SoylentNews

takyon [soylentnews.org] writes:

Optalysys claims to have implemented [top500.org] a convolutional neural network [wikipedia.org] (CNN) using its optical supercomputer/coprocessor:

Optalysys says they have built the world's first implementation of a convolutional neural network using its optical coprocessor.

The UK-based company has developed optical computing hardware that uses lasers and spatial light modulators [wikipedia.org] (SLMs) to perform complex numerical processing at extremely high speeds and using very little power. The technology has been implemented as an HPC coprocessor, which is meant to be attached to a conventional computer – a desktop system or a server. The company's first prototype was announced in 2014.

Although not much detail has been provided on this latest application of the technology, Optalysys says they have implemented a CNN based on the MNIST dataset of hand-drawn numerals. The dataset is comprised of 60,000 training characters and 10,000 testing characters. According to the company, its optical laser technology enables them to process this model several orders of magnitude faster than conventional electronic hardware and does so at a fraction of the energy consumption.

[...] Optalysys is one of a growing number of startups that is applying optical computing to AI. Those companies include Lighton, Light Intelligence, Fathom Computing, and Lightmatter. (We covered that latter two here [top500.org] and here [top500.org].) All are using various forms of optical technology to encode neural networks and are promising huge speedups and much better energy efficiency than traditional CPU/GPU-based machine learning.

MNIST database [wikipedia.org]. Press release [hpcwire.com].

Previously: Computing With Lasers Could Power Up Genomics and AI [soylentnews.org]
Optalysys, Back in the (Press Release) News [soylentnews.org]

Original Submission