Native GPU client in GPU3?

Moderators: slegrand, Site Moderators, PandeGroup

GPU3 native Linux support via OpenMM?

Postby p2501 » Thu May 27, 2010 5:39 pm

Any chances we could have that?
p2501
 
Posts: 140
Joined: Sun May 04, 2008 9:10 am

Re: GPU3 native Linux support via OpenMM?

Postby 7im » Thu May 27, 2010 5:44 pm

You'd need more up to date linux drivers from NV first.
How to provide enough information to get helpful support
Tell me and I forget. Teach me and I remember. Involve me and I learn.
User avatar
7im
 
Posts: 15235
Joined: Thu Nov 29, 2007 4:30 pm
Location: Arizona

Re: GPU3 native Linux support via OpenMM?

Postby Hyperlife » Thu May 27, 2010 6:05 pm

Dr. Pande has mentioned previously on a few occasions that he'd like to get OS X and Linux GPU clients out, after the Windows client stabilizes. That may be a while still, unfortunately. :(
Image
User avatar
Hyperlife
 
Posts: 447
Joined: Sun Dec 02, 2007 7:38 am

Re: Native GPU client in GPU3?

Postby p2501 » Fri May 28, 2010 5:48 am

Just lol. You guys crack me up, one asks about availablility of a native client and you move/merge it to a virtualization subforum. Very funny. Is that an official statement? Like "you guys have your wine thing going on, you won't ever get a real client so shut up already and crawl back into your hole"?

I don't care who said what in the past, only thing matters here is that Folding is missing out on performance because of this developement speed.
p2501
 
Posts: 140
Joined: Sun May 04, 2008 9:10 am

Re: GPU3 native Linux support via OpenMM?

Postby Mindmatter » Fri May 28, 2010 7:26 am

7im wrote:You'd need more up to date linux drivers from NV first.


The only requirement I've seen so far from Stanford is CUDA 3.0 and Nvidia has supported that on all platforms since November. They have even had drivers and libraries for OpenCL since September. Once again on all platforms. Either Stanford is waiting to make sure there are no show stoppers in the Windows GPU3 and then will release Linux and Mac CUDA 3 based clients, or they are waiting for the OpenMM/OpenCL client to be finished before expanding on to other platforms. The waiting for OpenMM part might make sense. They are making it sound like OpenMM will be the core for their clients for years to come, so why port to CUDA now when OpenMM is just around the corner. Unless they are planning on supporting both CUDA and OpenMM for the foreseeable future depending on WU requirements. Whatever happens I feel like we will get our native client fairly soon.

http://developer.nvidia.com/object/cuda ... loads.html
Mindmatter
 
Posts: 48
Joined: Tue May 27, 2008 1:53 pm

Re: Native GPU client in GPU3?

Postby bruce » Fri May 28, 2010 1:53 pm

p2501 wrote:I don't care who said what in the past, only thing matters here is that Folding is missing out on performance because of this developement speed.


That's true, but Stanford is developing OpenMM. Other people are developing the OpenCL support. Your complaint isn't in a place where those who build OpenCL for their GPU hardware are not reading this forum.

I'm not a developer, but it's certainly reasonable to suspect that the OpenCL environment is not sufficiently robust for them to produce a reliable port to OpenCL yet. OpenCL currently supports relatively simple activities and FAH is not a simple activity.

OpenMM has been ported to CUDA because that allows them to validate OpenMM outside of the lab. In the long run, the early port is wasted work, but CUDA is a dependable interface and this does allow Stanford to continue to develop OpenMM while waiting for an OpenCL environment that's robust enough to support challenging tasks.

Stanford spent a lot of time trying to get FAH-GPU to work on DirectX and in the final analysis, DX works well for graphics but not for serious computing. A lot depends on how well the DX compiler (or CUDA or OpenCL) can deal with code that creates complex GPU operations and how quickly the compiler developers respond to bug reports. I'm sure it also depends on how well the OpenCL compiler can produce well optimized GPU code for GPU-A or for GPU-B or for GPU-C. Even if it gets the correct results, if it's terribly inefficient (slow) nobody is going to use it.

Again: I'm not a developer so a lot of this is supposition, but Stanford is not going to release a product that isn't ready, and a CUDA port is one step in the right direction although it's not a final step of any kind.
bruce
 
Posts: 21005
Joined: Thu Nov 29, 2007 10:13 pm
Location: So. Cal.

Re: Native GPU client in GPU3?

