First AI chip designed with natural language processing could power energy-efficient artificial intelligence for robots and autonomous vehicles.
Statistically speaking, there are probably a fair number of engineers using ChatGPT in their day-to-day lives. With more than 180 million users, it’s safe bet that at least some of those people are engineers using the large language model (LLM) to draft emails or outline proposals. Outside of some fancy CAD features, that’s probably the most exposure working engineers have to natural language processing.
However, thanks to some recent research from Johns Hopkins University, that could soon change. An electrical and computer engineering graduate student and an undergraduate used ChatGPT4 to produce detailed instructions for building a spiking neural network chip, a design that operates much like the human brain.
A series of prompts took ChatGPT from mimicking a single biological neuron to linking them to form a network and eventually generating a full chip design for fabrication.
“This is the first AI chip that is designed by a machine using natural language processing. It is similar to us telling the computer ‘Make an AI neural network chip’ and the computer spits out a file used to manufacture the chip,” said Andreas Andreou in a post on the Johns Hopkins website.
Andreou runs the lab where the research was conducted and is a professor of electrical and computer engineering, as well as co-founder of the Center for Language and Speech Processing.
The chip’s network architecture has two layers of artificial neurons. Users can adjust the strength of the connections between the neurons using an 8-bit addressable weight system. The chip can be reconfigured and programmed using a Standard Peripheral Interface subsystem, which was also designed by ChatGPT using natural language prompts.
The researchers validated the design through simulation before sending it to Skywater’s foundry for fabrication. The chip is currently in the process of being printed using a 130-nanometer manufacturing CMOS process.
“Over the last 20 years, the semiconductor industry has made great progress in scaling down the feature size of physical structures on computer chips, enabling more complex designs in the same silicon area,” explained lead researcher Michael Tomlinson in the same post. “The latter advanced computer chips, in turn, support more sophisticated software Computer-Aided Design algorithms and the creation of more advanced computing hardware yielding the exponential growth in computing power that is powering today’s AI revolution.”
The research is available in preprint on arXiv.
