Home / Electronic-Photonic AI Accelerators
Electronic-Photonic AI Accelerators
Integrated photonics for ultra high-speed machine learning and signal processing.
Neural networks spend most of their time multiplying matrices by vectors. Light is well suited to this: a mesh of interferometers can multiply an optical vector by a matrix as the light passes through, with very low latency and without moving data back and forth to memory. We are exploring electronic-photonic AI accelerators that pair silicon photonics with CMOS electronics for ultra high-speed machine learning and signal processing.
Photonic compute core
Programmable photonic meshes, building on our SiROAP work, hold the weights and perform the multiply-accumulate operations in the optical domain. See Programmable Photonics.
CMOS electronic interface
Drivers, transimpedance amplifiers and data converters feed the optical core and digitize its outputs. Co-designing the photonic and electronic chiplets is a theme shared with our RF photonic channelizer.
Energy-efficient AI
We pair this work with analog in-memory computing and neuromorphic circuits to reduce the energy cost of AI. See In-Memory Computing.
Research Questions
- How do calibration and drift limit the precision of an analog optical matrix multiplier, and how cheaply can electronics correct it?
- What converter resolution and speed does the electronic interface need to keep up with the optical core?
- How should photonic and electronic chiplets be co-packaged for a complete accelerator?