WCG GPU Client Discussion Thread

[BinaryMage]

Announced last September, WCG is planning to release a GPU application, initially for the Help Conquer Cancer project.

Quote from forums:

Hello WCGrid members,

The September 2011 update to the Help Conquer Cancer project has been posted. In this update, we announce work presented earlier this year at the High-Performance Computing Symposium in Montreal: the development of an OpenCL (GPU) implementation of HCC. This GPU version is currently running in our lab. We are pleased with its performance, and anticipate its eventual launch on the World Community Grid.

Please be patient in the meantime. It will take time. A large-scale launch on a new hardware platform is tricky business for the World Community Grid staff, especially while juggling every other active project and new projects emerging from the pipeline.

Thank you to everyone for your enthusiasm and interest in HCC, and your contribution of CPU cycles that make this and other projects possible.

Christian A. Cumbaa
Research Associate, Ontario Cancer Institute
University Health Network

Timing-wise (thank you KieX) a beta period is slated to start sometime in January 2012. (If you want to participate, you'll need to opt-in on the WCG website)

If this goes through as planned, our whole crunching strategy will likely change. GPUs are far, far more efficient than CPUs, FLOPS-wise. (Floating Point Operations per Second) For example, an overclocked i7 2600k, costing about $310, puts out about 14 GFLOPS. (reference) A videocard at the same price point, the HD 6970, puts out 2700 GFLOPS - 192 times what the i7 does. Now, GFLOPS do not directly translate into BOINC crunching power - but they are related. An i7 puts out maybe 5,000 PPD on an average BOINC project. The 6970 will do 250,000 PPD - a 50-fold increase. (source) Now, these GPU credit numbers are not from WCG, and the actual numbers from WCG GPU crunching will probably differ slightly. But nonetheless, crunching with a videocard instead of a CPU will likely be more efficient by many orders of magnitude. The official supported list of cards is here, but any AMD card supporting OpenCL or nVidia card supporting CUDA should work.

Now, many of you are probably frustrated right now. Your 2600k you just bought for crunching - you now think was maybe not the best investment. I like to think of it like this. We crunch to help people. Not to win contests. Not to accumulate points. Your i7 isn't going to do any less work. It's not going to help people any less. But hopefully using videocards in the future will allow us to help people more. If the GPU application is, say, 40 times more efficient than the CPU one (in tasks finished/day at a certain price point), then that means, if we switch to buying GPUs instead, the project can obtain hopefully useful data much, much faster than it could otherwise. More efficient ways to help people? Sounds good to me, and that's the bottom line.

[(FIH) The Don]

this will be VERY interesting, may even make me invest in 2-3 more 5870s just for this, as they are dirt cheap now and will get cheaper

[Bow]

Sweet put my 6950's to work

[f22a4bandit]

This is awesome news! I wonder how points would scale in crossfire/SLI?

[(FIH) The Don]

im thinking it will be like folding, where sli/cf has to be disabled

[KieX]

Thanks for putting up the thread BinaryMage. Might wanna add the links from the Daily Numbers thread to your OP to keep it all in one place.

As far as information on start dates:

At this time we are not prepared to give you a date for the start of GPU for HCC but we can tell you this:

1. At this time, the plan is to start Beta testing before the end of January, 2012. This may change as conditions arise.

2. Because of the large variety of processors, we will probably have a longer than average Beta test.

Source

Looks like it's still some time away from full implementation for end users. Anyone willing to participate in the BETA, will need to remember to opt-in for it on the WCG website.

Now.. the scoring system.. that's going to be interesting. According to WCG the current points are based on a 1 GigaFLOP machine, running full time, producing 100 units of credit in 1 day (~14Boinc points)

That would certainly entail a massive PPD difference over anything a CPU can produce (GPU vs CPU GFlops). And I would not imagine they'll want to go the route of F@H and start differentiating points

A lot of informations still to come, but looks very promising.

http://boinc.berkeley.edu/flops.php

[f22a4bandit]

Besides the PPD output (speculation at this point) another great thing about using video cards is their universality. You don't have to worry about new sockets, it's just plug-and-play. I know our video cards aren't stressed that much anyway unless we're gaming, editing, etc., so this just makes using your machine that much easier.

[Chicken Patty]

F22 has a very good point. A lot easier to stack up points with the GPU's than it is with CPU's.

[F150_Raptor]

Your also going to have a hell of alot more heat and a higher power bill. Both of those were the reasons why I stopped using gpu's with F@H.

[Chicken Patty]

Originally Posted by F150_Raptor

Your also going to have a hell of alot more heat and a higher power bill. Both of those were the reasons why I stopped using gpu's with F@H.

That is very true. But I already fold as it is so for me it's not much of a big deal. Plus I have a fixed rate monthly. So unless it's a huge jump I don't think they'll tell me anything. Just bad for those who will get the big hit.

[KieX]

Originally Posted by Chicken Patty

F22 has a very good point. A lot easier to stack up points with the GPU's than it is with CPU's.

If the points are calculated the same way for GPU then yes.

At the moment it's only a small part of one project which will have support. Might be a lottery for the first Work Units. I'm feeling lukcy

[f22a4bandit]

Originally Posted by F150_Raptor

Your also going to have a hell of alot more heat and a higher power bill. Both of those were the reasons why I stopped using gpu's with F@H.

Great point, I don't know why I even thought of that! But for those like CP with a fixed electricity bill, it is a great advantage.

[KieX]

Originally Posted by F150_Raptor

Your also going to have a hell of alot more heat and a higher power bill. Both of those were the reasons why I stopped using gpu's with F@H.

