Why Optical Computing Matters for AI?
Today’s AI systems rely on power-hungry electronic processors that face fundamental limits in speed and energy consumption. Optical computing—using light instead of electricity—has long promised a revolutionary alternative, offering ultra-fast, low-power processing ideal for AI tasks like image recognition, autonomous driving, and large language models. However, a major obstacle has persisted: noise and precision limitations in analog optical systems have kept them from practical deployment.
The Breakthrough: Merging the Best of Digital and Analog.
Our team has developed a digital-analog hybrid optical processor (HOP) that achieves 16-bit precision in high-definition image processing—previously thought impossible for optical AI accelerators. Unlike conventional optical systems limited to ~4-bit precision, HOP could enable complex AI tasks (e.g., medical imaging, real-time object detection) without sacrificing speed or energy efficiency. By eliminating the need for power-hungry digital-to-analog converters, HOP could reduce the energy footprint of photonic computing devices.
In tests, HOP could directly process a 16-bit high-definition image with a low pixel error rate, and achieves same accuracy in handwriting recognition (MNIST) as a desktop computer does. The HOP has also been explored to be applied in a YOLO object detection model, where it detects the distant vehicles while the conventional analog optical computing would completely fail —a critical capability for autonomous driving. This reveals the significance of having sufficient numerical precision in real-world complex AI models.
Why This Matters for the Future?
As AI models grow larger and more energy-intensive, the world needs hardware that keeps up. Our work bridges the gap between the speed of light-based computing and the precision of digital electronics, paving the way for practical implementation of photonics for AI applications.
Future works?
We are currently working on scaling the proof-of-concept photonic chip to a larger scale and exploring novel architectures that could achieve higher noise tolerance and numerical precision.
prettypoly•5mo ago
The Breakthrough: Merging the Best of Digital and Analog. Our team has developed a digital-analog hybrid optical processor (HOP) that achieves 16-bit precision in high-definition image processing—previously thought impossible for optical AI accelerators. Unlike conventional optical systems limited to ~4-bit precision, HOP could enable complex AI tasks (e.g., medical imaging, real-time object detection) without sacrificing speed or energy efficiency. By eliminating the need for power-hungry digital-to-analog converters, HOP could reduce the energy footprint of photonic computing devices. In tests, HOP could directly process a 16-bit high-definition image with a low pixel error rate, and achieves same accuracy in handwriting recognition (MNIST) as a desktop computer does. The HOP has also been explored to be applied in a YOLO object detection model, where it detects the distant vehicles while the conventional analog optical computing would completely fail —a critical capability for autonomous driving. This reveals the significance of having sufficient numerical precision in real-world complex AI models.
Why This Matters for the Future? As AI models grow larger and more energy-intensive, the world needs hardware that keeps up. Our work bridges the gap between the speed of light-based computing and the precision of digital electronics, paving the way for practical implementation of photonics for AI applications.
Future works? We are currently working on scaling the proof-of-concept photonic chip to a larger scale and exploring novel architectures that could achieve higher noise tolerance and numerical precision.
Links: European Innovation Council Pathfinder Project DOLORES: https://doloresproject.eu