Shining Light onto a Chip Could Give Robots a Faster AI Brain Update
Cornell Tech researchers have built a receiver that rewrites a chip's memory using flashes of light instead of electrical wires, and it could one day let warehouse robots refresh their AI models on the fly.
Key points
- Cornell Tech researchers presented a light-based memory receiver at the IEEE/JSAP Symposium on VLSI Technology and Circuits last month.
- The device rewrites a chip's built-in memory by shining QR-code-like patterns of red light onto photosensitive cells, cutting out the power-hungry circuits traditional optical systems need.
- Dennis Sylvester of the University of Michigan called the approach "a clever way" of tackling one of AI hardware's biggest bottlenecks.
- The team says the technology could reduce energy use in data centres, self-driving cars and AI-powered robots.
- Photosensitive memory cells that are larger than standard SRAM cells remain the main obstacle before the design could appear in commercial hardware.
Imagine your robot assistant needing a software update. Right now, getting new AI instructions into a chip means shuffling data along copper wires, burning energy. A team at Cornell Tech thinks light can do that job better.
Postdoctoral researcher Yifan He and associate professor Jae-sun Seo built a receiver that sits inside a chip and rewrites its own memory when a beam of light hits it. The beam carries a pattern of bright and dark squares resembling a QR code. Each square switches a tiny memory cell on or off, encoding the numbers called parameters that tell an AI model how to behave.
Why does moving data with wires cost so much energy?
AI models are enormous collections of numbers. A chip's built-in memory, called SRAM (static random-access memory, the fast but small storage baked into a processor), can't hold them all, so the extra data lives in a separate, larger store called DRAM (dynamic random-access memory) connected by metal wires. Shuffling numbers across those wires uses a surprising amount of electricity, and it gets worse as AI systems grow.
Optical links carry data as light rather than electrical current and can move more data with less energy loss. Today's light-receiving circuits still need power-hungry analogue components to convert light pulses into digital ones and zeroes. The Cornell design skips that step entirely: light hits a photodiode, a tiny component that turns light into electrical current, built directly into each memory cell, and that current flips the cell's stored value straight from zero to one. No analogue middleman.
"That's one of the major bottlenecks," Seo told IEEE Spectrum, which first reported the research. Sylvester, an IEEE Fellow who chairs the University of Michigan's electrical and computer engineering department, called it "a really important problem" with "massive commercial implications."
Our earlier piece on the electricity problem at the heart of modern AI from 17 July 2026 laid out why every watt saved inside a data centre compounds quickly. This research is the same battle fought at the chip level.
What does this mean for robots and everyday gadgets?
Warehouse robots could get updated AI instructions without being wired to a central server, saving time and energy across every unit on the floor. Tiny microrobots, barely larger than a fingernail, could eventually carry useful AI smarts without the heavy memory hardware that currently makes that impractical.
Sylvester says the technology is far from shop shelves. The photosensitive memory cells are bigger than standard SRAM cells, so a chip using them holds less memory, a trade-off he says could cancel out the efficiency gains entirely.
Seo's team is shrinking the cells by optimising transistor sizes. They're also partnering with optics groups to build a transmitter that can flash the light pattern millions of times per second, reaching the gigabits-per-second speeds real applications need. The prototype He demonstrated in the lab emits a single, fixed 14-by-14-square pattern through a metal stencil.
How soon could this show up in real products?
Three to five years is Sylvester's honest window for edge AI broadly. This specific technology needs more work before it can match conventional chips. The current prototype proves the idea works. That's all it proves, and that's enough for now.
