Until recently, researchers have faced challenges in creating photonic memory for AI applications, often having to choose between crucial attributes such as speed and energy efficiency. An international team has now introduced an innovative approach that overcomes the existing limitations of optical memory, successfully combining features like non-volatility, multibit storage, high switching speed, low energy consumption, and increased endurance within a single framework.
In a groundbreaking achievement, a global team of electrical engineers has unveiled a new technique for photonic in-memory computing, suggesting that optical computing may become viable in the near future.
The collaborative team consists of researchers from the University of Pittsburgh Swanson School of Engineering, the University of California — Santa Barbara, the University of Cagliari, and the Tokyo Institute of Technology (now the Institute of Science Tokyo). Their findings have been published in the journal Nature Photonics today.
This project was coordinated by Nathan Youngblood, an assistant professor of electrical and computer engineering at Pitt, together with Paulo Pintus, formerly of UC Santa Barbara and currently an assistant professor at the University of Cagliari, Italy; and Yuya Shoji, an associate professor at the Institute of Science Tokyo, Japan.
For a long time, researchers have struggled to develop photonic memory suitable for AI processing, often achieving one important parameter, like speed, but at the expense of another, like energy efficiency. In this study, the international team presents a distinctive solution that tackles the current flaws of optical memory by merging features such as non-volatility, multibit storage, high switching speeds, low energy needs, and high endurance on one platform.
“The materials used for these memory cells have been around for decades but were mainly utilized in static optical purposes, like on-chip isolators, instead of in high-performance photonic memory,” Youngblood remarked. “This discovery is a vital enabling technology for creating faster, more efficient, and scalable optical computing architectures that can be directly programmed with CMOS (complementary metal-oxide semiconductor) technology, allowing integration with modern computer systems.”
“Moreover, our technology demonstrated an endurance three times better than existing non-volatile approaches, reaching 2.4 billion switching cycles at nanosecond speeds.”
The authors advocate for a resonance-based photonic architecture that exploits the non-reciprocal phase shift in magneto-optical materials to facilitate photonic in-memory computing.
Typically, photonic processing involves multiplying a fast-changing optical input vector with a matrix of fixed optical weights. However, implementing these weights on-chip using conventional methods and materials has been a complex task. By utilizing magneto-optic memory cells made of integrated cerium-substituted yttrium iron garnet (Ce:YIG) on silicon micro-ring resonators, these cells enable light to travel in both directions—much like sprinters racing in opposite directions on a track.
Computing by Controlling the Speed of Light
“It’s akin to the wind pushing against one sprinter while boosting the other’s speed,” explained Pintus, who spearheaded the experimental work at UC Santa Barbara. “By applying a magnetic field to the memory cells, we can manipulate the speed of light differently based on whether it moves clockwise or counterclockwise around the ring resonator, offering a level of control unattainable with standard non-magnetic materials.”
The team is currently focused on scaling from a single memory cell to a larger memory array that can accommodate even more data for various computing applications. The authors mention that the non-reciprocal magneto-optic memory cell provides an efficient non-volatile storage solution with the potential for limitless read/write endurance and sub-nanosecond programming speeds.
“We also believe that advancements in this technology could explore different mechanisms to enhance switching efficiency,” noted Shoji from Tokyo. “Additionally, newer fabrication techniques using materials beyond Ce:YIG and more precise deposition methods can further enhance the possibilities for non-reciprocal optical computing.”
Other contributors to this project are:
- John E. Bowers, distinguished faculty at the University of California at Santa Barbara
- Mario Dumont, graduate student researcher at the University of California at Santa Barbara
- Duanni Huang, former researcher at the University of California at Santa Barbara
- Galan Moody, faculty at the University of California at Santa Barbara
- Toshiya Murai, researcher at the National Institute of Advanced Industrial Science and Technology, Japan
- Vivswan Shah, graduate student researcher at the University of Pittsburgh
