Magneto-Optics Increases Photonic Processing Efficiency for AI

A platform for in-memory photonic computing using magneto-optic material has the potential to offer efficient, nonvolatile data processing with unlimited read/write capabilities and sub-nanosecond programming speeds. The magneto-optic memory platform was developed by researchers from the University of California, Santa Barbara (UCSB) and several international institutions. It addresses multiple limitations of existing photonic processing techniques.

Data-intensive applications like AI require ever greater computational power, while at the same time, the computing capacity of digital hardware has begun to flatten. Optical solutions use less energy and process data faster than traditional, electronic-based processing techniques.

In-memory computing, which uses photonic memories, allows data processing operations to be performed almost instantaneously. But current techniques for creating photonic memories have yet to achieve nonvolatility, multibit storage, high switching speed, low switching energy, and high endurance in a single platform.

The researchers developed a resonance-based photonic architecture that leverages the nonreciprocal phase shift in magneto-optic material. They used cerium-substituted yttrium iron garnet (Ce:YIG), a magneto-optic material whose optical properties change in response to external magnetic fields, on silicon microresonators. They used magnets to store data and control the propagation of light within the magneto-optic material.

A conceptual illustration of a photonic memory array. Courtesy of UCSB/Brian Long.

“These unique magneto-optical materials make it possible to use an external magnetic field to control the propagation of light through them,” researcher Paolo Pintus, an assistant professor at the University of Cagliari, said. “In this project, we use an electrical current to program micromagnets and store data.”

The researchers excited both the clockwise and counterclockwise modes of the microring resonator with a magneto-optic cladding layer of Ce:YIG. The interaction between the optical mode and the Ce:YIG material induced a nonreciprocal phase shift for the two counter-propagating modes, which appeared as a split resonance shift with opposite signs, dependent on the direction and strength of an applied magnetic field.

This approach to photonic in-memory computing was found to enable programming speeds of about 1 GHz as well as multilevel coding. The magneto-optic memory cells can be reprogrammed multiple times to perform different tasks.

The magneto-optic memory platform demonstrated switching speeds 100 times faster than the switching speeds of integrated photonic technology, while consuming about one-tenth the power of traditional photonic processing techniques.

The team demonstrated that magneto-optic memories could be rewritten more than 2.3 billion times, indicating a potentially unlimited lifespan. Existing optical memories, with their limited lifespans, can be rewritten up to 1000 times.

The magneto-optical memory platform leverages the properties of light to perform calculations at significantly higher speeds and with greater efficiency than can be achieved by using traditional electronics. “The magnets control the propagation of light within the Ce:YIG material, allowing us to perform complex operations, such as matrix-vector multiplication, which lies at the core of any neural network,” Pintus said.

In the field of deep learning, the computation required to train deep neural networks grew by over 300,000 times between 2015 and 2020, while the efficiency of graphics processing units grew by only 300-fold.

Several different photonic architectures have been proposed to address this bottleneck. The 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 an array of photonic memory cells, is limited by material- and device-level issues.

The new approach to encoding optical weights for in-memory photonic computing, using magneto-optic memory cells, offers several key advantages over existing architectures. By leveraging the nonreciprocal phase shift in magneto-optic materials, the researchers have achieved a platform for on-chip optical processing that is fast (1 nanosecond), efficient (143 femtojoule per bit), and robust (2.4 billion programming cycles).

The research was published in Nature Photonics (www.doi.org/10.1038/s41566-024-01549-1).