BRP3 CUDA application 1.06 for Linux released

Oliver Behnke
Oliver Behnke
Moderator
Administrator
Joined: 4 Sep 07
Posts: 942
Credit: 25166626
RAC: 0
Topic 195589

We just released a 32-bit Linux CUDA binary. It uses the same codebase as the Windows release, but contains a workaround for a scheduling bug in the current NVIDIA CUDA driver. This workaround leads to the fact that the app has to use a full CPU core, instead of just ~20%. While this is clearly not optimal it's probably better than having no Linux CUDA app at all (remember: the old ABP2 CUDA also used a full core but this app is much more efficient). As soon as NVIDIA released their next driver update (approx. 2-3 months from now) we'll switch back to the regular implementation that uses as less of a CPU core as possible.

In the meantime we'll continue to improve the GPU usage and accuracy. We hope to be able to release a Mac OS CUDA version in the not too distant future. We also started to look into an OpenCL implementation which will eventually enable the use of ATI/AMD GPUs for Einstein@Home.

Cheers,
Oliver

 

Einstein@Home Project

M. Schmitt
M. Schmitt
Joined: 27 Jun 05
Posts: 478
Credit: 15872262
RAC: 0

BRP3 CUDA application 1.06 for Linux released

Quote:

We also started to look into an OpenCL implementation which will eventually enable the use of ATI/AMD GPUs for Einstein@Home.

Cheers,
Oliver

Great news!
This will have influence to my decision what new card I will buy. :)

.clair.
.clair.
Joined: 20 Nov 06
Posts: 62
Credit: 1051176770
RAC: 0

I already have ATI 4650

I already have ATI 4650 cards,
So when it is written it will depend on the minimum specs for the app to run on them.
Dose einstein need double precision floating point,
I think that started with the 4800 series :(

Saenger
Saenger
Joined: 15 Feb 05
Posts: 403
Credit: 33009522
RAC: 0

RE: I already have ATI 4650

Quote:
I already have ATI 4650 cards,
So when it is written it will depend on the minimum specs for the app to run on them.
Dose einstein need double precision floating point,
I think that started with the 4800 series :(


It doesn't need for Cuda, otherwise I couldn't crunch with my GT240, so I don't see any need for OpenCL.

Grüße vom Sänger

M. Schmitt
M. Schmitt
Joined: 27 Jun 05
Posts: 478
Credit: 15872262
RAC: 0

RE: ... so I don't see any

Quote:
... so I don't see any need for OpenCL.

I definitively will buy an ATI/AMD card for my Linux workplace and if it can't be used here I will go somewhere else, where double precision is needed. Beside that a lot of people own high speed AMD/ATI cards. :)
Afaik the low end 6850 cards support only single precision too.

But we are getting off topic anyway. ;)

Kind regards

telegd
telegd
Joined: 17 Apr 07
Posts: 91
Credit: 10212522
RAC: 0

Much appreciated for the

Much appreciated for the Linux Cuda 32bit app.

I am getting an error message:
"CUDA version 3.1 capability needed"

My system self-reports:
"NVIDIA GPU 0: GeForce GT 220 (driver version unknown, CUDA version 3010, compute capability 1.2, 1024MB, 131 GFLOPS peak)"

Is there something wrong on my end? I am running Kubuntu Lucid on the Nvidia 256.53 driver (the last known good one that doesn't error out on Primegrid).

Saenger
Saenger
Joined: 15 Feb 05
Posts: 403
Credit: 33009522
RAC: 0

RE: Is there something

Quote:
Is there something wrong on my end? I am running Kubuntu Lucid on the Nvidia 256.53 driver (the last known good one that doesn't error out on Primegrid).


Quote from the thread BRP CUDA requirements:

Quote:
The BRP CUDA tasks require the following:
- CUDA driver >= 3.2 (>= 260.00)

Grüße vom Sänger

Oliver Behnke
Oliver Behnke
Moderator
Administrator
Joined: 4 Sep 07
Posts: 942
Credit: 25166626
RAC: 0

RE: Much appreciated for

Quote:

Much appreciated for the Linux Cuda 32bit app.

I am getting an error message:
"CUDA version 3.1 capability needed"

My system self-reports:
"NVIDIA GPU 0: GeForce GT 220 (driver version unknown, CUDA version 3010, compute capability 1.2, 1024MB, 131 GFLOPS peak)"

Is there something wrong on my end? I am running Kubuntu Lucid on the Nvidia 256.53 driver (the last known good one that doesn't error out on Primegrid).

The error message contained a wrong CUDA version number. We fixed the scheduler and it now correctly reports "CUDA version 3.2 capability needed".

Cheers,
Oliver

 

Einstein@Home Project

Metod, S56RKO
Metod, S56RKO
Joined: 11 Feb 05
Posts: 135
Credit: 808430349
RAC: 65800

RE: We just released a

Quote:
We just released a 32-bit Linux CUDA binary.

Please consider releasing binary which would depend on less shared libraries than the current one. On 64-bit linux some distributions don't include many 32-bit libraries, see description of my problem here.

Metod ...

Oliver Behnke
Oliver Behnke
Moderator
Administrator
Joined: 4 Sep 07
Posts: 942
Credit: 25166626
RAC: 0

RE: Please consider

Quote:

Please consider releasing binary which would depend on less shared libraries than the current one. On 64-bit linux some distributions don't include many 32-bit libraries, see description of my problem here.

We try of course to use as many static libs as possible. I'll look into this and see if things can be further improved.

Oliver

 

Einstein@Home Project

telegd
telegd
Joined: 17 Apr 07
Posts: 91
Credit: 10212522
RAC: 0

RE: The error message

Quote:

The error message contained a wrong CUDA version number. We fixed the scheduler and it now correctly reports "CUDA version 3.2 capability needed".

Cheers,
Oliver


Thanks for the clarification. I hadn't seen the required specs on the other thread. I am happy to say that the current Nvidia 260.19.29 (Ubuntu) driver works for PrimeGrid as well as E@H.

Although it still uses 1 CPU core, the speed is pretty decent. I am getting through a WU in about 6400 sec with a basic GT220 (and 1 core of an i7-860).

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.