Researchers at NYU Tandon School of Engineering have achieved a groundbreaking milestone in chip development by fabricating a microprocessing chip using plain English “conversations” with an AI model. This achievement has the potential to expedite chip design and enable individuals without specialized technical skills to participate in the process.
The research team utilized ChatGPT-4, a Large Language Model (LLM) designed to comprehend and generate human-like text, to engage in English conversations with two hardware engineers. Together, they designed a new microprocessor architecture and subsequently sent the designs for manufacturing.
Traditionally, hardware development, including chip design, begins with describing the intended functionality in regular language. Trained engineers then translate these descriptions into Hardware Description Languages (HDLs), such as Verilog, to create the necessary circuit elements for the hardware.
In this study, the LLM successfully generated functional Verilog code through iterative dialogue with the engineers. The resulting chip design, which included benchmarks and the processor, underwent manufacturing using the tapeout process, utilizing the Skywater 130nm shuttle, a specific semiconductor manufacturing service made accessible through Tiny Tapeout.
Hammond Pearce, research assistant professor at NYU Tandon, expressed the significance of the study, stating, “This research shows AI can benefit hardware fabrication too, especially when it’s used conversationally, where you can have a kind of back-and-forth to perfect the designs.” While AI models like OpenAI’s ChatGPT and Google’s Bard have previously generated software code, their application in hardware design had not been extensively explored until now.
The NYU Tandon research team, consisting of Professor Ramesh Karri, Institute Associate Professor Siddharth Garg, and doctoral student Jason Blocklove, employed LLMs to generate Verilog code for functional and verification purposes across eight hardware design examples. Their focus then shifted to the deep-dive case study of chip fabrication. The researchers discovered that incorporating interactive dialogue with a live engineer produced the most favorable outcomes.
If implemented in real-world scenarios, the utilization of LLM conversations in chip fabrication could reduce human errors during the HDL translation process, enhance productivity, shorten design time and time to market, and facilitate more innovative designs. Additionally, this approach could eliminate the requirement for chip designers to possess fluency in HDL, which is a relatively scarce skill and a significant barrier to entry for many individuals seeking chip design positions.
The researchers acknowledged the need for further testing to address security considerations associated with using AI in chip design.
The signing of the federal CHIPS Act into law in August 2022 reflects the United States’ effort to strengthen domestic semiconductor research and manufacturing. Currently, the U.S. accounts for only about 12% of the global semiconductor manufacturing capacity. The impact of chip shortages during the COVID pandemic highlighted the importance of bolstering chip production to ensure availability for various chip-dependent devices, including automobiles.
Source: NYU Tandon School of Engineering