In memory photonic computing for AI
An international team of researchers has developed a way to process AI in memory for photonic systems.
The team, led jointly by the University of Pittsburgh and the University of California at Santa Barbara, developed the new approach to encoding optical weights for in-memory photonic computing using magneto-optic memory cells. These cells were built from cerium-substituted yttrium iron garnet (Ce:YIG) on silicon micro-ring resonators and showed key advantages with fast (1 ns), efficient (143 fJ per bit) and robust (2.4 billion programming cycles) on-chip optical processing.
- $6m for in-fibre photonic AI
- Fraunhofer, GF team for new memory technologies
- A rewriteable photonic chip with phase change memory
A typical approach to photonic processing is to multiply a rapidly changing optical input vector with a matrix of fixed optical weights. However, encoding these weights on-chip using traditional methods and materials has proven challenging. By using magneto-optic memory cells comprised of heterogeneously integrated cerium-substituted yttrium iron garnet (Ce:YIG) on silicon micro-ring resonators, the cells cause light to propagate bidirectionally.
“The materials we use in developing these cells have been available for decades. However, they have primarily been used for static optical applications, such as on-chip isolators rather than a platform for high performance photonic memory,” said team leader Nathan Youngblood at Pittsburgh. “This discovery is a key enabling technology toward a faster, more efficient, and more scalable optical computing architecture that can be directly programmed with CMOS (complementary metal-oxide semiconductor) circuitry – which means it can be integrated into today’s computer technology.
“Additionally, our technology showed three orders of magnitude better endurance than other non-volatile approaches, with 2.4 billion switching cycles and nanosecond speeds.”
“By applying a magnetic field to the memory cells, we can control the speed of light differently depending on whether the light is flowing clockwise or counterclockwise around the ring resonator. This provides an additional level of control not possible in more conventional non-magnetic materials,” said Paulo Pintus, who led the experimental work at UC Santa Barbara and is now at the University of Cagliari, Italy.
The team is now working to scale up from a single memory cell to a large-scale memory array which can support even more data for computing applications. The non-reciprocal magneto-optic memory cell offers an efficient non-volatile storage solution that could provide high read/write endurance at sub-nanosecond programming speeds.
The team included researchers from AIST and Tokyo Institute of Technology in Japan.
“We also believe that future advances of this technology could use different effects to improve the switching efficiency,” said Yuya Shoji at Tokyo, “and that new fabrication techniques with materials other than Ce:YIG and more precise deposition can further advance the potential of non-reciprocal optical computing.”
The results were published in Nature Photonics
doi: 10.1038/s41566-024-01549-1; www.pittsburgh.edu
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
