Hybrid photonic/electronic computing

Modern computing operations, especially artificial intelligence and machine learning applications, are notoriously resource intensive, both in terms of computing power and electrical energy demands. Physical limitations on the ability to add more transistors to microchips are preventing hardware from keeping pace with the increasing demands. An MIT team featuring ISN-affiliated faculty member Dirk Englund and members of his research group has demonstrated the potential of a photonics-enhanced electronic network interface card (NIC) dubbed a Lightning that seamlessly pairs the strengths of photonics and electronics through the use of a reconfigurable count-action abstraction, allowing for sustained activity rather than the stop-and-go processing of previous systems.
Image
A computer chip featuring a lightning bolt, symbolizing MIT's Lightning smartNIC for high-speed neural network processing.

MIT researchers introduce Lightning, a reconfigurable photonic-electronic smartNIC that serves real-time deep neural network inference requests at 100 Gbps. Image: Alex Shipps/MIT CSAIL via Midjourney

Image
An assembled Lightning developer kit.

An assembled Lightning developer kit. Image: courtesy of the researchers.