Slashdot Mirror
16-Teraflops, £97m Cray To Replace IBM At UK Meteorological Office
Memetic writes: The UK weather forecasting service is replacing its IBM supercomputer with a Cray XC40 containing 17 petabytes of storage and capable of 16 TeraFLOPS. This is Cray's biggest contract outside the U.S. With 480,000 CPUs, it should be 13 times faster than the current system. It will weigh 140 tons. The aim is to enable more accurate modeling of the unstable UK climate, with UK-wide forecasts at a resolution of 1.5km run hourly, rather than every three hours, as currently happens. (Here's a similar system from the U.S.)
16 peta not tera FLOPS
It will spend its days predicting it's own global warming impact
16 TFlops ain't much to write home about. 480,000 CPUS? What are they? 6502s?
Turns out it's 16PFlops according to the BBC.
SJW n. One who posts facts.
What, are you suggesting they fucking read the articles they're going to post? Or more absurdly yet, be broadly informed about the general goings on in technology?
One might even imagine that this headline, the weekly articles about the latest multi-teraflop figures from single GPUs, and some working synapses might have raised a SIGREDFLAG or something.
Slashdot is getting worse by the minute.
Those guys are still around? I thought they were all eaten by dinosaurs. How many times have they gone bankrupt now?
I'm trying to teach myself to set people on fire with my mind... Is it hot in here?
I find it ironic that on one side they deny climat change, on the other - almost the same people build the ark.
All hope abandon ye who enter here.
I miss the days when supercomputers looked super. This one looks like a row of drinks machines.
As a British nerd my 2 favourite topics of conversation are the weather and super computers, so this is exciting news.
I was interested in what the change-over was, which was causing the performance increase, and how old the existing system is. This information seems to be missing.
What is included actually sounds a little disappointing:
13x faster
12x as many CPUs
4x mass (3x "heavier")
I would have thought that there would be either a process win (more transistors per unit area and all that fun) or a technology win (switching to GPUs or other vector processors, for example) but it sounds like they are building something only marginally better per computational resource. I suppose that the biggest win is just in density (12x CPUs in 4x mass is pretty substantial) but I was hoping for a little more detail. Or, given the shift in focus toward power and cooling costs, what impact this change will have on the energy consumption over the old machine.
Then again, I suppose this isn't a technical publication so the headline is the closest we will get and it is more there to dazzle than explain.
UK weather forecasts have become much more accurate over the last few decades as the computers that do the forecasting have become more powerful. This new machine will continue that trend.
http://www.metoffice.gov.uk/media/image/7/2/capIndPlot-600.jpg
Eh quite a bit of industry where even small impovements in weather forecasting are extremely valuable.
To tell which of the UK's three weather conditions (rainy, cloudy, or foggy) it's gonna be?
Now dat just Cray.
Cameron shot that one out of the water when he promised to victims of the floods last winter: "MONEY IS NO OBJECT."
Some of us have long memories.
Political debates have me rolling my eyes so much I think I got optical whiplash. I should sue. - Foamy The Squirrel
My imagination leads me to think that it is AWESOME.
Oh, and it uses a lot of lectricity stuff!
Better than 2000's ASCI White, but worse than 2002's Earth Simulator. 13 years back to the past!
Or maybe the actual performance is 16 PetaFLOPs, as the linked article states.
Hmmm. I wonder if you are just confirming what the parent comment said. The sheer linearity of that graph indeed hints that the improvements have mostly happened by just throwing more and more raw CPU power into the task, without breakthroughs in making the algorithms more accurate or efficient.
...I could predict the weather for the same price.
Rain, rain and more rain.
And neither mentions the CPU architecture, but if you go to the product brochure then you learn that they're Intel Xeon E5s (which doesn't narrow it down much). Interesting that they're using E5s and not E7s, but perhaps most of the compute is supposed to be done on the (unnamed, vaguely referenced) accelerators.
I am TheRaven on Soylent News
Interesting that they're using E5s and not E7s
Probably something to do with yields and availability - buying 480,000 CPUs in one go is going to cause consternation, regardless of who your supplier is :) Getting 480,000 E5s in half the time it would take to get 480,000 E7s means you have less liability on the books for the duration (you have to hold delivered stock and down payments as liabilities), and a better cash flow.
The hardware is the cheap part - coming up with better algorithms is akin to mathematical breakthroughs these days...
Even if that's true that the algorithms are pretty much unchanged, that the accuracy gets better when throwing resources at the problem probably means the algorithm is working as intended.
Seriously, all the useless stats... weighs as much as 11 double decker buses... I've asked the metoffice on twitter but they ignore more. How much electricity is this gonna need?
[FUCK BETA]
Don't tell the UK meteorlogical Office this. The trick to getting those TFLOPs on those 480,000 6502 CPUs is that Cray benchmarked them all with nothing but NOP instructions.
Trolling is a art,
True.
As long as the weather on the island of Islay stays the same so the flavour of Laphroaig never changes, the rest doesn't matter
Trolling is a art,
Does it run Linux?
