rpi_logo
program with CUDA
program with CUDA
log in

Advanced search

Message boards : Number crunching : program with CUDA

Author Message
[af>FRANCE>44>Nantes] Einstein-Rosen-Podolsky
Send message
Joined: 31 Jan 08
Posts: 5
Credit: 17,641,479
RAC: 0

Message 3989 - Posted: 3 Jul 2008, 21:41:23 UTC

Hello,
Know you "CUDA" ? It's a programmation in "C".
For severals months the CUDA is use with GPU Nvidia for to compute for foldingathome project.
With GPU Nvidia the computing is 100* more fast than a CPU.

Have you a project to compute with CUDA ?


http://www.nvidia.com/object/cuda_home.html#

Profile Crunch3r
Volunteer developer
Avatar
Send message
Joined: 17 Feb 08
Posts: 363
Credit: 258,227,990
RAC: 0

Message 4001 - Posted: 5 Jul 2008, 16:31:17 UTC - in response to Message 3989.

Hello,
Know you "CUDA" ? It's a programmation in "C".
For severals months the CUDA is use with GPU Nvidia for to compute for foldingathome project.


Yes i know CUDA... and atm it's pretty much useless for any BOINC project.
The problem is that BOINC does not recognize the GPU as a full "processor" ... at best it would see a "co-processor" which would still require one cpu core to run the gpu application...


With GPU Nvidia the computing is 100* more fast than a CPU.


That statement is completely wrong... by no means a gpu client would be any faster (probably way slower) now.

What they do at folding@home is to issue special wus suited for the GPU and PS3,
without doing that the speed of those apps would be worse than the cpu applications.

HTH




____________

Join Support science! Joinc Team BOINC United now!

Emanuel
Send message
Joined: 18 Nov 07
Posts: 280
Credit: 2,442,757
RAC: 0

Message 4011 - Posted: 7 Jul 2008, 12:20:15 UTC - in response to Message 4001.

I don't think you should write off CUDA quite that quickly. Yes, it could only be used as a co-processor by BOINC, but FAH is in the same situation as CUDA is not yet able to perform all the calculations they need done. Isn't there any significant part of the Milkyway@Home calculations that would benefit from parallelisation? If it could detect whether a CUDA device is present, it could choose between creating a thread that works on the CPU and creating one that works on the CUDA device - as long as initialisation and memory allocation overhead don't destroy the advantages CUDA might have.

Profile Keck_Komputers
Avatar
Send message
Joined: 29 Aug 07
Posts: 20
Credit: 979,019
RAC: 176

Message 4012 - Posted: 7 Jul 2008, 14:12:59 UTC

I don't think Crunch3r is writing CUDA off too quickly, just trying to inject some realism into the marketing hype.

GPUs sacrifice accuracy for speed. This can be fatal for most DC projects.

The quoted speed increases are normally burst speeds for single precision floats. Almost all projects need sustained double precision operations.

For some projects CUDA and other auxiliary resources will be a big help. Major changes in BOINC's scheduling system will make utilizing these resources easier for the projects. Some will be in the 6.2.x client and the rest in the 6.4.x client.
____________
BOINC WIKI

BOINCing since 2002/12/8

Profile Crunch3r
Volunteer developer
Avatar
Send message
Joined: 17 Feb 08
Posts: 363
Credit: 258,227,990
RAC: 0

Message 4016 - Posted: 7 Jul 2008, 20:08:28 UTC - in response to Message 4012.

I don't think Crunch3r is writing CUDA off too quickly, just trying to inject some realism into the marketing hype.

GPUs sacrifice accuracy for speed. This can be fatal for most DC projects.

The quoted speed increases are normally burst speeds for single precision floats. Almost all projects need sustained double precision operations.

For some projects CUDA and other auxiliary resources will be a big help. Major changes in BOINC's scheduling system will make utilizing these resources easier for the projects. Some will be in the 6.2.x client and the rest in the 6.4.x client.


That's exactly the point i was trying to make.

____________

Join Support science! Joinc Team BOINC United now!

Biggles
Send message
Joined: 7 Jun 08
Posts: 2
Credit: 7,342,557
RAC: 0

Message 4138 - Posted: 15 Jul 2008, 12:53:04 UTC - in response to Message 4012.

