UK splashes £45M on AI supercomputer to help crack fusion power

The UK government is splashing out £45 million (c $60 million) on a new AI-driven supercomputer designed to help scientists model the chaotic physics of nuclear fusion, with the system expected to come online this summer at the UK Atomic Energy Authority's (UKAEA) Culham campus. The machine, called Sunrise, is being pitched as …

[may_i]

AI will not speed things up

Useful fusion power will continue to remain 30 years in the future.

[I am David Jones]

Re: AI will not speed things up

You’re too pessimistic. Just imagine how exciting it would be to live in a world where fusion power continually remains a mere 5 years away!!

[Korev]

Re: AI will not speed things up

With AI it'll only be three decades away though

[Anonymous Coward]

Re: AI will not speed things up

Wouldn't it make more sense to help ITERate joint research, rather than Grot Britain going its own silly Brexity way? Oh…

(I suppose at least as a supercomputer centre, Sunrise can (hopefully) be used by and results shared with researchers elsewhere in Europe and beyond, but, given that tokamaks are very much not the cheapest doughnuts in the shop, it seems a bit silly to duplicate that particular effort separately.)

[Korev]

Re: AI will not speed things up

I missed that one, an incredibly dumb idea

[Anonymous Coward]

Re: AI will not speed things up

How did you miss Brexit ?!

[Rich 2]

Algorithms

One thing I find odd is the number of times someone takes an existing problem (such as nuclear fusion simulation, for example), then (apparently) just “adds some AI” and the result is immediately “better” (or at least “wow! It’s AI”)

I know stuff-all about nuclear fusion simulation but I bet the (non “AI”) algorithms used have taken many many years (decades) to develop. How come new “AI” algorithms (which are obviously “better”. Obviously) seem to be instantly available? Or at least available as soon as the new “AI” machine has been built? I’m thinking “AI” algorithms must be massively different when compared to the boring old algorithms so how come they seem to get developed and written in an afternoon?

Or is the “AI” bit just …well …total bollocks?

[elsergiovolador]

Re: Algorithms

I think the idea might be that public sector cannot hire competent people as that would involve massive change of pay scales. So the "hack" is to employ AI that supposedly has better reasoning skills than your average civil servant. Ironically you can see evidence of it by how enthusiastic the big wigs are about adoption of the AI. They see AI slop and they instantly think it is genius.

[LionelB]

Re: Algorithms

I think you missed the "The idea behind Sunrise is to combine high-performance computing with physics-informed AI models" bit. I sincerely doubt your "average civil servant" has much expertise in plasma physics.

[Jedit]

"Or is the “AI” bit just …well …total bollocks?"

AI can be faster for iteration, as it can filter out pointless lines of enquiry rather than simply brute force everything. But that doesn't help to solve an unsolved problem, as without an extant solution it can't determine which lines are pointless.

As ever, AI cannot do anything that a human has not already done.

[Headley_Grange]

Re: "Or is the “AI” bit just …well …total bollocks?"

Read up on the protein folding that DeepMind did about 6 years ago.

https://www.science.org/doi/10.1126/science.370.6521.1144

When aimed at specific problems like this, with lots of possibilities, iterations and dead ends then AI can be crazy good. Like another OP, I know bugger all about fusion but I bet there are problems of magnetic flux and plasma flow that are too complex for current computation that a suitably trained AI could have a crack at.

[Doctor Syntax]

Re: "Or is the “AI” bit just …well …total bollocks?"

I remain unconvinced about that one. Proteins do not start as stretched out chains of amino acids that fold up when released. They're built one amino acid at a time in a specific sequence. As more are added the degrees of freedom of the orientation of each new residue must be very quickly limited. Is guessing really computationally cheaper then emulating the actual process?

[LionelB]

Re: "Or is the “AI” bit just …well …total bollocks?"

> Is guessing really computationally cheaper then emulating the actual process?

It's not (blind) "guessing", and no.

Firstly, ML algorithms do not make blind guesses. The fact that they may involve a stochastic element seems to mislead many. Let's say they make highly informed guesses. Here's an analogy: imagine that I have a coin that I know to be 95% biased towards tails. I know that, because I've tossed this coin many, many times. Guess what? I'm going to guess it'll come up tails. More realistically, in my own work (I'm a researcher in computational neuroscience) I will sometimes make use of stochastic search algorithms, like simulated annealing. Those methods are powerful; they work. Modern ML methods such as transformers and various learning mechanisms are not dissimilar in many respects.

Secondly, emulating plasma physics is insanely computationally intensive. That's why they use massive supercomputers for it.