Not mentioned in TFA, and I haven't seen anyone talk about it yet in the comments here. Or maybe the answer is so obviously 'yes' that nobody even talks about it anymore.
I am not really here right now.
Oh look, it's another small minded Little Endlander with their "I don't understand it, it doesn't benefit me directly and it costs money, so it must be bad". See also HS2.
It benefits the UK economy massively. It allows shipping & aircraft companies to make sensible decisions like "Should we have the snowplows on standby tonight?" and "Should we wait in port while that storm passes?". It benefits farmers by giving them more accurate long-range forecasts so they can plant and harvest more efficiently. It even benefits you directly by letting your council plan their road gritting better.
I'm sick and fucking tired of stupid, small minded people in this country with their stupid, ill informed opinions. The UK is the 6th largest economy in the world by the way; perhaps we could celebrate that fact instead of whining about "Oh no someone is spending money!" The last reported GDP was £1.5 TRILLION. £97m is chump change.
predicting the weather will be a breeze ....
I would have hoped that they would have a cluster of Raspberry Pis to do this instead.
Donte Alistair Anderson Roberts - hi son!
Karma: Chameleon
England will be covered in Cray skies. No Sun.
97 million pounds is a pittance in a 731 billion budget. An Eurofighter Typhoon costs 110 million (marginal cost, not factoring R&D in).
This post contains no rudeness or derision of any kind. All arguments are friendly. Terms and exclusions may apply.
I think the problem isn't the lack of centralized computing resources but rather the lack of distributed sensors. The UK is quite small. If they would have spent the money on blanketing the country with sensors, they could give a much more localized and up to date weather forecast. I find that I get the best forecast for rain if I look at the radar map. But it requires quite a bit of time to read the map. I should be able to check on my phone, which has GPS anyway, and determine if it's going to rain, but I've never seen a weather app that takes your exact location into account. They all just give you the information for your city, which can vary quite a lot even within a few kilometers.
Anthropic principle: We see the universe the way it is because if it were different we would not be here to see it.
Yes, it runs linux.
Cray Linux® Environment (includes SUSE Linux SLES11, HSS and SMW software)
Extreme Scalability Mode (ESM) and Cluster Compatibility Mode (CCM)
system specs
There is a reason that organizations by supercomputers. There is no cloud in the world that can do what this computer does. None. Nada.
Cloud computing can run multiple copies of Office and host a website but when you need real horsepower, you get a supercomputer.
E7 is useful for areas where extremely large memory per core is mandatory (some parts of HPC)
In general, E5 strikes the balance between having adequate amounts of cache and SMP interconnect, compute capability (Haswell E5 is available, E7 is still Ivy bridge, AVX2 being a big thing there), and per-unit cost (E7 carries a huge premium for its benefits, most of which are generally not needed in HPC of this scale).
Even in places where you do see E7, it's usually in a special portion of the cluster for big-memory jobs that can't be split into multiple nodes as easily as most HPC workload, with the majority of the clusters employing something more like E5.
XML is like violence. If it doesn't solve the problem, use more.
It might also be that Intel has a bit too much capacity for E5s, and needs to utilize it. Unused semiconductor capacity is costly. Now don't get me wrong: this might simply be a case of more efficient capacity being available. E5s and E7s may be all made on the same equipment, but if said equipment makes E5s at half the cost of E7s, and you can sell them for more than half the cost of E7s, you really have more capacity in terms of what's sensible to use for ROI.
A successful API design takes a mixture of software design and pedagogy.
Of course the mention of 6502 was a joke, but let's see how close one could get. Let's say that you could get one FLOP in 1000 cycles on a legacy 6502. With 2MHz clock, we're talking 2kFLOPs per chip. With half a million of them, we get 1GFLOP. That's still 7 orders of magnitude away from where one needs to be... This tells us, indirectly, that the desktop processors we currently have are essentially the realm of 1980s science fiction :)
A successful API design takes a mixture of software design and pedagogy.
In the end, you have to do those additions and multiplications, there's nothing to be more efficient about. All those computations run on a grid, and the elements in the grid can approximate effects of various orders (think polynomial orders). Up to a certain point, increasing the order of individual elements decreases the net amount of computations done, since the increase in number of computations within an element is outcompensated by the decrease in the needed number of elements. At a certain point, your hardware is not accurate enough to go to higher order elements, and by doing multiprecision arithmetic you're decreasing the effective number of computations that could be done, so you're stuck there.
Unfortunately, brute computational power and inter-node link speeds are the only way to attack large, grid-based calculations such as weather forecasting, fluid dynamics, mechanics of solids, etc.
A successful API design takes a mixture of software design and pedagogy.
I should probably say that it speaks to the incredible flexibility and scalability of our grid-based methods that they even can be scaled in such a fashion. Some numerical methods simply don't scale at all, and throwing more computational power at them gives slower-than-linear increases in accuracy or decreases in computation time. For example, good luck with scaling up the grade-school long multiplication, or with single-polynomial approximations that span more than a dozen points...
