Breaking the Barrier to Training: A New Optical chip Accelerates Machine Learning

A machine-learning chip with an optical chip

The optical chip can be used to train machine learning hardware. Optica published their research today.

According to a McKinsey report, machine learning applications are now worth $165 billion per year. Before a machine is able to perform intelligent tasks, such as recognizing details in an image or analyzing it, it needs to be trained. The training of artificial intelligence systems such as Tesla’s autopilot, which uses supercomputer-like technology and costs millions in electricity consumption, is expensive.

The gap between the demand for AI and computer hardware is growing due to this \”appetite\”. Photonic integrated chips, also known as optical chips, are a promising solution for delivering higher computing performance. This is measured by TOPS/W, which stands for the number of operations per second. Photonic chips, while they have improved the core operations of machine intelligence for data classification, are still unable to improve front-end machine learning and training.