Machine learning (ML) models today are computationally expensive, requiring thousands of matrix multiplications during training. The most common way to address this is to offload them to a specially designed electronic accelerator. While these are substantially faster than general purpose CPUs, power consumption and heat emissions increase as models scale in complexity. Photonics is an emerging technology that uses light to operate at significantly higher speeds and efficiency than electronics. We designed an interface that demonstrates the viability of using a photonic integrated circuit (provided by GXC) as an ML accelerator. While our prototype is a functional accelerator, future iterations have the potential to increase performance by several orders of magnitude.

Team Members

Cole Choe
Anthony Lam
Ben Marsan
Raymond Ngyuen
Clark Simpler