Take back CTRL
7 min read
Britain’s AI future will soon be found… in a Bristol car park. James O’Malley reports
If the last year has proven anything, it is that artificial intelligence (AI) is the real deal. Forget cryptocurrencies, virtual reality and other technological baubles. As OpenAI’s ChatGPT has so clearly demonstrated, future generations may look back on 2023 and 2024 as a pivotal moment in the history of technology. And the race to seize the opportunities presented by AI has already begun.
“Just think for a moment about what that will mean for our country,” Rishi Sunak said in a speech last October to the Royal Society, just days before the start of the UK’s big AI safety summit. “The growth it will catalyse, the jobs it will create, the change it can deliver – for the better.”
The prime minister admits he is “unashamedly optimistic about the power of technology to make life better for everyone”, but what exactly will this glittering future look like? Should we expect whizz-bang robots, lasers and pyrotechnics to accompany such a dramatic step-change in digital capabilities?
The reality is perhaps slightly less eye-catching, as it turns out the future will soon be found inside a liquid-cooled, 40ft shipping container, placed in the car park at the National Composites Centre in Bristol.
This container is set to be the home of Isambard-AI, a new £225m supercomputer that, along with the Dawn supercomputer in Cambridge, will form the National AI Research Resource. The hope is that, together, they will bolster Britain’s position as an AI power-player, and the government hopes it will keep Britain at the forefront of the emerging technology.
“We’re looking to solve the biggest, most challenging societal and scientific problems in the world,” says Simon Appleby, AI business lead at Hewlett Packard Enterprise, the company tasked with delivering the new supercomputer.
To explain this, Appleby points to fields such as weather and climate modelling, fusion energy, pharmaceuticals, healthcare, and aerospace engineering – industries that could all benefit from an uplift in computing power to crunch through huge amounts of data.
But exactly how much power? Isambard-AI will reportedly eventually make use of 5,500 GPUs – the computer chips used to perform the technical calculations needed to train AI models – to give scientists up to 21 exaflops of computing power, or ‘compute’, as techie types call it. The ‘Grace Hopper’ chips, manufactured by US chipmaker Nvidia and named for the American computing pioneer, will together be capable of around 21 quintillion mathematical operations per second – about a billion times more powerful than a standard desktop PC.
The system’s origins can be traced back to March last year and the recommendations made by Professor Zoubin Ghahramani from Cambridge University, in an independent review of the “Future of Compute”. Ghahramani, who is also a senior research director at Google Brain, found our current supercomputers lacking.
“Existing compute capabilities are not fit to serve all users, particularly AI users, and are falling behind those of other advanced economies,” he wrote, pointing out that at the end of 2022, the UK only had a 1.3 per cent share of global compute capacity – and recommended the immediate creation of “continuous and sustained investment in accelerator-driven compute capabilities”.
Isambard-AI, then, is the government’s response, with the Department for Science, Innovation and Technology announcing last August that it had helped the UK Research and Innovation agency pay for £100m-worth of chips for the new supercomputer.
“The main takeaway was that we needed to rapidly increase our compute capability in the UK, both for AI and classical compute capabilities,” one policy official told The House.
The plan is that once Isambard-AI is up and running, scientists from the public and the private sectors will be able to apply for time to run their code on the system, following similar practices in other shared research facilities.
This means that, day to day, Isambard could be doing anything from modelling the climate, to training large language models that form the underlying technology inside ChatGPT and other generative AI tools.
We also understand that the new UK AI Safety Institute, announced at the AI safety summit held at Bletchley Park last year, will also be involved in the approval process, and will have the opportunity to examine code run on the system – though the specifics on this have not yet been worked out.
However, there is still the obvious question of why? Impressive as Isambard-AI is, won’t the likes of Amazon Web Services and Google Cloud inevitably offer similar capabilities, conceivably with even more GPUs to throw at computing problems?
Having a sovereign domestic capability that we are able to access whenever we want is a big strategic advantage for the UK
“AI is becoming the defining technology of the future, underpinning all other technologies and progress across various different sectors,” says the policy official. “Having a sovereign domestic capability that we are able to access whenever we want is a big strategic advantage for the UK.”
And this rationale makes sense. A domestic resource means Britain’s AI capabilities won’t be at the mercy of foreign governments and corporations. And it will make it easier to apply AI capabilities to sensitive datasets, such as health data, which may be legally restricted from leaving the country or being shared more widely. Plus, as Appleby notes, Isambard-AI is about more than just the raw power – it’s about the ecosystem of scientists and people around the facility supporting its work too.
Perhaps the most surprising thing for a significant project is how quickly it is getting off the ground. The first test phase of Isambard-AI, making use of just 168 of the 5,500 GPUs, should be online around March – with the rest of the system following in July and August. Why so fast?
“If you look at most data centres, they’re looking at three to four years, and the overheads and the costs of that are extreme,” says Appleby, who argues that it is the modular nature of Isambard-AI that makes it quick to turnaround, as the system can be assembled remotely and then transported to its final location, like a shipping container.
Because of the modular nature of the system it’s conceivable that, in the future, capacity could be increased by simply shipping in another container full of GPUs. The only real limitation is the availability of electricity to power the energy-hungry system – as well as, of course, money.
“If we did three or four of these alongside each other, on the same park, which is maybe one of the ambitions the government has in the future, we could easily get to 20,000 to 25,000 GPUs at Isambard-AI,” says Appleby.
When Isambard-AI comes online then, it will be a big moment. In an instant, it will be within the top 10 AI facilities of its type in the world. And it will be closely competing with similar facilities in Germany and Switzerland, as well as more broadly with systems in Japan, China and Dubai.
And that’s perhaps why, even before Isambard-AI is finished, there are more supercomputers on the horizon. In last year’s Autumn Statement, Jeremy Hunt announced an additional £500m investment over the next two years to help make Britain an “AI powerhouse”, as the Chancellor put it. And there are already plans for another Bristol supercomputer, Isambard 3, which will offer scientists more conventional CPU compute power, to complement Isambard-AI’s GPUs.
PoliticsHome Newsletters
Get the inside track on what MPs and Peers are talking about. Sign up to The House's morning email for the latest insight and reaction from Parliamentarians, policy-makers and organisations.
