Researchers at Cornell Tech have developed an optical receiver prototype that directly alters its own memory using light-based photocurrents, offering a potential path to ease mounting memory bottlenecks in artificial intelligence processors. Unveiled at the IEEE/JSAP Symposium on VLSI Technology & Circuits, the experimental design bypasses power-hungry analog circuits by receiving rapid flashes of digital, QR code-like light matrices to update AI model parameters on the fly.
As modern artificial intelligence models expand in size, hardware architectures face intense physical constraints. Processors rely on static random-access memory (SRAM) for rapid execution, but this built-in memory lacks the capacity to hold entire AI models independently. Consequently, systems store additional parameters in dynamic random-access memory (DRAM). Moving this volume of data between DRAM and the processor via traditional metal wires introduces energy losses and efficiency concerns as systems scale.
Optical links offer a high-bandwidth alternative to metal wires, but conventional optical receivers rely on power-hungry analog circuitry to translate light signals back into electronic bits. The Cornell Tech team’s approach aims to eliminate that conversion step. In their experimental setup, a transmitter beams data directly to an array of SRAM cells modified with photodiodes. When light strikes these individual photodiodes, it generates a current that flips binary values within the SRAM without intermediate analog translation.
“People are designing all sorts of different AI chips,” says Jae-sun Seo, an associate professor of electrical and computer engineering at Cornell Tech in New York City. Seo notes that managing data transfer between storage and processors remains one of the major bottlenecks in contemporary hardware design.
Overcoming Alignment and Scaling Hurdles
Implementing light-based memory updates outside a controlled laboratory environment presents engineering hurdles, notably maintaining accurate alignment between transmitters and receivers. Because physical components cannot guarantee perfect perpendicular alignment, the Cornell-developed chip incorporates a calibration circuit. The processor references a data frame containing positional information for each pixel of incoming data, ensuring reliable reception even if the optical link is slightly tilted.
“Ideally the best way is to have direct, point-to-point space between the transmitter and the receiver, but even if it’s slightly tilted, we have this calibration circuit,” Seo explains.
In its current state of development, the hardware functions as a proof-of-concept. During lab demonstrations, postdoctoral researcher Yifan He positioned an optical receiver almost a meter away from an LED emitting a red light beam through a static 14×14-bit metal mask matrix. To achieve practical implementation in real-world settings, the research team notes they will need to build an optical transmitter capable of shifting light matrices millions of times per second to sustain gigabit-per-second transfer rates.
Commercial Implications and Edge AI Applications
Industry observers note that while the technology targets a critical scaling problem, significant hurdles remain before commercialization. Dennis Sylvester, an IEEE Fellow and chair of the electrical and computer engineering department at the University of Michigan who was not involved in the project, describes the concept as a clever solution with massive commercial implications. However, Sylvester points out that the current photosensitive bit cells are larger than conventional SRAM cells, meaning chips built with this architecture would fit less memory, a trade-off that could cancel out the added efficiency of the light-based approach.
To address this trade-off, Seo’s team is working to shrink the bit cell footprint by optimizing transistor and circuit sizing and utilizing CMOS scaling techniques.
Looking ahead, researchers are targeting edge computing applications where power efficiency is paramount. AI-powered robots operating in automated warehouses and factories could potentially use optical data transmission to save time and energy when updating AI models. Microrobots, which are inherently memory-constrained due to their size, could also benefit from the tech if future iterations achieve 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 says.
Further developments from the research group are expected as they refine transmitter speeds and cell scaling. Readers interested in following updates on semiconductor design and optical computing research can consult proceedings from the IEEE/JSAP Symposium on VLSI Technology & Circuits or institutional updates from Cornell Tech.
Related reading