Chinese scientists have created a revolutionary chip with 3600 times more power than its analogs

Researchers from China’s Xinhua University have developed an all-analog ACCEL photonic chip that could be «revolutionary for artificial intelligence and machine vision technologies.» They described their achievement in an article published in the journal Nature.

The new chip does not use analog-to-digital converters, but directly sends encoded light waves to an analog computing chip to convert the image. Scientists note that this is very important for solving machine vision problems, such as autopilots and autonomous robots. The device uses electronic components to manipulate continuous physical quantities.

The chip has a system energy efficiency of 74.8 peta operations per second per watt and a computing speed of 4.6 peta operations per second. This is more than three times higher than that of current high-performance GPUs. Thanks to the fusion of optoelectronic computing and adaptive learning, ACCEL is very good at distinguishing between different objects in images — the accuracy was 85.5% for Fashion-MNIST tasks, 82% for ImageNet 3-grade classification, and 92.6% for frame-by-frame video recognition tasks.

*comparing the speed and efficiency of the ACCEL chip with other technologies

«Regardless of the way the number of operations is calculated, fully analog ACCEL experimentally reduces system latency and power consumption by a factor of several compared to digital neural networks on a modern GPU (NVIDIA A100) while achieving the same accuracy in practice,» the scientists noted.

According to the developers, the chip has shown high reliability even in low light, making it suitable for use in heavy and light industry, mobile devices, and autopilots. Low power consumption reduces heat generation, which in the future will make it possible to make the device much smaller without losing power.