ChatGPT4 And Hopkins Engineers Work To Design Brain-Inspired Chips

The electrical and computer engineers at Johns Hopkins are leading the way in developing a novel method for producing neural network chips, or neuromorphic accelerators, which have the potential to drive energy-efficient, real-time machine intelligence for the next generation of embodied systems, such as robots and autonomous cars.

Michael Tomlinson, a graduate student in electrical and computer engineering, and Joe Li, an undergraduate, both work at the Andreou Lab. They used ChatGPT4 and natural language prompts to generate comprehensive instructions for building a spiking neural network device, which functions similarly to the human brain. By following instructions to ChatGPT4, they were able to produce a complete chip design that could be manufactured. They began by simulating a single biological neuron and then connected more to make a network.

“This is the first artificial intelligence chip created by a computer employing natural language processing. “It’s like instructing a computer to “create an artificial intelligence neural network chip,” and the computer produces the necessary files to build the chip,” explained Andreas Andreou, an electrical and computer engineering professor, co-founder of the Centre for Language and Speech Processing, and member of the Kavli Neuroscience Discovery Institute and the newly established Data Science and AI Institute at Johns Hopkins University.

The project was started during last summer’s NSF-funded 2023 Neuromorphic Cognition Engineering Workshop. It can be found at arXiv, the preprint site.

The resulting network architecture on the chip resembles a tiny silicon brain with two layers of networked neurons. Through the use of an 8-bit addressable weight system, the user can modify the strength of these connections, enabling the chip to create learning weights that control its behaviour and functionality. A user-friendly interface similar to a remote control is termed the Standard Peripheral Interface (SPI) sub-system, which is used for reconfiguration and programmability. ChatGPT also created this SPI sub-system with natural language prompts.

As a proof of concept, Tomlinson clarified, they created a straightforward neural network chip with straightforward coding. The team used extensive software simulations for validation before submitting the chip for manufacture. This allowed them to make sure the final design would function as planned and allowed them to iterate on the design and fix any problems.

The completed design was electronically sent to the chip fabrication company Skywater “foundry,” where it is presently being “printed” via a reasonably affordable 130 nm CMOS manufacturing method.

“While this is just a small step towards large-scale automatically synthesised practical hardware AI systems, it demonstrates that AI can be employed to create advanced AI hardware systems that in turn would help accelerate AI technology development and deployment,” Tomlinson stated. “The semiconductor industry has made significant strides in the last 20 years to reduce the physical structure of computer chips’ feature sizes, allowing for more sophisticated designs within the same silicon space. The later, more powerful computer chips enabled the development of more sophisticated hardware and software, including Computer-Aided Design algorithms, which in turn produced the exponential rise in processing power that is driving the AI revolution of today.”