Bid for a world-class AI computer, then build it in record time and run it all? Certainly – just don’t look down. After the government’s recent launch of its AI Opportunities Action Plan, Professor Simon McIntosh-Smith, Director of the Bristol Centre for Supercomputing (BriCS), describes the 30–year path that led to Isambard-AI.
Getting Isambard-AI built has been like having kids or walking a tightrope. If you really thought about what you were attempting, you’d talk yourself out of it. So, you just do it and tell yourself it’ll be fine. And so far, it is.
If you stepped inside my office at the University you’d find a makeshift museum of Bristol tech artefacts dating from the early 1990s to now, guarded by a life-size cutout of Isambard Kingdom Brunel, the namesake for our series of supercomputers. After arriving in the city in ’94, I worked in the semi-conductor industry, then in various high-tech start-ups, producing advanced graphics processors. Then I co-founded Clear Speed; in my latter years there I worked with various universities, including Bristol. My collaborators said they were looking for someone to come and work in high-performance computing – would I be interested? So, when a post came up in 2009, I jumped across and never looked back.
The High-Performance Computing Research Group, which I created in 2010, has grown rapidly. When the government put out a call for a new AI facility in summer 2023, we put in a proposal. They did due diligence on us, then asked: ‘If money was no object, what’s your limit?’ I’ve never had that happen before in my entire career! Our limit was the power supply we had at our supercomputer site at the National Composites Centre (NCC) – up to five megawatts. ‘Okay’, they said, ‘build us a five-megawatt AI supercomputer – how much would that cost?’ The answer was £211 million and the total running cost over five or six years is over £300 million, so it’s about a third of a billion pounds in total.
1k to 128Gb to…?
It’s all a far cry from when I was in the school computer club in the 1980s, using a BBC Micro B. My family couldn’t afford a home computer, but I was round my friends’ houses on their ZX Spectrums and so on, typing in programmes out of the backs of magazines – that sounds crazy now, doesn’t it?!
The ZX-81 had 1k of memory and the ZX Spectrum started with about 16k; the Commodore 64 had 64k. These computers had an unbelievably tiny fraction of the power we’ve got now: one of the most recent bits of hardware on my office shelf has a GPU with 96 gigabytes of memory on the chip, and a CPU with another 128 gigabytes of memory. That’s eight million times more memory in about 40 years – Moore’s Law at work.
The graphics chip I worked on in 2000 had 70 million transistors – one of the biggest chips that had ever been made at the time. We’ve gone from that to 70 billion transistors in 2024. Today AI is very much driving the incredibly rapid rate of improvement in processor technologies.
Getting to grips with AI
Artificial intelligence is at the start of its hype cycle. It’ll start to calm down once people stop expecting it to be able to do everything, and then it’ll become widely adopted for the things that it’s really good at. Today people have similar concerns about AI as they once did about computers in general, or with the early internet; but typically these new technologies eventually become net contributors to our quality of life.
As with the internet, there could be misinformation or deepfakes, and there’s already lots of work underway to find ways of defending against them. The AI Safety Institute, a new government agency that will tackle a lot of these things head-on, is an early user of Isambard-AI.
We expect Isambard-AI to evolve quite rapidly. There are research groups in AI-based drug design and discovery who are among our first users. There’s a genetic modelling project that could lead to a better understanding of how cancer works – a computational challenge that was beyond this team until Isambard-AI came along. We’re having conversations about climate modeling and other potential uses every day.
Isambard-AI allows us to do things only Meta, Amazon or Google could do until now – it’s brilliant to be putting these tools and capabilities into the hands of scientists across the UK.
Heating up
Within six months, Isambard-AI part one was up and running, while part two – the main part – is due online this summer. It’s really exciting to challenge prevailing wisdom. For example, we’ve learned a lot about how to build a supercomputer much more quickly than usual. The quicker you can do it, the more science you can get done sooner – and the faster taxpayers’ money can bring benefits.
The University’s energy supply is all renewable and we’re building one of the most energy-efficient supercomputers ever. The waste heat from the system is all captured in the form of hot water. In Scandinavia, they plug this waste hot water into district heating circuits for local homes. We’re in talks with our local council about whether we can do the same. That’s the next challenge.
When I eventually retire, my ambition would be for BriCS to be running multiple future generations of Isambard, providing an exceptional service supporting amazing science around the UK, with waste energy helping the local community. Then, I’ll be happy to leave BriCS in someone else’s capable hands to run!