How advances in photonic memory enable faster optical computing

Advances in technologies such as autonomous driving and computer vision are driving a surge in demand for computing power. Optical computing has attracted widespread attention from academia and industry for its high throughput, high energy efficiency, and low latency. However, current optical computing chips are limited in power consumption and size, which hinders the scalability of optical computing networks. Due to the rise of non-volatile integrated photonics, optical computing devices can achieve in-memory computing with zero static power consumption. Phase change materials (PCMs) have become promising candidates for realizing photonic storage and non-volatile neuromorphic photonic chips. PCMs provide high refractive index contrast between different states and reversible transitions, making them ideal for large-scale non-volatile optical computing chips.

While the promise of non-volatile integrated optical computing chips is enticing, it comes with challenges. The need for frequent and rapid switching, essential for online training, is a hurdle that researchers are determined to overcome. Opening a path toward fast and efficient training is a critical step in unlocking the full potential of photonic computing chips.

Recently, researchers from Zhejiang University, West Lake University and the Institute of Microelectronics of the Chinese Academy of Sciences have made some research breakthroughs. According to Advanced Photonics, they have developed a five-bit photonic memory capable of fast volatile modulation and proposed a solution for non-volatile photonic networks that support fast training. This was achieved by integrating low-loss PCM antimonite (Sb2S3) into a silicon photonic platform.

Optical convolution kernel based on volatile modulation compatible photonic memory

The photonic memory uses the carrier dispersion effect of PIN diodes to achieve volatility modulation, with a fast response time of less than 40 nanoseconds, retaining the stored weight information. After training, the photonic memory uses PIN diodes as microheaters to achieve multi-level reversible phase transitions of Sb2S3, thereby storing the trained weights in the photonic computing network. This leads to an incredibly energy-efficient photonic computing process.

Using the demonstrated photonic memory and working principle, the research team simulated an optical convolution kernel structure. Remarkably, they achieved over 95% accuracy in recognizing the MNIST dataset, demonstrating the feasibility of fast training through volatile modulation and weight storage through 5-bit non-volatile modulation. This work establishes a new paradigm for photonic memory and provides a promising solution for implementing non-volatile devices in fast-trained optical neural networks. With these advances, the future of optical computing will be brighter and more promising than ever.