Tsinghua University scientists have developed a light-driven AI chip, OFE², utilizing photonic circuits to process data at 12.5 GHz.
New optical computing chip enables ultra-fast, low-latency parallel processing for quantitative trading.
Click Here to Add Gadgets360 As A Trusted Source
Artificial Intelligence (AI) systems involve various technologies, such as image recognition and language translation, and they require a huge amount of energy. Scientists at Tsinghua University have developed an AI chip powered by light that can handle data at 12.5 GHz, which is an unprecedented speed for optical computing. Essentially, it is less energy-consuming than a regular electronic device because it completes complicated pattern-recognition tasks by directing light beams through tiny on-chip structures. Eventually, this breakthrough might free AI from being the cause of energy shortages while still being able to analyse images and data.
Optical Feature Extraction Engine (OFE²)
According to the paper published in Advanced Photonics Nexus, the Tsinghua team's device, called an Optical Feature Extraction Engine (OFE²), uses photonic circuits to split input data into parallel light channels. The synchronised beams pass through a tiny diffraction plate etched on the chip, performing a matrix-vector multiplication as the light waves interfere. In tests, OFE² clocked 12.5 GHz, completing one such computation in just 250.5 picoseconds – the fastest result yet for an optical processor.
Applications and Impact
In imaging tests, OFE² extracted edge features that improved classification (for example, identifying organs in medical scans). It was also able to make decisions on buy/sell in microseconds, and it was fast enough to act as the speed of light since it processed live financial data. Improvements in other light-based AI chips are similar: one design with etched Fresnel lenses reduced the power requirement of image-convolution operations by 10-100x. Scientists believe that by moving some of the important calculations in photonics, a new generation of real-time, low-energy AI will be initiated.
