» Current | 2020 | 2019 | 2018 | Subscribe

Linley Newsletter

Lightmatter Realizes Photonic AI

September 22, 2020

Author: Bob Wheeler

Burning up to 450W, data-center AI chips are reaching thermal limits, and recent process nodes offer little power reduction. In some data centers, power limits now prevent operators from filling racks. Against this backdrop, startup Lightmatter argues new compute elements are necessary to greatly improve energy efficiency, or else data-center throughput will stall. Compute is only half the problem, however, as storing and moving data consumes at least half the power.

Lightmatter is among the small group of photonic-computing startups that have identified machine-learning inference as a strong application fit. Of these companies, only two intend to handle standard neural networks using silicon photonics, and both have their roots in MIT research. Until recently, Lightelligence was more visible, having disclosed a prototype accelerator last year. At last month’s Hot Chips, Lightmatter emerged from stealth by presenting details of its first device. Although Mars is a test chip rather than a product, it includes all the components a production design requires, showing that a complete product could quickly follow.

What immediately stands out is Lightmatter’s holistic approach to building an AI accelerator, whereas related academic research principally focuses on photonics. To build a usable inference chip, the startup developed a silicon-photonic die, a digital die, and an interposer connected in a 3D stack. It also created the software to handle neural networks from the Pytorch and TensorFlow frameworks as well as the ONNX exchange format. The resulting accelerator moves the technology from lab experiment to prototype.

Subscribers can view the full article in the Microprocessor Report.

Subscribe to the Microprocessor Report and always get the full story!

Purchase the full article

Free Newsletter

Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products
Subscribe to our Newsletter »


Linley Fall Processor Conference 2020
October 20-22 and 27-29, 2020 (All Times Pacific)
Virtual Event
More Events »