New Release: GPU2 Client (open beta) for HD2xxx/HD3xxx Cards

[werty316]

The time has finally come peeps:


We're (finally) ready to start the QA process outside Stanford & ATI for the next generation GPU client (nicknamed FAH GPU2). We've gone through a limited alpha and a larger closed beta test, and we think the client is now ready for an open beta, i.e. a public release of the software that's still in an early stage. That's not to say that it's flawless, but there don't appear to be any "show stoppers" so far.

If you have a 2xxx or 3xxx card, please feel free to check out the new GPU client? You can download it from here

http://www.stanford.edu/~beberg/Folding@home-R600-systray.zip

and a FAQ is here

http://folding.stanford.edu/English/FAQ-ATI2

We will next put this on our main download page, assuming that this stage of beta testing looks good.

Please keep in mind that this is beta software. The client/core may be unstable (although it looks ok so far), the points benchmarking may change due to software or policy changes, etc. With that in mind, we'd appreciate any comments you may have -- please post them in this forum section. Thanks!

Source: http://folding.typepad.com/news/2008/04/gpu2-open-beta.html

[PP Mguire]

Why dont we have any nvidia clients?

[werty316]

http://www.bjorn3d.com/forum/showpost.php?p=101385&postcount=6

Nvidia can't be bothered to support F@H, since Stanford refuses to use CUDA to write their application in. This is because Stanford writes their program in DX10 code so they will not be locked into a vender specific GPU. Nvidia's drivers are reportedly still to buggy and for whatever reason they have presently refused to fix them to allow Stanford to code with them. http://forum.folding-community.org/ftopic16684-0-asc-195.html

I wish I had some stats on how many people went out and bought ATI cards just to fold on, I've seen a surprising number of users doing this with used/new ATI cards. Maybe it would convince Nvidia it was worth their time.

[Miker]

This is stupid.

"DX10 code so they will not be locked into a vender specific GPU."

With out Cuda there is no Nvidia, only ATi, so it is vender specific.

[Kougar]

I suppose that is one way to look at it. However DirectX code isn't vender specific, and if NVIDIA doesn't think F@H is worth their time while ATI believes that they are then there really isn't much to do about it. NVIDA does things their own way.

I have been hearing this reason given for more than a year... but the crux of it is ATI is simply providing a good deal of help to the Stanford F@H team. Nvidia is apparently not providing much at all. Second time I've seen mention of Nvidia drivers having bug issues.

Also, the project is not only working with ATI, but is also working with NV to make GPU clients. However, the ATI program seems to be moving a lot faster from a lot of help from ATI. There are some holdups with NV drivers that Stanford cannot overcome without more help. Source (http://foldingforum.org/viewtopic.php?f=38&t=1981)

[PP Mguire]

Tesla would pwn anything TBH.

['Amp!']

Tesla would pwn anything TBH.

Not without support it wouldn't

[PP Mguire]

Lol well i meant with support of course.

[raidersforever01]

I love the new gpu client i am getting 1500ppd if only i could OC more to better feed the gpu.

[werty316]

Is your HD3850 overclocked? 256MB ror 512MB model?
What is your PPD with your HD3850? (I would guess somewhere around 1100-1300 perhaps?)

[raidersforever01]

My 3850 is running at stock with 256mb of ram. I don't think my CPU is doing a good job of feeding it i only get about 80% out of the gpu and about 1420 ppd. I even had it under clocked for a day and was still getting 1420 ppd. So i don't know if its being weird because its new or my cpu can't keep up.

Edit: after messing withit some more i am 99.9% sure my cpu is holding me down. When i underclock to 600mhz i get 80-85% from the gpu but when its at stock 669 i get 70-80% gpu activity.

[trueg50]

Why dont we have any nvidia clients?

Wanna take a lucky guess, between ATI and Nvidia who does a lot of donations, funding etc.. ?

Yea, thats why despite CUDA being much easier, it isn't going to happen any time soon.

[werty316]

Damn raiders, a ppd of ~1420 is quite sweet.

It does seem that the new cards are dependent on the speed of the CPU based on the user feedback over at the FAH forums.

[raidersforever01]

My overclock is quite high probably why i am higher then most on the forums their but even then i can't feed the gpu completely.

Edit: Figured out how to max out the GPU i just run 2 instances of the gpu client now i average about 2100ppd when i am letting the computer sit idle with only the gpu working.

[ZhengHe]

Extremetech has written a fairly extensive article (http://www.extremetech.com/article2/0,1697,2284042,00.asp) on the new GPU2 client if anyone would like to take a look. There have also been a number of updates to Vijay Pande's (http://folding.typepad.com/) blog in the last few days concerning the new client, for anyone who would like to keep current.

[Kougar]

Raider, it is not exactly a CPU limitation. Actually the code they are using right now is to small to fully max out 3800 GPUs. Please read this post from Vijay's blog, this was dated from yesterday:

There have been some misunderstandings on how the GPU2 core works. In particular, for small proteins like villin on GPU's with large number of stream processors (SP's) like the 3850 or 3870, the protein is too small to use a larger number of SP's unless the CPU is very fast. Some people have guessed that there is some internal SP limit. This is incorrect; the problem is that small proteins can't be parallelized amongst a large number of SP's.

We are working to release larger proteins (about 2x the number of atoms) as they are more interesting scientifically and use the GPU's (even the high end ones) much closer to 100%. The exciting part for us is that the larger proteins run at almost the same speed as the slower ones on GPU's (whereas on CPU's, they're 4x slower); this is where the GPU2 code should shine. In parallel, Mike Houston at AMD is working to optimize CAL such that it has lower CPU overhead.

For now, we're pushing out villin WU's as a test (good to know that the code is working well), but we expect the larger WU's to be going out soon (say a week or two, pending internal testing).

[werty316]

It'll be interesting to see how the new GPU client will evolve down the road.

[raidersforever01]

Nice so more ppd once the bigger WU's come out.