT, individuals could one day design their own CPU architectures at home.
