The American team has developed a new type of optical neural network chip

2024-01-23

Recently, researchers from the University of Pennsylvania in the United States announced the successful development of a powerful new optical chip that can process nearly 2 billion images per second, providing faster and more energy-efficient neural networks for the next generation of deep learning systems.

It is reported that this chip consists of an integrated end-to-end photon deep neural network (PDNN), which does not require components like memory that would slow down traditional computer chips. The magic lies in the fact that it does not process electrical signals, but rather information in the form of light. It uses optical fibers or "waveguides" as neurons, which are stacked in multiple layers, each focusing on a specific type of classification.

In testing, the team created a chip with a surface area of 0.01 square inches (9.3 square millimeters) and used it to classify a set of handwritten characters that looked like letters. After training on relevant datasets, the chip achieved an accuracy of 93.8% for classifying image sets containing two types of characters, and 89.8% for classifying image sets containing four types of characters.

The most impressive thing is that the chip can classify each character in 0.57 nanoseconds and process 1.75 billion images per second. The team stated that this speed advantage stems from the chip's ability to convert information into light for processing, which gives it several advantages over existing computer chips. In summary, this type of chip eliminates four major time-consuming obstacles in traditional computer chips: the conversion of optical signals to electrical signals, converting input data into binary format, large storage modules, and clock based calculations.

Firstly, this type of chip relies on "computation by propagation" to process information, which means that unlike clock based systems, its computation occurs when light propagates through the chip. Secondly, this operation also skips the step of converting optical signals into electrical signals, as its chip can directly read and process optical signals, which brings a technology that greatly accelerates information processing.

In addition, it also has an advantage that the information being processed no longer needs to be stored, which means there is no need to send data to memory and space, and no memory components are needed. Therefore, it saves both time and space, and is more secure, preventing any possible data leakage.

Researchers also point out that it can do more than just classify images. The team stated that they plan to further expand the scale of chip production and adjust technical details and parameters to handle other types of data. They currently understand how to convert numerous data types, such as electronic types (such as images, audio, speech), optical types, etc., while maintaining high-speed processing performance.

Research shows that robotic surgery is safer and reduces patient recovery time by 20%