Best GPU Projects?

[Mumak]

So some results

Milkyway@Home:
----------------
HD7950: ~37 / ~47 secs per WU; ~80 / ~107 pts per WU

Einstein@Home BRP4:
---------------------
650 Ti: ~3000 secs for 2 WUs
660 Ti: ~1200 secs for 1 WU, ~2200 secs for 2 WUs
HD7770: ~2300 secs for 1 WU, ~4000 secs for 2 WUs
HD7950: ~1050 secs for 1 WU
Each GPU WU is worth 500 pts

GPUs are clocked:
HD7950 - 1120 MHz
HD7770 - 1120 MHz
660 Ti - default
650 Ti - 1020 MHz

[Gandalf]

Mumak - do you run these WUs virgin or do they go through the app_config file?
If app_config, what settings are you using?

[Mumak]

Mumak - do you run these WUs virgin or do they go through the app_config file?
If app_config, what settings are you using?

For MW@H I run a single WU per GPU, so no app_config.

For E@H there's no need to use app_config for running multiple WUs/GPU, since you can specify this in the E@H Preferences page - "GPU utilization factor of BRP apps" (set to 0.5 for 2 WUs/GPU).

For GPUGrid it doesn't make much sense to run multiple WUs/GPU, since they already utilize the GPU pretty high (>90% depending on GPU). Though I use app_config there to reserve a full CPU core for each WU on Kepler (by default the requested CPU reservation is much lower there, but on Kepler it seems the tasks require much more CPU resources). So there I use:

<app_config>

<app>
<name>acemdlong</name>
<max_concurrent>1</max_concurrent>
<gpu_versions>
<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>

<app>
<name>acemdshort</name>
<max_concurrent>1</max_concurrent>
<gpu_versions>
<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>

<app>
<name>acemd2</name>
<max_concurrent>1</max_concurrent>
<gpu_versions>
<gpu_usage>1</gpu_usage>
<cpu_usage>1</cpu_usage>
</gpu_versions>
</app>

</app_config>

[Snow Crash]

Snow - How's POEM working? any apparent shortages? Is Docking@Home GPU based?

POEM appears to have a 100 WU in process limit - it is not stated on their site but I never get more than that and once I start reporting them in I fill back up to 100. Running 6 at a time via app_config on my 7950 because a single WU only uses 40% GPU ... each additional concurrent WU seems to take about another 10%. Six at a time takes about 26 minutes. Points are pretty good (if that matters) at 2,925.23 per WU but CPU usage is pretty heavy - they have it set to 1 GPU + 1 CPU but I found you really only need between .5 and .75 CPU for each GPU WU.
app_config ...

<app_config>

   <app>
      <!-- the name tag is the name of the application you want to control with this app_config file -->
      <name>poemcl</name>

      <!-- set max_concurrent to the total number of both CPU and GPU WUs to run at a time on your rig -->
      <max_concurrent>5</max_concurrent>

      <!-- gpu_version tag is where you take control of how BOINC allocates your resources for this app -->
      <!-- this line is unnecessary but adds control especially for those with two different cards in a rig -->
      <gpu_versions>

         <!-- gpu_usage calucation is 1 divided by the number of WUs you wish to run on any single card -->
         <!-- don't change this when adding a card as each will try to run this qty (if you have enough CPU) --> 
         <gpu_usage>0.20</gpu_usage>

         <!-- cpu_usage is how much cpu resources BOINC will reserve to run each gpu WU-->
         <!-- calculation is the number of CPU cores/threads for all GPU work divided by the number of GPU WUs -->
         <cpu_usage>0.75</cpu_usage>

      </gpu_versions>

   </app>

</app_config>

Docking@Home is a CPU only project ... I just need 25k to make one of my stat goals (#2 on team XS for number of projects w/ 25K plus)

Docking@Home is a project which uses Internet-connected computers to perform scientific calculations that aid in the creation of new and improved medicines. The project aims to help cure diseases such as Human Immunodeficiency Virus (HIV). Docking@Home is a collaboration between the University of Delaware, The Scripps Research Institute, and the University of California - Berkeley. It is part of the Dynamically Adaptive Protein-Ligand Docking System project and is supported by the National Science Foundation.

[stoneageman]

As team captain for GPUgrid, I'd just like to point out that unlike every other project XS is involved with, we are at #1 position there. So if you have an Nvidia card (500 series or better) come join our success & help keep us there

[Mumak]

As team captain for GPUgrid, I'd just like to point out that unlike every other project XS is involved with, we are at #1 position there. So if you have an Nvidia card (500 series or better) come join our success & help keep us there

Are Noelia tasks finally working there? I had so many issues with those, that I didn't want to continue until it's fixed...

[stoneageman]

I've not seen a Noelia task for 3 weeks

[Gandalf]

As team captain for GPUgrid, I'd just like to point out that unlike every other project XS is involved with, we are at #1 position there. So if you have an Nvidia card (500 series or better) come join our success & help keep us there

Stone man - has Nvidia resolved the one WU per GPU issue?

P.S. - nice avatar!

[Snow Crash]

As team captain for GPUgrid, I'd just like to point out that unlike every other project XS is involved with, we are at #1 position there. So if you have an Nvidia card (500 series or better) come join our success & help keep us there

As soon as I get some personal stats settled, I'll be back with a 670 + 660Ti combo

[Evantaur]

Stone man - has Nvidia resolved the one WU per GPU issue?

you run long tasks with 100% utilization, should not be a problem there

will see if i have money to buy a couple of titans, most likely not

[anubis]

Computing considered - Titan should be the biggest-baddest kid in town at the moment?

[stoneageman]

Stone man - has Nvidia resolved the one WU per GPU issue?

P.S. - nice avatar!

From the testing done by others, overall there is nothing to be gained at present by running more than 1 wu per gpu.

[Mumak]

Titan doesn't work for most projects yet. Seems that current CUDA 4 and 5 apps don't like it and they were not fixed yet. So for example you won't be able to run GPUGrid on Titans.
Though I think Einstein@Home BRP4 app (CUDA 3.1) is able to run on Titans, however the performance is not so great.

Here's a performance comparison of different GPUs for E@H: http://www.dskag.at/images/Research/...rmancelist.pdf

[Evantaur]

you won't be able to run GPUGrid on Titans.

well then http://youtu.be/r-bA9FYB8HY?t=49s sums it up for me pretty well:P

[Mumak]

I checked GPU usage when running MW@H tasks and it seems there might be some room left for running multiple WUs. So I changed it to run 2 WUs/GPU, more doesn't seem to make sense. The app name for app_config is "milkyway".

[Evantaur]

E@h seems to ask 0.5 cores to run well so only running 2 (wcg > everything after all)

[Moo]

supercoolin is tearing up the milkway!!!!

[Mumak]

My 2x7950 are doing ~560,000 pts/day there

[Otis11]

Found another - anyone knowedgeable about Primegrid? Here's all I got:

Primegrid - Cryptography(Trying to find out various different primes in an attempt to improve encryption algorithms, as well as other algorithms)

[D_A]

I ran it for a little while. It's a small budget affair but the work has some important applications if you're into cryptography.