Can GPT‑4 Really Design a Chip? Inside NYU’s AI‑Generated Processor Breakthrough

Some commentators claim GPT‑4 resembles an early version of artificial general intelligence (AGI). Cerebras CEO Andrew Feldman estimated GPT‑4 contains roughly one hundred trillion parameters, comparable to the number of synapses in a human brain.

The human brain has about 800‑1000 billion neurons and a similar number of synapses, which govern all thoughts and actions over a lifetime. GPT‑4’s parameter count raises the question of whether it can approach human‑level capabilities.

Recent research from the NYU Tandon School of Engineering demonstrates a striking application: two hardware engineers conversed with GPT‑4 in ordinary English to design a new microprocessor architecture. The dialogue lasted 19 rounds, during which the model produced a complete Verilog description of an 8‑bit accumulator‑based processor.

After the conversational design, the Verilog code was sent to Skywater’s 130 nm “tape‑out” service. The chip was fabricated, benchmarked, and validated, marking the world’s first fully AI‑generated hardware description language (HDL) that resulted in a physical silicon device.

The researchers note that AI‑driven, conversational hardware design can reduce human errors in HDL translation, boost productivity, shorten design cycles, and enable more creative architectures. However, they also caution that further testing is needed to address safety and reliability concerns when applying LLMs to chip design.