Programmable Photonic Latch Could Deliver the Performance, Efficiency Gains Necessary for AI Growth

Researchers at Nokia Bell Labs have developed a programmable photonic latch — a component that, they hope, will be integral to delivering improved performance and efficiency for machine learning and artificial intelligence (ML and AI) systems.

"While optical communications and computing have seen significant progress over the past decades, data storage has been predominantly implemented using electronic memory," explains author Farshid Ashtiani of the work. "Having a fast optical memory that can be used with optical processing systems, as well as other optical systems used in communications or sensing, would make them more efficient in terms of energy and throughput."

"Large language models like ChatGPT rely on massive amounts of simple mathematical operations, such as multiplication and addition, performed iteratively to learn and generate answers," Ashtiani continues. "Our memory technology could store and retrieve data for such systems at high speeds, enabling much faster operations. While a commercial optical computer is still a distant goal, our high-speed optical memory technology is a step toward this future."

The device developed by Ashtiani and colleagues is known as an integrated photonic latch, based on the set-reset latch in traditional electronics and capable of switching between set and reset states based on its inputs — storing a single bit of data per latch. Rather than electronics, though, the new part is based on optics and is theoretically capable of operating at speeds far in excess of those achievable using existing electronic parts.

"Because each memory unit has an independent input light source, it is possible to have several memory units working independently without affecting each other through optical power loss propagation," Ashtiani adds of the design's benefits. "The memory units can also be co-designed with the existing silicon photonic systems and be built reliably and with very high yields."

The team's prototype has been tested to picosecond response times, for the storage of multiple bits per component, and in the presence of fluctuations in power and input — though all using a programmable photonic platform, a lab-based precursor to producing a dedicated component. The next step, the researchers say, is to attempt to scale the technology, fabricate dedicated parts, and find a way to integrate both the silicon photonics and the control electronics into a single manufacturing step.

The project is documented in a paper published in the journal Optics Express, under open-access terms.