Nice! I never really been a huge fan of folding. I like Einstein@Home. Not sure why though....
I use an AMD X2 5600+ with an 8800GT. It runs all the time with next to no performance impact - kinda suprising since the CPU is pegged at 100% at all times. I let it use the GPU after 10 mins of me not using the PC.
I'm all for giving spare CPU cycles to science don't get me wrong. But I can't imagine how much a farm would cost in electric alone.
Coolio article, the Bit-Tech staff should host a rack full of one of those systems that you just built. ;)
It's a shame that there wasn't any graph comparing the performance of different CPUs on the CPU client, not everyone can easily run more than two graphics cards.
Corrections:
First page: "Stanford updates the folding points system regular basis..."
Third page: A GTS 250 is "much faster" than a 4870 X2 at folding? I'm not so sure, the difference is <300ppd. The GTS 250 is a much better value for folding and it does draw far less power, but they fold pretty much the same.
Fourth Page: On the charts, GTX 250 should probably be GTS 250
I can't say i really care much for F@H (please don't kill me now), but isn't the whole point of the project that they dont have to use a supercomputer? Why build a dedicated supercomputer then, aside from leaderboard ePeenage?
Anyway, i have no problem with designing over-the-top rigs, so why don't you use PCIe-risers to make space and fit 7 GTX295 cards.
You'd probably need a secong PSU to juice everything up, but iirc there used to be a PSU that natively supported daisy-chaining.
... or you could get the single-PCB watercooled version :D
I read this article a couple of months ago in issue 70 of CPC & thought it was good at the time, but after reading through it again today I thought I'd, (very quickly) check out how a homemade 'supercomputer' would compare against the real thing - and I'm STAGGERED by how powerful and cheap it is by comparrison!
According to the chap who runs Atlas, an nVidea 295 GTX is worth 1.788 TeraFLOPS - which means that CPC's / bit-tech's £2k machine generates approx 7.152 TeraFLOPS.
That works out at approx £280 per TeraFLOPS.
Southapton University has just ordered, and not yet recieved, it's own 'proper' 8,000 core IBM supercomputer that will produce 74 TeraFLOPS - at a cost of £3m!
That's approx £40,500 per TeraFLOPS - or approx 145 times more expensive for the same processing power!
I appreciate the purpose built machine for Southapton University will probably be more stable and will definitely have far better longevity, (and probably has service and support included in the price) but seriously, using CPC / bit-tech's method you could get the same power for around £21k as a £3m machine.
For that money, even if the GPU cards wore out every six months you could just replace / upgrade them - and still have a vastly more powerful machine for a fraction of the price.
Am I missing something stupid or grossly over-simplifying things? Their machine is designed for similar functions, including medical research....
Actually, re-reading the article, the machine is £1.8m with the extra £1.2m for the service & support - but my point still stands, as the overall cost is the same and with the home built system you wouldn't need anywhere close to that level of service & support.
Maybe you CPC / bit-tech guys should start moonlighting, building homemade supercomputers for institutions....
Maybe you CPC / bit-tech guys should start moonlighting, building homemade supercomputers for institutions....
I think nvidia already does this kind of thing with their Tesla brand. Also the FLOPS figures are misleading and probably aren't directly comparable but yeah you right in that you can get some pretty powerful desktop gear these days.
Am I the only one who saw that seven card monstrosity and wished they used 7 water cooled (AKA single slot) GTX 295's just to break the 100,000 PPD mark on ONE machine?
Thanks general22 - I wasn't aware of the Tesla Brand, but have had a look & think I can see why there's such a disparity:
The Tesla C1060 produces 933 GFlops of Single precision floating point performance, but only 78 GFlops of Double precision floating point performance, (I'm not going to pretend to know what that means, but it's a rounghly 12x disparity in the overal FLOPS numbers).
As you said, I probably wasn't making a like-for-like comparison but comparing single precision with double precision numbers.
Still, like you said, that's still some pretty amazing desktop gear for comparitively little money...
Gremlin i think the current version of Nvidia CUDA drivers only allow 8 gpus to work. Got to admit tho an 14 gpu system would rule aswell as consuming a hell of alot of leccy and also heating up your whole neighbourhood!!
Originally Posted by Gremlin Am I the only one who saw that seven card monstrosity and wished they used 7 water cooled (AKA single slot) GTX 295's just to break the 100,000 PPD mark on ONE machine?
There's a challenge for you guys at CPC/Bit!
While that would be a lot of fun to build, unfortunately, CUDA currently supports a maximum of 8 GPUs per Windows install at the mo, so it's impossible.
What about running everything under Linux in Wine? Seems like you could run a couple VM's to get all 14 GPU cores running if the limitation is CUDA... I was under the impression that the P6T7 was the issue.
Also, added bonus of the Linux install, the CPU can get MUCH better PPD numbers from the ability to run the A2 core (where Windows is currently limited to the A1 core IIRC).
Originally Posted by JaredC01 What about running everything under Linux in Wine? Seems like you could run a couple VM's to get all 14 GPU cores running if the limitation is CUDA... I was under the impression that the P6T7 was the issue.
How would that help? You can't fold on nVidia GPUs unless they're CUDA-capable (i.e. 8-series onwards) and have a CUDA driver, so you're always going to be restricted by the CUDA limitation.
huh and people wonder why we are in a power crisis lol =p ya I would hate to see their power bill thats for sure :) but hey its great because it goes to a good cause but man I would invest into solar power for those setups lol
Good article. I too would like to see the watercooled 14 gpu behemoth. It's got to be done just as a proof of concept to see if the VM route will work. Probably need dual PSU's too. :)
That said, Silverstone showed 4x 8800GX2/295's Folding in the Raven2, but did not show any thermals. That could be a good case for those without deep pockets as its speculated to be about £120.
