Cornell Tech обновляет AI роботов светом: оптический чип вместо энергозатратных схем

Cornell Tech researchers unveiled in June 2026 an optical receiver that updates AI model parameters directly in chip memory using flashes of light shaped like QR-like matrices — without power-hungry analog circuits. The development was presented at the IEEE/JSAP Symposium on VLSI Technology & Circuits.

How light rewrites chip memory

The receiver directly changes its own memory using photocurrents created by a beam directed at it. In the demonstration, postdoc Yifan He placed the receiver's lens almost a meter from a red LED, and a connected monitor instantly displayed an array of squares resembling a QR code. But unlike a QR code, which hides a simple link, this optical code carries neural network parameters.

In the system, DRAM sits next to the transmitter, while the receiver is built into the processor's SRAM. The transmitter shines light onto an array of SRAM cells that have photodiodes added to them: light hitting each photodiode creates a current and flips binary values in memory. Today's optical receivers lose the advantage of light because of power-hungry analog circuits, but Cornell Tech's new technology accepts digital matrices directly and remains fully digital.

*Authors — postdoc Yifan He and associate professor Jae-sun Seo, Cornell Tech (New York)

*Where presented — IEEE/JSAP Symposium on VLSI Technology & Circuits, June 2026

*Transmitter prototype — a static 14×14-bit matrix through a metal mask

*Goal — millions of matrix updates per second, gigabit-per-second transfer

*Independent expert — Dennis Sylvester, IEEE Fellow, University of Michigan

Why the chip is still far from a product

Photosensitive cells are larger than ordinary SRAM cells, so such a chip fits less memory — and this trade-off could eat up the entire energy-efficiency gain. Dennis Sylvester of the University of Michigan, who was not involved in the work, warns of this. The Cornell Tech team is already working to shrink the cells through transistor optimization and CMOS scaling.

The transmitter prototype shown in the lab is so far only a proof of concept: it emits a static 14×14-bit matrix through a metal mask. Real-world applications will need an optical transmitter capable of changing the matrix millions of times per second and transmitting gigabits per second; the authors are working on this together with optical research groups.

"People are designing all kinds of AI chips," says Jae-sun Seo, associate professor of electrical and computer engineering at Cornell Tech.

According to him, electrical connections between DRAM and the processor create cost and efficiency problems when scaling: "That's one of the major bottlenecks."

Why this matters for robots and edge AI

Seo and He are targeting robotics and other edge applications. One example is warehouses and factories with AI robots: optical data transfer would save time and energy when updating models in each robot. Microrobots, which are tightly constrained on memory because of their size, could also benefit, though they would need an even more compact design.

The independent expert rates the technology's potential highly: according to him, the solution has "massive commercial implications."

"Edge AI is a big growth area, and in three, four, five years it will be talked about almost as much as data centers — as intelligence increasingly migrates into our devices," —

Dennis Sylvester, IEEE Fellow, University of Michigan.

What this means

Delivering data optically straight into chip memory is an attempt to remove one of the major bottlenecks in AI hardware: the costly, power-hungry exchange between DRAM and the processor. A product is still far off, but the direction targets the growing edge AI market, where every watt and every millisecond matters.

Frequently asked questions

How does light change chip memory?

Photodiodes are built into SRAM cells. When light from the transmitter hits them, it generates a current that flips binary values in memory — so the model's parameters are updated directly, without analog converters.

Why isn't the technology ready for sale yet?

Photosensitive cells are larger than standard SRAM cells, so a chip fits less memory. According to Dennis Sylvester of the University of Michigan, this trade-off could cancel out the energy-efficiency gain until the cells are made smaller.

Where could this be applied?

The authors cite warehouses and factories with AI robots, as well as microrobots with limited memory. Optically updating models would save time and energy compared with transferring data over metal wires.