A US research team has developed a new type of photonic memory computing device

Recently, a research team from the University of California, Santa Barbara has successfully developed a new type of photonic memory computing device that integrates non reciprocal magneto-optical technology. This device achieves high-speed, high-energy efficiency, and ultra-high durability photon computing by utilizing the non reciprocal phase shift phenomenon. The research findings, titled "Integrated non recurrent magneto optics with ultra high endurance for photonic in memory computing," were published in Nature Photonics.

Photon computing has become one of the important directions for the future development of artificial intelligence and machine learning due to its advantages of high speed and low energy consumption. However, the current photon processing architecture faces challenges such as slow storage array update speed, high energy consumption, and insufficient durability. The non reciprocal magneto-optical technology proposed by the research team has successfully solved these bottlenecks by integrating cerium doped yttrium iron garnet with silicon micro ring resonators. By utilizing the non reciprocal phase shift properties of this material, researchers have demonstrated fast programming (1 nanosecond), low energy consumption (143 femjoules per bit), and excellent durability (programmable 2.4 billion cycles) of photonic memory cells.

Figure a. Schematic diagram of computing architecture and unit devices; d. Schematic diagram of memory unit.

The core of this technology is to encode optical weights through the non reciprocal phase shift effect generated by magneto-optical materials in micro ring resonators. Unlike existing photon weights based on thermal or plasmonic dispersion effects, non reciprocal magneto-optical weights not only improve programming speed, but also significantly enhance the device's fatigue resistance and multi-level storage capability. The research team also pointed out that the photon computing platform using this new architecture is expected to provide higher computational efficiency for matrix vector multiplication (MVM) in artificial intelligence.

The photon memory unit demonstrated in this study can update weights at a very high programming speed with high-speed response and low energy consumption, greatly reducing the overall energy burden of the system. Especially in applications such as deep learning that require large-scale computing, this technology can significantly reduce the computational bottleneck of traditional electrical architectures through non-volatile, multi bit storage, further promoting the development of future computing architectures towards more efficient and green directions.

Based on the future development prospects of this technology, researchers believe that by further optimizing the integration of materials, such as utilizing spin orbit torque or spin torque transfer effects, it is possible to achieve higher switching efficiency. In addition, with the advancement of single-chip integration technology between cerium doped yttrium iron garnet and silicon photonic devices, this technology has enormous potential for future applications in fields such as photon computing and magnetic storage.

Source: Opticsky