Is there going to be a review of the P6T7? It looks to be a very novel board, even if it is X58.
I used to run a GPU client but it never got to 100% it always bombed out and started a new one. In the end I gave up with it. Seems to me that folding on GFX cards doesn't work very well.
Comments 1 to 25 of 48
ReplyI use an AMD X2 5600+ with an 8800GT. It runs all the time with next to no performance impact - kinda suprising since the CPU is pegged at 100% at all times. I let it use the GPU after 10 mins of me not using the PC.
I'm all for giving spare CPU cycles to science don't get me wrong. But I can't imagine how much a farm would cost in electric alone.
It's a shame that there wasn't any graph comparing the performance of different CPUs on the CPU client, not everyone can easily run more than two graphics cards.
Corrections:
First page: "Stanford updates the folding points system regular basis..."
Third page: A GTS 250 is "much faster" than a 4870 X2 at folding? I'm not so sure, the difference is <300ppd. The GTS 250 is a much better value for folding and it does draw far less power, but they fold pretty much the same.
Fourth Page: On the charts, GTX 250 should probably be GTS 250
Sextuple would be six cards, not seven.
Septuple or heptuple (if you're into older forms) would be seven.
Anyway, i have no problem with designing over-the-top rigs, so why don't you use PCIe-risers to make space and fit 7 GTX295 cards.
You'd probably need a secong PSU to juice everything up, but iirc there used to be a PSU that natively supported daisy-chaining.
... or you could get the single-PCB watercooled version :D
haha! i've missed those, i'm gonna give you some rep... but i hate to encourage people lol
According to the chap who runs Atlas, an nVidea 295 GTX is worth 1.788 TeraFLOPS - which means that CPC's / bit-tech's £2k machine generates approx 7.152 TeraFLOPS.
That works out at approx £280 per TeraFLOPS.
Southapton University has just ordered, and not yet recieved, it's own 'proper' 8,000 core IBM supercomputer that will produce 74 TeraFLOPS - at a cost of £3m!
http://www.channelregister.co.uk/2009/07/31/southampton_uni_idataplex/
That's approx £40,500 per TeraFLOPS - or approx 145 times more expensive for the same processing power!
I appreciate the purpose built machine for Southapton University will probably be more stable and will definitely have far better longevity, (and probably has service and support included in the price) but seriously, using CPC / bit-tech's method you could get the same power for around £21k as a £3m machine.
For that money, even if the GPU cards wore out every six months you could just replace / upgrade them - and still have a vastly more powerful machine for a fraction of the price.
Am I missing something stupid or grossly over-simplifying things? Their machine is designed for similar functions, including medical research....
Maybe you CPC / bit-tech guys should start moonlighting, building homemade supercomputers for institutions....
I think nvidia already does this kind of thing with their Tesla brand. Also the FLOPS figures are misleading and probably aren't directly comparable but yeah you right in that you can get some pretty powerful desktop gear these days.
There's a challenge for you guys at CPC/Bit!
The Tesla C1060 produces 933 GFlops of Single precision floating point performance, but only 78 GFlops of Double precision floating point performance, (I'm not going to pretend to know what that means, but it's a rounghly 12x disparity in the overal FLOPS numbers).
As you said, I probably wasn't making a like-for-like comparison but comparing single precision with double precision numbers.
Still, like you said, that's still some pretty amazing desktop gear for comparitively little money...
great guide. but i would have thought quad cores or i7 running SMP clients may generate more PPD than only letting the CPU feed GPUs
2x : 2GB (KIT 2x1GB) DDR2 800MHz PC6400 CL5 PATRIOT Signature Line
2x: AMD Athlon II X2 240 Dual-Core (65W), 2800MHz, BOX, socket AM3 (Regor)
8x: ASUS EN9600GSO/HTDP, 384MB DDR3
2x: MSI K9A2 Platinum - AMD 790FX/SB600, DDR2 1066, 4xPCIe v2.0 x16, scAM2+
2x: OCZ Silencer 750W QUAD, ATX V2.2, 4x pci-e
2x: OEM Microsoft Windows XP Home Edition ENG, SP3
2x: SEAGATE Barracuda 7200.12 250GB, SATA II NCQ 8MB cache
Total Cost: 1 271,99 EUR (retail price; 03 Aug 2009)
Total PPD: 29488
1271 Euros = 1 079 British pounds
Just a HW junkie doing some calculations... nvm. :|
While that would be a lot of fun to build, unfortunately, CUDA currently supports a maximum of 8 GPUs per Windows install at the mo, so it's impossible.
Also, added bonus of the Linux install, the CPU can get MUCH better PPD numbers from the ability to run the A2 core (where Windows is currently limited to the A1 core IIRC).
How would that help? You can't fold on nVidia GPUs unless they're CUDA-capable (i.e. 8-series onwards) and have a CUDA driver, so you're always going to be restricted by the CUDA limitation.
That said, Silverstone showed 4x 8800GX2/295's Folding in the Raven2, but did not show any thermals. That could be a good case for those without deep pockets as its speculated to be about £120.
Is there going to be a review of the P6T7? It looks to be a very novel board, even if it is X58.
And whats with all the bots about recently?
-
« Previous
-
1
-
2
-
Next »
Discuss in the forums