Forums :
Wish list :
CUDA
Message board moderation
Author | Message |
---|---|
![]() Send message Joined: 16 Jul 07 Posts: 3 Credit: 2,025,631 RAC: 0 |
Your university is the first CUDA Excellence Center in the world since a few days and receive $500.000 from nVidia. Do you plan to put Cosmology@Home on CUDA ? |
![]() Volunteer moderator Project administrator Project developer ![]() Send message Joined: 1 Apr 07 Posts: 662 Credit: 13,742 RAC: 0 |
Your university is the first CUDA Excellence Center in the world since a few days and receive $500.000 from nVidia. Do you plan to put Cosmology@Home on CUDA ? Are there currently plans to do this? No. Have we been discussing this for the future? Definitely. Scott Kruger Project Administrator, Cosmology@Home |
![]() Send message Joined: 16 Jul 07 Posts: 3 Credit: 2,025,631 RAC: 0 |
Good. Thank you for your quick answer. |
jziembinski Send message Joined: 26 Oct 07 Posts: 2 Credit: 58,650 RAC: 0 |
thats mean one thing - advancement... |
![]() Send message Joined: 29 Oct 07 Posts: 3 Credit: 4,022,806 RAC: 0 |
some news of computing on CUDA ? Or on OpenCL ? With the third version of CUDA wich come Q4 2009, we can use CUDA and OpenCL CUDA is Nvidia and OpenCL is ATI. |
![]() Volunteer moderator Volunteer tester ![]() Send message Joined: 15 Jun 07 Posts: 345 Credit: 50,500 RAC: 0 |
CUDA is Nvidia Correction, OpenCL is a standard on its own. CUDA will eventually do OpenCL as well, possibly already at CUDA 2.2 (still in beta). CAL is AMD/ATI at this moment, but that's not supported. ATI support will be added to BOINC 6.8, but under what form exactly is still unknown. |
![]() Send message Joined: 29 Oct 07 Posts: 3 Credit: 4,022,806 RAC: 0 |
i think for cosmology it's like milkyway, for a best precsion we should use RADEON 4770 or 48xx, etc, graphics cards works in double precision. is'nt it ? |
Will Send message Joined: 15 Jun 09 Posts: 2 Credit: 864,121 RAC: 0 |
The ATI GPU programming system is now called Stream. Would like to see Cosmologyathome pursue both CUDA and Stream algorythms. They have a Stream system for Milkyway@home and it processes what have been 2-3 hour work units in ~2 minutes. That is some huge improvement in productivity. Cheers. |
Will Send message Joined: 15 Jun 09 Posts: 2 Credit: 864,121 RAC: 0 |
Nevermind! Nothing worse than a know-it-all that knows nothing. I see that CAL is the language for the "Stream" system. Sorry for the "correction". Will is now going to go kick himself :) A little knowledge is a dangerous thing...and I have so little knowledge. |
JLConawayII Send message Joined: 27 Apr 10 Posts: 19 Credit: 1,086,670 RAC: 0 |
I would also love to see CUDA support for this project. There are certain less worthy CUDA projects running on my systems that I would gladly replace with this one. |
![]() ![]() Send message Joined: 19 Jul 08 Posts: 19 Credit: 2,963,018 RAC: 0 |
+1 for GPU processing :) |
.clair. Send message Joined: 4 Nov 07 Posts: 629 Credit: 12,068,402 RAC: 0 |
It would take a very big rewrite of the software, and it depends if the calculations being carried out by it would gain benifit from masivly multithreded processing that a GPU can do. |
kararom Send message Joined: 9 Jan 09 Posts: 69 Credit: 29,506,700 RAC: 0 |
--1 for GPU processing... |
![]() Send message Joined: 26 Oct 11 Posts: 55 Credit: 568,606 RAC: 858 |
I think I've seen some Fortran compilers that can compile to CUDA. Not very cheap, though, so it might take significant donations first. The main feature needed for programs that can run well on GPUs is many sections of the program that can run in any order because none of them depend on the results of the sections that will run at the same time. If you need double precision, expect the suitable Nvidia-based boards to be restricted to be GTX 2xx series, just one laptop-compatible board, and the desktop-compatible boards based on Fermi chips. |
![]() Send message Joined: 26 Oct 11 Posts: 55 Credit: 568,606 RAC: 858 |
I just finished a web search for such compilers. Apparantly, there is only one available, and their pricing is complicated enough that negotiating with them about BOINC use is probably needed. http://www.pgroup.com/resources/cudafortran.htm http://developer.nvidia.com/cuda-fortran http://forums.nvidia.com/index.php?showtopic=156856 Howerver, a Fortran to CUDA translation program is apparantly available: http://www.esrl.noaa.gov/gsd/ab/ac/F2C-ACC.html |
![]() Volunteer moderator Project administrator Project scientist ![]() Send message Joined: 24 Jun 07 Posts: 192 Credit: 15,273 RAC: 0 |
Hi - Thanks for these pointers! I did not know about the Fortran to CUDA translator - will have to look into that. Thanks, Ben Creator of Cosmology@Home |
![]() Send message Joined: 26 Oct 11 Posts: 55 Credit: 568,606 RAC: 858 |
For any users who want to run OpenCL GPU workunits: The 7.0.* series of BOINC versions appear to be the first ones that support doing this very well. All still in beta test, though. I've been using 7.0.3 lately while waiting for some sign that a later version is better - such as release from beta testing or some BOINC project I'm participating in saying that they need some feature from a later version. |