Cornell Tech optical receiver aims to update robot and edge AI memory with light

Cornell Tech researchers have demonstrated an optical receiver that can alter on-chip memory using photocurrents from beamed light arrays, a design aimed at updating AI model parameters without energy-heavy electrical DRAM links, IEEE Spectrum reported. Postdoctoral researcher Yifan He and associate professor Jae-sun Seo presented the approach at the IEEE/JSAP Symposium on VLSI Technology and Circuits. In the lab setup, an LED shines a QR-code-like matrix of light onto a receiver almost a meter away. Photodiodes in modified SRAM cells convert the light into currents that flip binary values, so model parameters could be written optically rather than through conventional metal interconnects. Seo told IEEE Spectrum that AI processors often lack room for full model parameters and rely on external DRAM, making electrical movement of weights a major bottleneck as systems scale. Optical links can move data at high bandwidth with less energy loss, but typical receivers still depend on power-hungry analog circuits. The Cornell design instead targets fully digital optical communication by flashing digital light matrices onto the SRAM array. University of Michigan IEEE Fellow Dennis Sylvester, who was not involved in the work, told IEEE Spectrum the memory bottleneck has massive commercial implications and called the solution a clever approach, while also noting the present photosensitive bit cells are larger than conventional SRAM cells, a density trade-off that could offset efficiency gains until cells shrink. The lab transmitter is a proof of concept that emits a static 14-by-14-bit matrix through a metal mask. The team said it is working with optics groups on a transmitter that can change the matrix millions of times per second. Seo and He pointed to warehouse and factory robots that need coordinated model updates, and to memory-constrained microrobots, as possible edge uses if the hardware matures.