Top scientists on Monday urged Rishi Sunak to invest $600m in building a new supercomputer capable of training up advanced artificial intelligence programmes.
These machines are necessary to train the complicated programs needed for artificial intelligence tools, such as ChatGPT, which was developed by Silicon Valley start-up OpenAI. These chatbots need huge amounts of computer power when being developed.
But that’s not what Chat GPT does. They want more computing they go rent some more computer time off Amazon or Google.
You don’t need – nor want actually – more shiny shiny kit to make AIs. You want more time. So, rent the time not buy the shiny shiny.
Putin controls more powerful computers than Britain, Sunak warned
He also has a lot more games and can stay up till midnight whenever he wants.
“Britain has fallen behind Russia and China in the global supercomputing race”
The Chinese computer recently set a new record time of two hours fourteen for the Electronic Marathon. Meanwhile, the British entry fell into a puddle, a few hundred meters from the starting line. Commentator Ron Obvious said: “Our computers just don’t get the same training, they’re well below the level of fitness that they need to compete at this level.”
So the future looks bleak for the UK and with only one bronze metal at the last Cyber Olympics things aren’t going to improve soon.
I’m sorry Dave. I’m afraid I can’t do that.
Academic funding same as the tax incentives (whether you are for or against them) has not yet caught up with the fact that unless your in the hyperscaler game compute/digital is now opex not capex. Took the fat controller over 10 years to allow cloud spend as part of the R&D tax credit.
Leaving the “training” of AI to computers sounds like asking Herr Hilter to carry out the entry assessments for a synagogue.
Better to let an average four year old sit down and have a chat with it for, oh, dunno, 14 years?
Who makes these supercomputers and where ?
The answer is 42. What is the question?
I’m sure that GCHQ has rather more computing power than I would like them to have, thank you and ChatGPT isn’t AI any more than Nuclear Fusion is just around the corner.
The website formerly fondly known as El Reg has been banging on tediously about how the British government doesn’t have a “semiconductor strategy”.
We need a “semiconductor strategy”, you see, otherwise something bad will happen.
Nobody’s really sure what that bad thing is, and we’ve somehow managed to get by without a “semiconductor strategy” to date. But we’ve got to get spaffing NOW, and put the future of the British IT industry in the hands of the same people who made lockdowns such a great success.
Hence the claim that Russia has better computers than we do. If this was still the 1970’s, and we were in a James Bond film, that might even be a valid complaint.
We should worry less about Russian computers and worry more about the fact that we blew our wad on sanctions, and all we got was this rapidly deteriorating economic situation with no relief, no victory, and no peace in sight.
For one thing, cloud and super computing is very energy intensive. Datacentres use a lot of electricity. Which is a bit of a problem for the energy-self-impoverished UK and Europe.
Don’t worry Steve. British researchers are working on the environmentally friendly, fully recyclable, wind powered wooden supercomputer as we speak. That’s what the 600 mil’s needed for.
Ah… So they ask for 600 mil, get 300, overrun up to over a billion, and in true Great British IT Projects tradition… It will be a decade late and not fit for the intended purpose if it works at all.
Hmm. The Telegraph says “According to the expert review presented to ministers, such a system could cost in the region of $600m. Scientists said the UK needed a computer with 3,000 processors, likely powerful graphics chips, to seriously compete in artificial intelligence.”
So, 600m US for 3,000 cores, largely straight into nVidia’s hungry, waiting sofa, for basically vector processors.
The Future of Compute report says “we define compute or advanced compute as computer systems where processing power, memory, data storage and network are assembled at scale to tackle computational tasks beyond the capabilities of everyday computers”.
What they actually talking about are single systems, in specific locations, that are whored out to users on the Lyons Ye Olde Tea Shoppe Logistics Model, ie. bureaus. They want to build ADP or Kalamazoo all over again, but this time it’s different, ‘cos it’s AI, see? Oddly, the Great and the Good giving evidence and writing reports get paid by the likes of Microsoft, Google, SalesForce et al.
The network bit doesn’t seem to be discussed at all. So, 27m-ish dwellings in the UK, ADSL/VDSL access probably around 90% or so, thus about 24m domestic routers (really modems, I suppose), each with CPU and RAM. Around 150~200 connected to each cabinet.
No routing takes place between modems/dwellings at the cabinet.
Change the network routing, upgrade the capabilities of each modem/router, and have a massively distributed cloud/compute capability. Throw in some FPGA type stuff for flexible (ish) hardware acceleration on the fly (maybe), and job done.
Stop fucking about engaging in fucking arms races, and do something different, for fuck’s sake.
I may be getting slightly annoyed with this sort of shit.
Good job their not asking for Quantum Computers, thats got to be worth at least a billion in funding, easily.
