Today the UK government’s national innovation agency — Innovate UK — has awarded a £1.1m grant to Oxford-based optical computing startup Lumai, another sign that governments are increasingly investing in the strategic tech of the future. The startup says its technology will deliver computation speeds that are “orders of magnitude” faster than anything else out there.
Lumai is building optical computing technology that it says will deliver 10^17 operations per second (that’s a 10 with 17 zeroes after it). For comparison, US graphic cards company NVIDIA’s most powerful graphics processing unit (GPU) does just over 100trn, or 10^14, operations per second.
Lumai’s founder and CEO Tim Weil says that these kinds of speeds will be “transformational” for the AI industry — and that his company plans to sell to both high-performance computing providers and to bigger AI companies that need to train off of large datasets.
How the tech works
Weil explains that Lumai achieves this kind of computing performance by applying optical computing, also known as photonic computing, to as much of the AI computing process as possible.
Photonics is essentially a way of using light to transfer information in computing rather than semiconducting materials like silicon. Other European startups like Spain-based iPronics are working on integrated photonic processors, which apply optical processing to a distinct part of a system, but Lumai says its system uses optical computing across the whole AI workflow.
“You need memory, you need compute, you need training, you need the communication pieces as well. What we're doing is making as many of those elements as possible optical,” Weil explains. “What we're creating is much more advanced computer speed using much lower energy than what GPUs can do at the moment.”
This is significant because massive AI models like OpenAI require a huge amount of computing power: to build its machine learning model GPT-3 the company teamed up with Microsoft to build a supercomputer with 10k GPUs.
As more and more companies start building AI-powered products, there’s going to be a big need for more efficient computing to keep up with demand for resources. Lumai says that its tech will help meet that demand while using less energy than traditional computing techniques.
Weil says that the grant will help hire key talent to bring its optical computing technology to the market, and that he’s confident the team can begin delivering products “within a few years”.
He adds that the backing of the UK government’s innovation agency is evidence that the country wants to be ”at the heart of what we see is going to be a world-leading technology that’s transforming what's going on”.
As well as being applied to the compute-intensive process of training the large language models (LLMs) — which use deep learning algorithms to process vast amounts of text — that power GPT-3 and other generative AI models, Weil says that self-driving technology will be another natural application for its optical computing in the future.
“You have (optical) input coming in — through seeing things. The AI needs to be trained very well to do the right thing. You need instantaneous decisions on the back of huge amounts of data coming in so that you don't crash,” he explains. “It needs to be extremely reliable and low cost — so all the things we're doing.”
It’s early days for Lumai — the company was spun out of Oxford University less than a year ago — and its big claims will take some time to be validated. As it tries to build the infrastructure for AI computation, it’s a small player swimming in a pond full of big fish like NVIDIA, Silicon Valley-based AMD and the UK’s Graphcore.
But optical computing is one of those technologies that could represent a technological step-change — and the UK government wants Lumai to be one of the companies to bring it to the masses.