A research team at the University of California, Santa Barbara has developed a photonic platform for in-memory computing that addresses challenges such as slow switching speeds and programmability limitations. Their findings are published in the journal Nature Photonics .
For decades, the circuits that run computers and smartphones have been shrinking in size while improving in power. But Moore’s Law is approaching its physical limits, as the maximum number of transistors on a chip and the heat generated by densely packed transistors limit further performance improvements.
This stability of computing power poses a challenge for data-intensive applications such as machine learning and artificial intelligence, which require ever-increasing processing power. To solve this problem, new technologies are needed. Photonics, which consumes less power and has less latency than electronics, is a promising solution.
One of the most promising approaches is in-memory computing, which relies on photonic memories that transmit optical signals to perform high-speed operations. But efforts to develop such memories have so far been limited by slow switching speeds and limited programmability.
The advance in this field was made possible through a collaboration led by Paolo Pintus, assistant professor at the University of Cagliari, John Bowers, professor of Electrical and Computer Engineering (ECE) at the University of California, Santa Barbara, and Garan Moody, associate professor of ECE at the University of California, Santa Barbara. The project also involved Nathan Youngblood of the University of Pittsburgh, Yuya Shoji of Tokyo University of Science, and Mario Dumont, a postdoctoral researcher in the Bowers lab.
The researchers used yttrium-substituted cerium iron garnet (YIG), a magneto-optical material whose optical properties can be dynamically affected by an external magnetic field. The team developed a new type of magneto-optical memory that uses tiny magnets to store information and control the transmission of light.
The platform uses light to perform calculations much faster and more efficiently than traditional electronics, and the new memory achieves switching speeds 100 times faster than state-of-the-art optical integration technologies, can be reprogrammed for different tasks, and consumes about 10 times less power.
Moreover, the team demonstrated that the magneto-optical memory can be rewritten more than 2.3 billion times, giving it a virtually infinite lifespan: in contrast, current optical memories typically only withstand a maximum of 1,000 write cycles.
These unique magneto-optical materials make it possible to control the propagation of light using an external magnetic field. In this project, an electric current is used to program the micro-magnets and store data. The magnets control the propagation of light inside the Ce:YIG material, allowing it to perform complex operations such as matrix-vector multiplication, which is at the core of any neural network .
Paolo Pintas, project scientist, University of California, Santa Barbara
The authors say that their results represent a major advancement in optical computing technology that may enable practical applications in the near future.
Journal References:
Pintus, P., et al . (2024) Integrated nonreciprocal magneto-optics with ultra-high durability for photonic memory computing. Nature Photonics . doi.org/10.1038/s41566-024-01549-1.
sauce:
University of California, Santa Barbara