I don't think Crunch3r is writing CUDA off too quickly, just trying to inject some realism into the marketing hype.

GPUs sacrifice accuracy for speed. This can be fatal for most DC projects.

The quoted speed increases are normally burst speeds for single precision floats. Almost all projects need sustained double precision operations.

For some projects CUDA and other auxiliary resources will be a big help. Major changes in BOINC's scheduling system will make utilizing these resources easier for the projects. Some will be in the 6.2.x client and the rest in the 6.4.x client.


The Radeon HD 4800 series supports double precision, with a few caveats.

The other points stand though, stream processing isn't yet ready for BOINC.

Profile Stefan Ledwina
Avatar
Send message
Joined: 28 Aug 07
Posts: 16
Credit: 27,436,750
RAC: 16

Message 4306 - Posted: 21 Jul 2008, 15:19:56 UTC - in response to Message 4138.
Last modified: 21 Jul 2008, 15:20:43 UTC



stream processing isn't yet ready for BOINC.


Ever heard about gpugrid? ;-)
____________


pixelicious.at - My little photoblog

John McLeod VII
Avatar
Send message
Joined: 27 Aug 07
Posts: 85
Credit: 405,705
RAC: 0

Message 4314 - Posted: 21 Jul 2008, 22:01:58 UTC

There is a planned update to BOINC that allows the CPU scheduler of BOINC to use the GPU (or any other extra processor on the system) if the project specifies its use. This would be instead of, or in addition to using multiple CPUs.
____________


BOINC WIKI

Profile Saenger
Avatar
Send message
Joined: 28 Aug 07
Posts: 130
Credit: 9,837,124
RAC: 17,594

Message 4453 - Posted: 24 Jul 2008, 18:19:05 UTC - in response to Message 4306.



stream processing isn't yet ready for BOINC.


Ever heard about gpugrid? ;-)

Just crunched my first WU.
It needed a core besides the GPU and paused while it wasn't on duty by the resource share.
____________
Grüße vom Sänger

Biggles
Send message
Joined: 7 Jun 08
Posts: 2
Credit: 7,342,557
RAC: 0

Message 4478 - Posted: 25 Jul 2008, 17:07:00 UTC - in response to Message 4306.



stream processing isn't yet ready for BOINC.


Ever heard about gpugrid? ;-)


I'm fully aware of it, but Saenger covered part of it - at the moment a CPU task controls the use of the GPU, which means that when that task isn't running the GPU lies idle.

Add to that only Linux 64-bit support, requiring specific drivers and many cards not being supported and what you find is that stream processing, whilst now a reality, isn't yet ready for BOINC. I'd love it if it was, because there's huge amounts of untouched computing power out there, but it's not quite ready.

John McLeod VII
Avatar
Send message
Joined: 27 Aug 07
Posts: 85
Credit: 405,705
RAC: 0

Message 4482 - Posted: 25 Jul 2008, 22:28:34 UTC - in response to Message 4478.



stream processing isn't yet ready for BOINC.


Ever heard about gpugrid? ;-)


I'm fully aware of it, but Saenger covered part of it - at the moment a CPU task controls the use of the GPU, which means that when that task isn't running the GPU lies idle.

Add to that only Linux 64-bit support, requiring specific drivers and many cards not being supported and what you find is that stream processing, whilst now a reality, isn't yet ready for BOINC. I'd love it if it was, because there's huge amounts of untouched computing power out there, but it's not quite ready.

Scheduling the GPU requires a BOINC client upgrade. The code for that is being worked on.
____________


BOINC WIKI

[af>FRANCE>44>Nantes] Einstein-Rosen-Podolsky
Send message
Joined: 31 Jan 08
Posts: 5
Credit: 17,641,479
RAC: 0

Message 4753 - Posted: 17 Aug 2008, 10:06:52 UTC

OK, so milkyway is not really capable to work in CUDA.

Or optimised some parts of computing for CUDA.

ps: i'm french


Post to thread

Message boards : Number crunching : program with CUDA


Main page · Your account · Message boards


Copyright © 2018 AstroInformatics Group