Professor Simon McIntosh-Smith
๐ค SpeakerAppearances Over Time
Podcast Appearances
They could be up to a million GPU hours.
So that's the sort of thing that if you were buying that time in the cloud, that would be costing you millions of dollars.
You can get that as a grant to run an exciting project on Isambard AI today.
I mentioned the project looking at pictures of skin cancer.
It had a hypothesis that some of these algorithms had bias built in that where they work better on paler skin and not so well on darker skin.
And they've actually already proven that is the case.
and they've already improved the training data so that it's more diverse.
We run the algorithms and they've got a lot better.
So one of the first breakthroughs that we've had was already starting to significantly improve how accurate these algorithms work on a wider range of skin color.
It does.
Yeah, I'm always thinking about what comes next.
So, Isambard AI is actually the fourth in the line of different supercomputers that we've been building.
We had Isambard 1, 2 and 3, now Isambard AI.
We're already considering what would we do next.
The government's very keen actually to try and get significantly more compute than this even for AI by 2030.
So, there's a lot more to come.
And we've got space nearby and more power nearby where we could build even bigger versions of Isambard nearby very quickly.
We're even considering things like different sorts of architecture that maybe will be optimized specifically for AI inference and maybe not just everything running on the same GPUs, but different kinds of hardware.
different kinds of problems.
We've even got some early discussions about possible quantum computing projects that would be connected to AI in some way.