Postby Sidicas » Fri May 28, 2010 3:09 pm

bruce wrote:
I'm not a developer, but it's certainly reasonable to suspect that the OpenCL environment is not sufficiently robust for them to produce a reliable port to OpenCL yet. OpenCL currently supports relatively simple activities and FAH is not a simple activity.

As far as I can tell with the OpenCL stuff that I've run... The robustness is definitely there already...
The problem, IMO, is performance.. I haven't seen a single OpenCL app that wouldn't have run faster if it was written in CUDA instead.

Right now, from the public developer tools that I'm using... You could write something in OpenCL and it will run on everything and be much easier to code.. Or you could write it in CUDA and have it limited to nvidia GPUs but run 2x-10x faster.. I have yet to see a single example where OpenCL performance = CUDA performance, in the overwhelming majority of developer examples provided, it's not even close to matching CUDA performance..

Hopefully compilers will improve, but right now OpenCL is just a huge trade-off.. You might gain code flexibility, but you're sacrificing efficiency and performance... Stuff written in OpenCL have much less bang in them...That's pretty evident.. Still as of right now, all the cool new bleeding edge stuff nvidia demos is written in CUDA first for obvious reasons..

Not sure how OpenCL is on ATI right now though so maybe somebody else can chime in who's using ATI's new development suite.
Sidicas
 
Posts: 233
Joined: Sun Feb 17, 2008 4:46 pm

Re: Native GPU client in GPU3?

Postby bruce » Fri May 28, 2010 6:27 pm

Sidicas wrote:As far as I can tell with the OpenCL stuff that I've run... The robustness is definitely there already...

... You could write something in OpenCL and it will run on everything and be much easier to code.. Or you could write it in CUDA and have it limited to nvidia GPUs but run 2x-10x faster...


Maybe the robustness is lacking for complex code or maybe it's already there. Let's assume you are 100% correct. Do you think that releasing a version of OpenMM which run 2x to 10x SLOWER than the one they just released makes any sense at all? Clearly the donors would object violently. The new code has to be an improvement over existing code or there's no point in releasing it.

The only possible advantage to an OpenCL implementation with rotten performance would be for someone who has a GPU which supports OpenCL which is not currently supported by FAH - - i.e.- something other than ATI or NV. -- but that also assumes there is a large quantity of untapped GPU resources that could be used by FAH. That also assumes that the results obtained from brand-X GPUs is better performance in PPD/Watt than just running the CPU client. That also assumes that everything is bug-free.

If you split up the problem into pieces, you've got OpenCL for ATI, OpenCL for NV, OpenCL for brand-X (and multiply that by Windows/Linux/MacOS) and you've also got OpenMM. If all 9+ versions of OpenCL fall into the "wait until performance is fixed" category, you can still field-test OpenMM on top of CUDA. All you need is a OpenMM-CUDA interface which is probably very very close to an OpenMM-OpenCL interface that has already been written. Proceed directly to a field test of OpenMM.

Besides the field-test of OpenMM, the support for Fermi gets into production without waiting for OpenCL.
bruce
 
Posts: 21005
Joined: Thu Nov 29, 2007 10:13 pm
Location: So. Cal.

Re: Native GPU client in GPU3?

Postby powerarmour » Fri May 28, 2010 6:38 pm

What about just releasing a OpenMM/CUDA version for Linux?, OpenCL can wait if it's not ready.
Image
powerarmour
 
Posts: 173
Joined: Wed Oct 29, 2008 1:00 am
Location: Surrey, UK

Re: Native GPU client in GPU3?

Postby bruce » Fri May 28, 2010 8:06 pm

powerarmour wrote:What about just releasing a OpenMM/CUDA version for Linux?, OpenCL can wait if it's not ready.


Hasn't that already been answered?
1) Linux drivers are not at the minimum level. New versions of Windows drivers are being released at a prodigious rate, which probably means NV is still finding a lot of bugs either in CUDA 3 or in the drivers so the minimum level for FAH stability is probably still somewhere in the future, not somewhere in the past.
2) It's not ready for release anywhere. The current client is beta. During beta, they try to shake out most of the bugs on one platform before starting a beta on another platform.

My hunch is that there will be a Linux-CUDA-NV version before there's a (yourOS)-OpenCL-(yourGPU) version, but that like predicting which development path is going to win a race and I don't have any useful handicapping information. (Of course everybody with ATI hardware is betting on (yourOS)-OpenCL-ATI beating out Linux-CUDA.)
bruce
 
Posts: 21005
Joined: Thu Nov 29, 2007 10:13 pm
Location: So. Cal.

