‘Ang on…..

£1 billion’s real money:

A new supercomputer providing more accurate forecasts of severe weather is to receive £1.2 billion from the Government towards its development.


The supercomputer itself is expected to cost £854 million

I thought we’d got to the point where a supercomputer was simply lashing thousands of PC boards together? Is there really a computer out there that will cost £800 million? And if there is, won’t it still cost £80 million in 3 years time?

23 thoughts on “‘Ang on…..”

  1. “The supercomputer itself is expected to cost £854 million”

    A quick google says the worlds ‘biggest’ supercomputer is Summit at the US Department of Energy. Interestingly;

    “The $200 million supercomputer is an IBM AC922 system utilizing 4,608 compute servers containing two 22-core IBM Power9 processors and six Nvidia Tesla V100 graphics processing unit accelerators each.”

    $200m is a lot less than more than a billion. Odd that.

  2. Presumably, based on the previous track record of government IT projects, the final cost will be 12 billion & it won’t work.

  3. Wiki – “In September 2003 the Met Office moved its headquarters from Bracknell in Berkshire to a purpose-built £80m structure at Exeter Business Park”.

    Anyone care to guess ‘weather’ (ha ha) their forecasts will be any better for the spunking of £1.2 billion?

  4. Things have changed a little in the supercomputing world since Summit, Sierra and their ilk were first put out for RFP. The I/O subsystem now needs to be so quick that it must be part of the supercomputer itself, rather than relying on external storage, and that increases cost massively. However, without that, you can’t get into the exaflop range. A supercomputer in the exaflops range is required for training AI models for granular weather forecasting and climate change modelling. Whether you think that’s worth it is up to you to decide, but the cloud computing bill is going up already as they realise their own supercomputer can’t do what AWS can do in the same amount of time, so it’s probably a net money saver. AWS will still be worth it for supporting websites and apps, and for archival data storage, but it’s an expensive place to train an exascale AI model.

  5. Bloke in North Dorset

    This isn’t really about forecasting in the general sense, they appear to be reaching the limits at about 6 days ahead. This is about the area that a forecast covers. I haven’t seen the detailed figures as the stories around this are quite simplistic but, IIRC, current forecasts are based on bins of about 25 sq km and the want to get down to something 1 sq km bins. The BBC implies they are getting down to individual street level, but that’s probably journalistic licence.

    That’s quite a lot more calculations hence the need for ever bigger computers

  6. The law of diminishing returns applies. A desktop PC can forecast reliably maybe two days ahead. A $1m computer might give you three days. A $1bn computer gives you three and a half days.

    If all the weather-sensitive industries change their working practices to incorporate these new, more accurate forecasts, then the country as a whole could perhaps save £1.2bn. But again, diminishing returns apply. If I know there’ll be a huge storm in 3½ days’ time, I can squeeze an extra half-day of fishing out of my fleet. But that extra half-day of notice isn’t going to make a material difference to most industries.

  7. Long story short: while sorta you lash hundreds of thousands of PC boards together, that’s only in principle. When you have that amount of compute you need lots and lots of memory and it had better be able to survive memory errors: restarting your 33 day computation every three days because of a memory error isn’t useful.

    And then you need inter PC board communications, and PC boards don’t come with any.

    Then you have ta integrate truly massive IO systems (as noted above) and PC boards don’t come with that either.

    Then there’s the cost of big boxes and lots of power and cooling and….

    So the good news is yes: you build big ole computers by aggregating lotsa smallish computers, but if you built PCs to be the building blocks of such systems they’d be quite a bit more expensive.

  8. “More accurate forecasts of severe weather” is the timely excuse to justify the kit.

    The Met Office have been here before; the last upgrade was to allow their climate models to be integrated with the forecasting models, which lead to the “barbeque summer” long range forecast, whereupon it promptly pissed down for 18 months or so.

  9. It doesn’t matter how much they spend on the hardware, if they don’t seriously improve their distinctly shonky software models the forecasts aren’t going to improve – they’ll just get the wrong results more quickly.

  10. One thing that they can do with faster computers is perform the same calculation multiple times with slightly different inputs, to test sensitivity to initial conditions. I used to joke with my actuarial colleagues that there’s no point in giving actuaries 10x faster computers, because they’d just run the same calculation 10 times and take the average.

  11. When the Met Office left Bracknell a large percentage of the more experienced forecasters took early retirement because they didn’t want the faff of moving to Exeter.
    There was an attempt to get them to move with the promise that house prices were lower in the West Country but (of course) as soon as the local estate agents heard that several hundred people would be looking for properties the prices shot up.

  12. The giveaway is the words “climate modelling” – what’s the point since we know all climate models are set up to forecast imminent disaster unless we return to feudalism, can we keep our 1bn fucking pounds please?

  13. Putting a few bucks into decent software would be better.

    The lack of good programmers leads to brute force computing.

  14. And the other side . . . what value will it provide. These projects are more “fun with meteorology” than anything useful.

    Current systems provide warning of severe weather systems days out. With the new system, they can provide warning of severe weather systems days out.

    I.e., you get no practical value from your billion pounds. Telling you 5 days out instead of 4 days out is worthless.

  15. Then there’s the cost of big boxes and lots of power and cooling and….”

    And make sure it has a damn good UPS – it would be a shame (NOT) if the Met office forecast failed because all that “Renewable Energy” went on strike due to lack of wind & sun, and took the grid down…

  16. Someone is taking the piss

    Edinburgh Uni’s new supercomputer last year total cost was £79 million and five times faster than any other in UK


    Oh yes, I remember that “barbecue summer” I had to use strimmer to cut the sodden grass, followed by a few “warm winters” with several feet of snow

    Following these failures Met announced they’d no longer release medium/long range forecasts

  17. Yeah, but building it in 6 years won’t make some functionary’s CV look absolutely ripping. “Defined, designed and led a billion-pound project to support the Met Office into a world-class 30-year future” (or words to that sort of effect) sounds better than “decided to shelve the project and save some taxpayers some money”…..

  18. As someone who lives by the weather forecast, especially in the summer, I would refute the idea that the current forecasting system is accurate beyond about 48 hours. You can normally rely on that to be OK. Beyond that, it becomes an increasing crap shoot. Not only that, the forecast for a point say 5 days ahead will change multiple times before that day arrives, which hardly makes it a forecast at all.

    Thats the trouble with these computer generated forecasts, they just keep rerunning the model, oh the outputs changed, out goes all the data to the websites, everything gets updated. You can watch the model runs during the day when the weather is catchy, the prediction changes by the hour. I look at the weather forecast at 9am, go out, by lunch time it may have altered completely, from wet to dry or vice versa. And by tea time it may have gone back again. Its utterly useless as a predictive tool.

    The computer model process also rarely predicts step changes in weather, when you transition from a series of storms say, to a spell of settled weather. Or when a hot dry spell occurs. All the models tend to stick with the current system and try to adjust it a bit, using long term averages I guess. No computer modelling system predicted the long hot summer of 2018 for example. Or indeed any of the very cold winters around 2010.

  19. In unsettled weather, if I want to know if a can cycle to the village shop and back (3m) before the next heavy show I look at the radar animation for the two hours past and make my own forecast [ https://www.netweather.tv/live-weather/radar ]. If I want to know what to expect fo the next day or two I look at the wind and precipitation models [ https://www.ventusky.com/?p=51.0;-2.5;5&l=radar ]. Anything beyond that, look at the jetstream [ https://www.netweather.tv/charts-and-data/jetstream ], south of jetsteam=good, north=bad.

Leave a Reply

Your email address will not be published. Required fields are marked *