[Anonymous Coward]

Re: Algorithms

I think the idea is focused on using specialized non-LLM AI methods to either reduce the testing space for physical prototypes, or increase the search space for influential parameters over which to optimize a design. In the Sandia case, it seems it helped them dig themselves out of some 'human intuition' corner they'd painted themselves in earlier (or somesuch), thanks to AI's unbridled irrationality (iiuc) -- ymmv.

[makeclean]

Re: Algorithms

Well for a number of years working on using AI in fusion, it mostly started out as machine learning based approaches or some Physics Informed Neural Network (PINN) type stuff, but increasingly HPC has been used to generate large datasets over a large parameter space, and then using different flavours of AI methods such as Fourier Neural Operators (FNO) as in this approach https://www.ukaea.org/news/modelling-a-star-in-a-jar-in-seconds/ which uses deep learning to learn the rules of the equations that define the physics that is of interest with subsequent very fast inference, meaning that for example very fast surrogate models can be used in place of traditional tools. There are people using AI to help analyze vast quantities of experimental data in order to look for patterns in the myriad signals that come from the experiment, for example to find correlations in different aspects of the plasma. Of course the machine can also be used to do traditional simulations but using GPU accelerated versions of the codes which allow for bigger and more complex problems to be solved with less time needed.

[Doctor Syntax]

Re: Algorithms

With added AI it costs more so it must be better. Obvious, innit?

[LionelB]

Re: Algorithms

Um, if you read the article, the point is that it's supposed to cost less in the long run, because it'll help narrow down which simulations are actually worth running – and running plasma models is ridiculously computationally expensive, and hence money expensive.

[LionelB]

Re: Algorithms

> I know stuff-all about nuclear fusion simulation but I bet the (non “AI”) algorithms used have taken many many years (decades) to develop. How come new “AI” algorithms (which are obviously “better”. Obviously) seem to be instantly available?

Perhaps you missed the bit: "The idea behind Sunrise is to combine high-performance computing with physics-informed AI models". And the AI machinery (transformers, a range of learning mechanisms, etc., etc.) have been around for a while. It may not be be as arduous as you imagine to tie those things together – there is previous for that, in other domains such as protein folding and drug discovery to name a couple.

Although this article seems to have triggered the usual suspects, bear in mind that we are not talking LLM slop here. AI (okay, call ML if you're easily rage-baited) has it's uses, and this may potentially be one.

> I know stuff-all about nuclear fusion simulation…

Well there you go. And how much do you know, when it comes down to it, about AI technologies?

[elsergiovolador]

Computer

It's great that we are buying a computer.

It's bad that we are not making a computer.

[Like a badger]

Not much "so what" here?

Most of this seems to have come straight from the relevant government press release, and I'm left wondering what £45m buys the taxpayer (other than about 150 metres of HS2)?

So does £45m actually buy any worthwhile scale of super computer?

How does 6.76 exaflops compare to other energy research computing around the world?

Is it likely enough to achieve anything worthwhile?

Or is this just part of a broader "£40m here, £40m there" programme in which our poorly qualified government hand out insignificant sums to pretend that the UK invests in science?

[elsergiovolador]

Re: Not much "so what" here?

Also computer is one thing, but how are they going to solve PEBCAK.

[Andy The Hat]

Is 1.4MW particularly beefy for a thunderous, brain stomping, Om-challenging AI entity that's not based on ARM or RISCV power sippers? Sounds a bit wussy to me.

[HuBo]

Looks like 1.4MW may slot this Sunrise somewhere between Top500's #22 Venado (GH200) with 98 PF/s at 1.7MW and #33 El Dorado (MI300A) with 68 PF/s at 1.1MW. Using AMD GPUs sounds like the right choice to me here, to hedge AI bets by maintaining proper FP64 performance (vs Ozaki). MI430X in particular would be very nice for this imho.

[Roj Blake]

Crysis

Can it run it?

[GioCiampa]

Re: Crysis

Only as a Beowulf cluster...

[Anonymous Coward]

Re: Crysis

It can! But not Vista

[Tron]

Brit AI will solve this first.

If they spend enough on this system, it will prove beyond all doubt that it will never be commercially viable.

The Nobel prize for Weltschmerz will be coming to Blighty.

[Doctor Syntax]

Now that we measure computing on $PREFIXwatthours we should be able to make a direct comparison between what this consumes over its lifetime and what the finished product, should it exist, produces.

[annodomini2]

So they're adding 128GB of RAM to the existing one