Re: Native GPU client in GPU3?

Postby powerarmour » Fri May 28, 2010 10:12 pm

bruce wrote:Hasn't that already been answered?


Apologies and thanks for the info, though I'm sure I won't be the last person to ask that question! :)

I'm just getting twitchy to give something new for my Linux machines to do that's all, I don't mind being patient as long as a nice usable native client comes out sometime... :wink:
powerarmour
 
Posts: 173
Joined: Wed Oct 29, 2008 1:00 am
Location: Surrey, UK

Re: Native GPU client in GPU3?

Postby dashaund » Wed Jan 12, 2011 9:04 pm

Any new updates on this? I've been having a time trying to get the Windows GPU3 client to run under Ubuntu. I've tried 3 different version (9.04, 10.4, & 10.10), and different CUDA version (3.0 & 3.2), different wrapper versions (3.0, 3.1, 3.2), and two different Wine version (1.2 & 1.3). I finally got it to fold, but then it locks up Gnome, even with auto-mator. Unfortunately, I've had to go back to Windows 7, even though I REALLY want to use Ubuntu 10.10. GPU folding is the killer app for me. Please, Pande project! Let's get a native Linux GPU folding app released! It seems as if CUDA is stable enough on Linux to be able to do this now.
dashaund
 
Posts: 1
Joined: Mon Jan 10, 2011 4:09 pm

Re: Native GPU client in GPU3?

Postby Sidicas » Mon Feb 21, 2011 5:00 pm

dashaund wrote:Any new updates on this? I've been having a time trying to get the Windows GPU3 client to run under Ubuntu. I've tried 3 different version (9.04, 10.4, & 10.10), and different CUDA version (3.0 & 3.2), different wrapper versions (3.0, 3.1, 3.2), and two different Wine version (1.2 & 1.3). I finally got it to fold, but then it locks up Gnome, even with auto-mator. Unfortunately, I've had to go back to Windows 7, even though I REALLY want to use Ubuntu 10.10. GPU folding is the killer app for me. Please, Pande project! Let's get a native Linux GPU folding app released! It seems as if CUDA is stable enough on Linux to be able to do this now.


Define "locks up gnome".. You mean your computer freezes or does gnome become less responsive? Does it respond to Ctrl+Al+Backspace? Have you tried increasing the priority of XServer? Are you using Compiz Fusion? What graphics card do you have? I have a lot of suspicions, the first of course being that you might be running out of video memory and are having a huge amount of bus traffic to/from RAM because of it (using system memory as GPU memory).. Gnome these days has a lot of hardware accelerated graphics and can take a sizable chunk out of your GPU's memory. KDE4 is the same way really... In fact, all Desktops these days including Windows are going the way of hardware accelerated graphics that eat up video memory.. It might just be as simple as a gnome configuration setting.

I think we can figure that problem out.. Feel free to start a new thread here...

Native client would be nice, but really it would be the same thing the wrapper is already doing. The only difference is it would just have a much more user friendly installation..

The wrapper just connects libraries.. Which unfortunately means that you'd likely have this exact same problem with a native Linux GPU client. The only thing that should be different from Windows is a slightly higher CPU usage which is due to wrapper overhead.. Anything else are most likely Linux specific issues that we should be able to resolve..
Sidicas
 
Posts: 233
Joined: Sun Feb 17, 2008 4:46 pm

Re: Native GPU client in GPU3?

Postby VijayPande » Mon Feb 21, 2011 11:48 pm

This is currently a lower priority, below getting v7 out and an OpenCL core for ATI. Once those items are out and in good shape, I think it would be a good time to for us to reconsider this one.
Prof. Vijay Pande, PhD
Departments of Chemistry, Structural Biology, and Computer Science
Chair, Biophysics
Director, Folding@home Distributed Computing Project
Stanford University
User avatar
VijayPande
Pande Group Member
 
Posts: 2662
Joined: Fri Nov 30, 2007 6:25 am
Location: Stanford

Re: Native GPU client in GPU3?

Postby kromberg » Wed Mar 23, 2011 6:11 pm

VijayPande wrote:This is currently a lower priority, below getting v7 out and an OpenCL core for ATI. Once those items are out and in good shape, I think it would be a good time to for us to reconsider this one.


Any idea when that would be?
kromberg
 
Posts: 90
Joined: Sat Nov 07, 2009 4:36 pm

PreviousNext

Return to unOfficial Linux GPU (WINE wrapper) (3rd party support)

Who is online

Users browsing this forum: No registered users and 1 guest

cron