Although that's also partly due to the bigadv bonus system, right? In F@H until recently an i7 could outdo a couple of GPUs.

I don't think WCG being IBM sponsored want to go down that route of varying bonuses, so it may be that GPU's really are better than CPU for this type of project.

From my GPU folding days I can certainly vouch for 500W+ for dual GPU setups being way less efficient than a good CPU. All speculation until there's any confirmation, of course. But here's to hoping

[BinaryMage]

Originally Posted by F150_Raptor

Your also going to have a hell of alot more heat and a higher power bill. Both of those were the reasons why I stopped using gpu's with F@H.

True indeed. There are differences, though. First, BOINC assigns credit roughly based on how much data you crunched, while F@H does not necessarily. GPUs are far more efficient in FLOPS/watt than CPUs, so BOINC crunching with GPUs is more efficient in work/watt and credit/watt, while F@H crunching with GPUs is more efficient in work/watt but not credit/watt. In addition, F@H's GPU system is mainly build for CUDA, and doesn't use the more powerful AMD GPUs (speaking in compute power) as effectively as it could.

[FordGT90Concept]

The reason I switched from F@H to BOINC is because BOINC didn't use GPUs. The reason is simple: CPUs are more accurate. An incorrect value often gets caught immediately in a CPU where a GPU requires testing on a separate machine in order to verify it's correctness and still not with 100% certainty.

We're doing science here: slow, correct results are more desirable than fast, incorrect results.

If TPU starts adding GPU work to its project list, I'll have to stop running BOINC for TPU. I refuse to compromise science for an arbitrary, altogether meaningless number.

[BinaryMage]

Originally Posted by FordGT90Concept

The reason I switched from F@H to BOINC is because BOINC didn't use GPUs. The reason is simple: CPUs are more accurate. An incorrect value often gets caught immediately in a CPU where a GPU requires testing on a separate machine in order to verify it's correctness and still not with 100% certainty.

We're doing science here: slow, correct results are more desirable than fast, incorrect results.

If TPU starts adding GPU work to its project list, I'll have to stop running BOINC for TPU. I refuse to compromise science for an arbitrary, altogether meaningless number.

Really? Where did you hear that? I did some quick Google-fu and came up with nothing...

And anyway, BOINC tasks for both GPUs and CPUs are double or triple-checked, with multiple computers performing the same tasks. I doubt BOINC would use GPUs if they didn't produce valid data.

[ChaoticAtmosphere]

Well electricity is no concern for me as it is included in my rent and it's why I have a 10,000BTU air conditioner in my room dedicated to cooling my rig.....almost like a server room

More good news, I plan on buying a second HD6870 in 2 weeks (Or maybe I should buy the bulldozer, hmmmm decisions decisions.....)

Hell, I'll load up on 6970's or 7xxx's if each one can do more work than an overclocked 2600k.

[KieX]

Originally Posted by FordGT90Concept

The reason I switched from F@H to BOINC is because BOINC didn't use GPUs. The reason is simple: CPUs are more accurate. An incorrect value often gets caught immediately in a CPU where a GPU requires testing on a separate machine in order to verify it's correctness and still not with 100% certainty.

We're doing science here: slow, correct results are more desirable than fast, incorrect results.

If TPU starts adding GPU work to its project list, I'll have to stop running BOINC for TPU. I refuse to compromise science for an arbitrary, altogether meaningless number.

Might want to take a look at the PDF with their initial findings: http://www.cs.toronto.edu/~juris/WCG/UPDATE-SEP2011.pdf

It includes a section on their findings with regards to accuracy. Like you said, GPU is always going to trade off some accuracy but it seems that the researchers find this to be a tolerable margin in order to run simultaneously with CPU WU.

[Chicken Patty]

I won't argue FORD on the accuracy/GPU subject because he does have a valid point, but keep in mind that the higher than meaningless # is, the more research that is getting done.

Now is this just a personal thing that would not like to Crunch for a team that uses GPU's as well, or you refuse to crunch with GPU's? I mean, no one here can obviously hold you back, but as a consistant and dedicated cruncher you are, I'm sure no one here would like to see you go. I know I sure as heck wouldn't. Not only that, but you've been of great help to the community in other things not just the research.

[ChaoticAtmosphere]

Hmmmmm, Maybe a bulldozer then.....

[(FIH) The Don]

Originally Posted by twilyth

Hell, I'll load up on 6970's or 7xxx's if each one can do more work than an overclocked 2600k.

i hope that this gpu project wont have nvidia favored like it is in f@h

i hope for some equal competition in this, just for once

[Bow]

Originally Posted by (FIH) The Don

i hope that this gpu project wont have nvidia favored like it is in f@h

i hope for some equal competition in this, just for once

[Chicken Patty]

Originally Posted by ChaoticAtmosphere

Hmmmmm, Maybe a bulldozer then.....

I'm really dying to have some money left over and maybe build a BD rig. Just dying to see what it can do.

[BinaryMage]

Originally Posted by (FIH) The Don

i hope that this gpu project wont have nvidia favored like it is in f@h

i hope for some equal competition in this, just for once

The opposite will likely be true. AMD cards, by raw compute power, are way more powerful than comparatively priced nVidia ones, usually by about a factor of two. (For example, the HD 6990 has roughly 5 TFLOPs versus the GTX 590's 2.5) Folding@home, which was initially designed for CUDA (nVidia), hasn't ever really used AMD cards very efficiently. BOINC, however, does. Assuming this follows similiar trends as the other BOINC projects, you will get roughly double the crunching ability and thus credit from an AMD GPU versus a similarly priced nVidia one. Hope that cheers you up!