A successful API design takes a mixture of software design and pedagogy.
And how is improved forecasting going to really help here, when you get past the platitudes? Is the transportation and rescue infrastructure up to par to cope with the evacuations prior to a forecast flooding? I somehow doubt it is. But feel free to prove me wrong, of course.
A successful API design takes a mixture of software design and pedagogy.
The grid that NASA used for those Neptune weather predictions probably had a cell the size of a large Earth country, or a small Earth continent. Neptune is fucking big.
A successful API design takes a mixture of software design and pedagogy.
The current NWS computer is only capable of 0.21 petaflops. There is an upgrade to bring it up to 0.8 petaflops, After Sandy (1.5 years ago) Congress gave money for a new computer but nothing seems to be happening with that money. Sandy's forecast was good not because of the American forecasts but because of the European forecast. I believe American forecasts were wrong in predicting Sandy's direction because America lacks of a decent supercomputer for forecasting.
Sure they could do that, but there's simply no cloud provider out there who has sufficient connectivity for the needs of a supercomputing system. The stuff one runs on a supercomputer would completely saturate the normal "cloud" datacenter interconnect, while leaving the nodes hopelessly underutilized. Serving web apps and doing large-scale computations have very different scalability requirements. That's why it's easy to scale a big cloud storage/app serving facility, while it's really hard to scale a supercomputer.
A successful API design takes a mixture of software design and pedagogy.
I see. Still, with the big data craze, I can imagine we are hitting such issues already and are either working around them through better software architecture and algorithms or invest in the hardware and offer it as a premium service.
"Software for Cray® XC Series Supercomputers"
Not everything in the world can be solved by clever software.
To have a right to do a thing is not at all the same as to be right in doing it
The so-called big data can replicate the "data" to each node, thus alleviating the interconnect requirements - most "big data" analytics is highly parallelizable with no interconnect. When you're doing big matrix inversions, the communications needs scale with the number of matrix elements...
A successful API design takes a mixture of software design and pedagogy.
You do more than rescue. When you know the storm is coming you prepare ahead of time. With 3-5 days notice, Councils, police cancel overtime. All vehicles are out of the garage/repair shop. Priority on getting sandbags in place, clearing all drains and drain covers.
Then the general public are warned. Less events are on, or they are cancelled. Less people travel, everyones been to the shops two days before.
And away from storms, farmers know 5 days in advance what they're doing; warm humid weather means preparing for blight, etc. Less fertlilizers, less pesticides are wasted.
People still grumble about the bad weather, but harvests and lives aren't lost.
Anyone who believes exponential growth can go on forever in a finite world is either a madman or an economist
Actually, this is misconception. The cloud can probably deliver 16Pflops. The problem with the cloud is not computation power. It is communication bandwidth and latency.
What makes a supercomputer is the balance between processing capability, communication capability and IO speed. For many applications, you need to be able to synchronize the processors with very little overhead. Many scientific application work under the following patterns: do a small computation, make a small communication with your neighbors, rince repeat for 10 hours. If you do not have balance capabilities, you are wasting lots of ressources. This is the type of computation the cloud can not really help you with.
Now if your application is: get 1MB of data, compute for 2 week, send 1MB of data. Then the cloud will be fine. Unfortunately, not many scientific applications follow that model.
Exactly, The Cray Y-MP that I was drooling over in 1988 has processors with a 167 MHz clock and 512MB of main memory. Now you can fit a faster CPU in your shirt pocket.
A NOP on the 6502 takes two cycles, so it was only half a million.
Nature uses a grid based algorithm to run the weather, so it shouldn't be a surprise that it works.
You can use a real cloud to see if it's going to rain.
The test for speed is not 16Pflops of raw computation but 16Pflop on the Linpack test suite. And no the cloud cannot do 16Pflops as they measure it on a supercomputer. You may be able to spin up more nodes to get more cpu power but I cannot spin up 100 new network connects and get 100x the bandwidth. Or get the sub microsecond latency of a supercomputer no matter how many connections you have.
Supercomputer are in a class by themselves.
so all the people using the amazon ec2 to run bitcoin and later altcoin/flavor of the month alt coin clones to make money aren't using cpu power? i looked into altcoins and it is pretty clear each coin launch is a huge way to launder 10 million in money easily.. and ec2 cloud computing is recommended for that use. criminals also use it to get gold and silver in exchange for their mined coins.
https://www.gnu.org/philosophy/free-sw.html
Flawed assumption: everyone is a feckless whiny turd like you.
Confucius say, "Find worm in apple - bad. Find half a worm - worse."
You checked it but we still know it's you, skid.
Confucius say, "Find worm in apple - bad. Find half a worm - worse."
It's not just a simple mistake that anyone could have made. If you know anything about computers at all, the error in the title, when you read it, is about as subtle as someone smacking you across the face.
If Soulskill doesn't know the difference between TFLOP and PFLOP, what is he doing posting articles here?