In Memory Photonic Computing for AI
An innovative approach to processing artificial intelligence in-memory for photonic systems has been explored by an international team of researchers. This team, co-led by experts from the University of Pittsburgh and the University of California at Santa Barbara, has developed a groundbreaking method for encoding optical weights that can drastically enhance in-memory photonic computing. Their pioneering technique leverages magneto-optic memory cells, which are constructed from cerium-substituted yttrium iron garnet (Ce:YIG) on silicon micro-ring resonators, to deliver superior speed, efficiency, and robustness in on-chip optical processing.
Traditionally, photonic processing involves the multiplication of a rapidly fluctuating optical input vector with a fixed matrix of optical weights. Despite its promise, encoding these weights directly on-chip using conventional materials and methods has posed significant difficulties. However, the new approach using Ce:YIG integrated on silicon micro-ring resonators addresses these challenges by enabling bidirectional propagation of light, thus facilitating more efficient processing.
“The materials used in these innovative cells have been accessible for decades, primarily serving static optical functions like on-chip isolators,” explained Nathan Youngblood from the University of Pittsburgh, the leader of the team. “Yet, this discovery marks a turning point, creating a platform for high-performance photonic memory, paving the way for faster, more efficient, and highly scalable optical computing architectures. Moreover, our technology seamlessly integrates with CMOS (complementary metal-oxide semiconductor) circuitry, making it compatible with today’s computer technology.”
Youngblood also noted the remarkable endurance of the technology, asserting that it boasts three orders of magnitude superior durability compared to other non-volatile approaches, offering 2.4 billion switching cycles with nanosecond speeds.
The research offers a substantial advantage by applying a magnetic field to the memory cells. This field can modulate the speed of light, contingent upon whether light is traveling clockwise or counterclockwise around the ring resonator, introducing a level of control that surpasses standard non-magnetic materials. Paulo Pintus, who led the experimental work at UC Santa Barbara and is now affiliated with the University of Cagliari in Italy, emphasized the unique control provided by this technique.
Currently, the team is focused on transitioning from a single memory cell to a large-scale memory array capable of supporting a broader array of data for computing applications. This scalable magneto-optic memory cell presents an efficient non-volatile storage solution, offering high read/write endurance and ultra-fast programming speeds at sub-nanosecond levels.
The collaborative effort includes esteemed researchers from AIST and the Tokyo Institute of Technology in Japan, who are instrumental in enhancing the technology’s potential. Yuya Shoji from Tokyo highlighted the future prospects, suggesting that upcoming advancements could leverage different effects to amplify switching efficiency. Shoji also pointed to the promise of new fabrication techniques and alternative materials, which, in conjunction with more precise deposition methods, could significantly elevate the capabilities of non-reciprocal optical computing.
As this revolutionary technology continues to mature, it holds promise for transforming AI computing by ensuring rapid, reliable, and efficient processing capabilities at the photonic level. With projects underway to scale the memory arrays, we can anticipate substantial contributions to the evolving landscape of computing technologies.
