Breakthrough AI convolutional neural network using an optical coprocessor

Optalysys has created convolutional neural network using its optical coprocessor.

The UK-based company has developed optical computing hardware that uses lasers and spatial light modulators (SLMs) to perform complex numerical processing at extremely high speeds and using very little power.

The 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.

In 2015 and 2016, Optalysys had completed a 320 gigaflop prototype and was talking about reaching about 9 petaflops in 2017, 300 petaflops in 2020 and then 17 exaflops in 2022. Optalysys is no longer publishing any statements about future performance targets.

They applied optical processing to the highly complex and computationally demanding area of CNNs with initial accuracy rates of over 70%.

Besides AI, the company has used its technology in a number of other domains, including bioinformatics, weather simulations, and mathematical processing. This application work led to the following proof-of-concept projects:

Project Genesys: delivered a working prototype of a genetic search system for The Earlham Institute (to perform sequence searches of metagenomic reads sequenced from the Human Microbiome Project Mock Community.

Project Escape: a collaboration with the European Centre for Medium-Range Weather Forecasts (ECMWF) and 10 other organizations to develop next-generation weather prediction algorithms.

Project Equate: an investigation undertaken with the US Defense Advanced Research Projects Agency (DARPA) to explore complex mathematical processing that underlies large-scale simulations like plasma modeling and fluid dynamics.

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.

Optalysys was formed to solve some of the fundamental limitations of conventional computing. This comes at a time when the unquenchable thirst for faster processing ever expanding volumes of data comes face to face with the impending demise of Moore’s Law.

Our optical co-processors use low power laser light, rather than electricity, to perform processor intensive convolution functions in parallel at incredibly high speeds and resolutions, yet taking only a fraction of the energy used by electronic processors.

Such functions are the basis of Convolutional Neural Networks (CNNs). CNNs are a burgeoning area of Deep Learning, used to classify and recognise images in applications including autonomous vehicles to recognise and make intelligent judgements on signs or obstacles on the road; in medical image analysis to identify potential causes of a patient’s symptoms so surgeons can make better informed decisions and in weather forecasting to more accurately predict the next natural disaster.

Our technology is hugely scalable – our first systems already offer an order of magnitude performance increase over top end graphics processors. Our upcoming systems will see this capability push beyond the limits of what can be realised through conventional computing, by employing the most naturally parallel processing medium there is – light. Through our unique, patented designs, our systems will be available in the cloud; High Performance Computing (HPC) systems and desktop computers. Miniaturized systems will also be integrated into mobile systems and even hand held devices, using component technologies that are available today.