Sigh, I’m glad I lived through the 80’s when Britain had a computing industry, and the Sir Clive’s of the world simply got on with things without dipping their hand in the Taxpayers’ pocket.
Seriously? $200k per core? There’s some impressive pork stuffing going on in that request, or a journo has forgotten a couple of zeros.
We could get the NHS computer system on the job. It’s not working for the NHS so must have some free time.
Ducky: the VDSL modems are shit, and always will be. Little memory, which is really the thing that kills them for distributed computing, and a slow MIPS-based CPU so they can make it on a really cheap process. Many of them don’t even have the horsepower to run a decent QoS scheduler, which is why buffering is a thing on streaming and gamers pull their hair out because of delays.
There is no incentive for anyone to change that, least of all the modem/router manufacturers.
BiW – probably, it was lifted from the Tel’s report, and I assumed they’d muddled up processors with cores. That said, AMD 64-core Ryzen is up on Amazon for 7.5 large US. That’d get you to about 350,000 US for that number of GP cores in about 46 processors. The rest of it could easily be swallowed by RAM, disk and interconnects. Otherwise, they did mean processors, may be Nvidia Grace’ or H100s, but there’s still all the storage and buses and switches and blinkenleds. And kettle leads. And a couple of sheds.
TG – if we must have a chip strategy, an AI strategy, or a Future Compute strategy or a smart coffee percolator strategy, then that’s my bloody point.
–But that’s not what Chat GPT does. They want more computing they go rent some more computer time off Amazon or Google. . . You don’t need – nor want actually – more shiny shiny kit to make AIs. You want more time. So, rent the time not buy the shiny shiny.
You’re taking the journo woo too literally; just because Matthew Field says “. . .artificial intelligence tools, such as ChatGPT . . .” – doesn’t mean that government linked sciencers want to make or use ChatGPT-II.
ChatGPT takes a close look at what its users are doing (some AI apps even own what is done with them), and Google and Amazon are no doubt taking a close look at ChatGPT and the users. Not necessarily a good setup for sensitive operations. Sometimes you need to own the shiny shiny.
I forget what they call it but AWS has this feature where you can buy processing when it’s going spare. So, you set up processes and tell AWS about them, and when they have idle capacity they run things for you. It’s super cheap and would be a better way to run scientific stuff than buying a load of servers.
“The Met Office’s Cray supercomputer, which was at one point one of the world’s most advanced machines, is now ranked number 86 and is scheduled to be replaced.”
Having said the above, let’s not waste any more money on those twats. Getting the weather prediction wrong through a filter of anti-capitalist climate hysteria can be got on the cheap.
We should worry less about Russian computers and worry more about the fact that we blew our wad on sanctions, and all we got was this rapidly deteriorating economic situation with no relief, no victory, and no peace in sight.
Lol, Steve, you could use ChatGPT to couch all your comments in terms of “Ukraine’s all our fault” embitterment.
Maybe you are.
PJF – we need a BellingChatGPT to produce a steady stream of optimistic Commando comics style stories and associated war pr0n for our population of credulous adult children.
It’d be cheaper than paying the BBC to do it.
Bloke in Brum,
“Sigh, I’m glad I lived through the 80’s when Britain had a computing industry, and the Sir Clive’s of the world simply got on with things without dipping their hand in the Taxpayers’ pocket.”
Well… a whole lot of money went into Sinclair Radionics from the National Enterprise Board. Meanwhile, Sinclair set up Science of Cambridge with Chris Curry. So he was on the board of 2 companies, one with a big government ownership, the other was his thing. So, he worked this nice because he’d get people at Radionics to try things, do R&D, but of course, he was then learning from that and applying it.
I can imagine that it can make sense to have a national strategy for some relatively small number of things where it looks like having – or avoiding, strategies can be for any purpose – some capability would be a Very Good Thing but alas we don’t have it.
The strategy shouldn’t try to predict the future in detail. If we need supercomputers, then – well, y’know – it’s not hard. You can buy the things. If what you need isn’t a supercomputer but a bunch of distributed compute, then you can rent from AWS. You can also rent time on other people’s supercomputers.
Now, if you’re confident that you need access to a supercomputer in times of war, then presumably that’s a supercomputer which is paid for, owned and operated by the armed forces, so you make the case and whack it into their budget.
So for the University of Blechh to have access to a supercomputer of the right sort so it can do research into machine learning (or whatever) it simply needs to have sufficient budget to buy and operate or rent time elsewhere.
A National Supercomputing Strategy could be something like “we’ll take the research budget for our universities and repartition it so that supercomputing budget takes a larger portion and is prioritized higher”.
There y’go.
For industry, you’d look at what blocks companies from using supercomputers and see if any of it is the fault of government – and remedy that.