Optical Technology Enables Updating AI Robots on the Fly

At the IEEE/JSAP VLSI Technology & Circuits Symposium, a new optical receiver design was presented, developed by Cornell Tech postdoctoral researcher Yifan He and Professor Jae-sun Seo. The receiver uses photodiodes embedded in the processor's static random-access memory (SRAM) cells: light from an LED array resembling a QR code generates a photocurrent that directly changes bits in memory. This approach allows transmitting AI model parameters without using analog circuits, which typically consume a lot of energy. In current systems, additional data is stored in dynamic random-access memory (DRAM), and transferring it over electrical interconnects creates an energy efficiency bottleneck. Optical channels offer high bandwidth with lower losses, but existing receivers require energy-hungry analog converters. The new circuit bypasses this problem by operating entirely digitally. So far, it is a 14x14-bit prototype, but the researchers are collaborating with optics groups to build a fast transmitter capable of updating the array millions of times per second. The technology could find applications in edge AI robotics—for example, updating AI models on warehouse robots or in microrobots with limited memory. Dennis Sylvester from the University of Michigan noted the commercial potential of the development but pointed out that the photosensitive cells are currently larger than ordinary SRAM cells, reducing memory density. The team is working on shrinking the cells through transistor optimization and CMOS scaling.