Researchers at Cornell Tech have created an optical receiver that can directly modify memory using light, potentially offering a more energy-efficient way to update artificial intelligence models in robots and data centers.
The device, presented last month at the IEEE/JSAP Symposium on VLSI Technology and Circuits, uses an array of light similar to a QR code to change binary values in static random-access memory cells. Unlike conventional QR code scanning, where light hits an image sensor as a first step to decoding data, this receiver uses the photocurrents generated by the light beam to directly alter its own memory.
Addressing Memory Bottlenecks
AI processors often lack sufficient on-chip memory to store all the parameters that make up AI models, forcing systems to rely on dynamic random-access memory (DRAM) for additional storage. The electrical connections used to move data between DRAM and the processor create cost and efficiency problems as systems scale up.
"That's one of the major bottlenecks," said Jae-sun Seo, an associate professor of electrical and computer engineering at Cornell Tech in New York City.
Optical links can move data at high bandwidth with less energy loss than metal wires. However, conventional optical receivers undercut that advantage by using power-hungry analog circuits to convert light into electronic bits. The new technology bypasses those analog circuits entirely, enabling fully digital optical communication.
How the System Works
In the new design, DRAM sits with the transmitter while the receiver is integrated into the processor's SRAM. The transmitter beams data to an array of SRAM cells that have been modified to include photodiodes. When light hits each photodiode, it creates a current that flips binary values in the SRAM.
Calibration is necessary because the light source and receiver may not be perfectly aligned or perpendicular. The chip uses a reference data frame containing information about the expected position of each pixel of data to ensure accurate reception.
"Ideally the best way is to have direct, point-to-point space between the transmitter and the receiver," Seo said. "But even if it's slightly tilted, we have this calibration circuit."
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Current Limitations and Future Work
The proof-of-concept transmitter in the lab emits a static 14x14-bit matrix through a metal mask over the light. For real-world applications, the researchers say they need to build an optical transmitter capable of altering the light matrix millions of times per second, transferring gigabits per second. They are working with optics research groups to develop such a transmitter.
Dennis Sylvester, an IEEE Fellow who chairs the University of Michigan's electrical and computer engineering department and was not involved in the work, noted that the technology is likely far from commercialization. The individual photosensitive bit cells are larger than SRAM bit cells in conventional chips, meaning the chip can fit less memory. That trade-off could cancel out the added efficiency of the light-based approach.
"This is a really important problem," Sylvester said. "It's got massive commercial implications. This solution is a clever way of dealing with it."
Seo said the group is working to shrink the bit cells by optimizing transistor and circuit sizes and leveraging CMOS scaling.
Potential Applications in Robotics
Seo and He are exploring uses for the technology in robotics and other edge applications. AI-powered robots in warehouses and factories could use optical data transmission to save time and energy when updating AI models in each robot. Microrobots, which are inherently memory-constrained due to their size, could also benefit from the technology, though it would require a more size-conscious design.
"Edge AI is a big growth area, and in three, four, five years, you're going to hear as much about that as you are with data centers, probably, as the intelligence migrates more and more into these devices that we have," Sylvester said.
The research was conducted by Yifan He, a postdoctoral researcher at Cornell Tech, and Jae-sun Seo.