Message boards : BOINC client : Fermis shader count - how to
Message board moderation
Author | Message |
---|---|
Send message Joined: 19 Aug 08 Posts: 87 ![]() |
The number of shaders are being reported incorrectly for Fermi cards by Boinc. To get the correct shader count multiply the multiprocessors count by 32 for GF100 cards, and by 48 for GF104. GF100 cards are GTX480, GTX470 and GTX465. These have 480, 448 and 352 shaders respectively. The GTX460 has 336 shaders grouped in clusters of (3x16) 48, unlike the GF100 cards that are in groups of 32 (2x16). # Using device 0 # There is 1 device supporting CUDA # Device 0: "GeForce GTX 470" # Clock rate: 1.41 GHz # Total amount of global memory: 1341718528 bytes # Number of multiprocessors: 14 # Number of cores: 112 (this bit is wrong, still) Number should be 14 x 32 = 448. |
![]() Send message Joined: 29 Aug 05 Posts: 15585 ![]() |
Which science application said those things? As BOINC doesn't report any of that. All BOINC does is check for a certain library file on your computer, to generally tell you what GPU you have in your system and report the general information it finds on that. Example given: ATI GPU 0: ATI Radeon HD 4700/4800 (RV740/RV770) (CAL version 1.4.696, 1024MB, 1000 GFLOPS peak) It doesn't precisely say what kind of Radeon I have, it doesn't say how many stream processors are on it. So it's a science application that reported this in your case and then comes the question which drivers you used, which CUDA library files you have installed and are used by the application (some use their own). |
Send message Joined: 5 Oct 06 Posts: 5142 ![]() |
The number of shaders are being reported incorrectly for Fermi cards by Boinc. We got that one sorted out back in March: message 31859. So everything from BOINC v6.10.45 onwards should be reporting correctly. The snip you posted looks like a GPUGrid stderr_out snip: and it was GPUGrid that first alerted BOINC to the mis-reporting problem (as you can see from the thread I've linked). You can also see that you were an active participant in that earlier thread. So I'm not sure what the new problem is, exactly. |
Send message Joined: 19 Aug 08 Posts: 87 ![]() |
Sorry, the issue is with the apps/driver/cuda, not Boinc. |
Send message Joined: 19 Aug 08 Posts: 87 ![]() |
NVIDIA GPU 0: GeForce GTX 460 (driver version 25856, CUDA version 3010, compute capability 2.1, 738MB, 363 GFLOPS peak) - The reported 363 GFLOPS peak seems a bit low for a card with 336shaders and 675MHz cores. It should be 907. http://en.wikipedia.org/wiki/GeForce_400_Series |
Send message Joined: 9 Sep 10 Posts: 1 ![]() |
I just posted the same question to the BOINC Alpha mailing list about the 363 GFLOPS that should be reported as 907. What determines that number? It can't be the driver, can it? I've gone through several versions and now I'm running on the leaked 260.52 drivers for my 460 and 6.11.7 still reports that number as 363. |
![]() Send message Joined: 29 Aug 05 Posts: 15585 ![]() |
I just posted the same question to the BOINC Alpha mailing list about the 363 GFLOPS that should be reported as 907. What determines that number? It can't be the driver, can it? I've gone through several versions and now I'm running on the leaked 260.52 drivers for my 460 and 6.11.7 still reports that number as 363. Until Nvidia releases an API that tells external programs how many shader processors per GPU their cards have, it's very difficult to give out correct values. At this moment BOINC estimates that all GF10x cards have 480 (or 48) shader processors. This is wrong, since some have different numbers. But until Nvidia releases a file (in programming language called an API (application programming interface)) that says exactly which GPU has how many, it is undoable for the developers guestimate the correct value of peak GFlops. Adding the values into an internal file is undoable as that would mean that they had to release a new BOINC at each time that Nvidia releases a new derivation of their videocards. And even then it's not a good idea, seeing how it can even depend on which 460 you bought to get to the correct number, as there are 2 versions of the GPU available. So you all have to wait, or complain at the Nvidia developer forums. |
Copyright © 2025 University of California.
Permission is granted to copy, distribute and/or modify this document
under the terms of the GNU Free Documentation License,
Version 1.2 or any later version published by the Free Software Foundation.