MENU

UK switches on its AI supercomputer

UK switches on its AI supercomputer

Business news |
By Nick Flaherty



The UK Secretary of State for science, innovation and technology Peter Kyle has switched on the second phase of the UK’s AI supercomputer, Isambard AI.

The machine, now has over 5400 GPUs and is by far the largest machine in the UK and ranks sixth in Europe (and 11th in the world).

The system, provided as modules by HPE, was installed in just over a year and uses the GraceHopper GH200 superchips that combine the ARM-based Grace processor with the Hopper GPU. There is also 25 petabytes of flash memory for scratch memory, and HPE’s Slingshot 11 interconnect with 15,000 end points in the 12 cabinets.

“This is one of the most energy efficient systems in the world,” Neil McDonald Executive VP for the server business at HPE tells eeNews Europe. “This is a very strategic capability being built out here to maintain competitiveness on the world stage.”

Funding of £225m covers the next five years, and the first call for projects would fill the first year of operation, says Prof Simon McIntosh-Smith, director of the Bristol Centre for Supercomputing (above) at the University of Bristol. However this demonstrates how this is a long way from the UK government’s target of 100,000 GPUs in the UK by 2030 to provide 420 exaflops of AI performance for sovereign AI training and inference.

“The sovereign AI issue is real and as geopolitics is more complicated it is even more real,” said McIntosh-Smith. “We could train ChatGPT in the UK with Isambard AI, we couldn’t do that before and the data would never have to leave the UK. The latest big models with 1tn parameters we could train in two or three weeks.”

There is another AI supercomputer planned for Scotland to replace the Archer system, with an investment of £750m.  

“It’s an ecosystem. I was really excited about the system in Edinburgh with more general purpose supercomputing as well as AI and there will be other things the government will announce,” McIntosh-Smith told eeNews Europe.

He also points out that the site has the capacity to expand its power envelope by 6x, which would provide a 10x increase in performance with the next generation GraceBlackwell superchips as the availability eases up.

UK Secretary of State for science, innovation and technology Peter Kyle switches on phase two of Isambard-AI

There is also an issue of control of access to the system. The UK Department of Science and Technology (DSIT) will determine who gets to use the system, allocating time to large enterprises, small companies and researchers as part of the overall industrial strategy.

“There aren’t many dedicated AI research facilities in the world,” he said. “The government recently ran a sovereign AI call and with that call we could fill up the machine for a year. How much of the machine will be available is up to the government, and DSIT is free to decide how they want it split.” 

Power consumption is another issue. Even with direct liquid cooling, Isambard AI uses 5MW of power, even with direct liquid cooling. Plumbing has been installed so that the waste energy can be used to heat the neighbouring National Composites Centre in Bristol, and McIntosh-Smith is in discussion with the local government on district heating to 15,000 homes.

However the power consumption will not be a limit on expansion. “We can get 30MW in the next few years so I have my eye on that to deliver a big chunk of the growth,” said Macintosh-Smith.

“Isambard is an incubator for AI to prove their business, this is the starting point of the journey,” said David Hogan VP of Europe for Nvidia. “We have challenges in the grid and distribution of energy, but with post industrial sites with access to energy, you can get around that. That aligns with other countries around Europe.”

This is highlighted by the release a large language model (LLM) later this year developed on the “Alps” supercomputer at the Swiss National Supercomputing Centre (CSCS). This uses 10,000 GraceHopper superchips.

www.bristol.ac.uk; www.hpe.com; www.nvidia.com

If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :    eeNews on Google News

Share:

Linked Articles
10s