CUDA question.

Message boards : Questions and problems : CUDA question.
Message board moderation

To post messages, you must log in.

AuthorMessage
Christopher Costa

Send message
Joined: 25 Aug 10
Posts: 4
United States
Message 36356 - Posted: 6 Jan 2011, 22:30:37 UTC

I have 256MB of VRAM, but BOINC only reports that I have 244MB. Einstein@home requires 250MB. I know that one way to solve this is to rollback my drivers, but as the drivers that would work are significantly older I would rather not do that. My first question is,is there ANY other way around this without rolling back the drivers, and, why the is there a discrepancy between the VRAM I actually have and what BOINC detects? Thanks.
ID: 36356 · Report as offensive
Richard Haselgrove
Volunteer tester
Help desk expert

Send message
Joined: 5 Oct 06
Posts: 5082
United Kingdom
Message 36360 - Posted: 7 Jan 2011, 0:02:20 UTC - in response to Message 36356.  

I have 256MB of VRAM, but BOINC only reports that I have 244MB. Einstein@home requires 250MB. I know that one way to solve this is to rollback my drivers, but as the drivers that would work are significantly older I would rather not do that. My first question is,is there ANY other way around this without rolling back the drivers, and, why the is there a discrepancy between the VRAM I actually have and what BOINC detects? Thanks.

I suspect that BOINC doesn't detect the VRAM size directly, but instead queries the driver (via the Applications Programming Interface), and reports the answers it gets.

As to why different drivers report a different RAM size, you'll have to ask NVidia. I can only presume that newer drivers reserve a bit of space for their own private use. So, even if BOINC was able in some way to subvert NVidia's programming, and report the raw underlying hardware memory size, it still wouldn't do any good. Either the memory would still be reserved for other uses, and unavailable for Einstein: or you could try to use it anyway, trample all over somebody else's code/data, and probably end up with a crash, bluescreen, or some other nasty.

Also, check reports at Einstein that their app actually uses (maybe only at certain stages of the calculation) significantly more than the 250MB they 'declare' as being required. But AFAIK, those tests haven't yet been rerun for the new version of the application deployed today. You may be in luck with that, but I wouldn't hold my breath.
ID: 36360 · Report as offensive

Message boards : Questions and problems : CUDA question.

Copyright © 2024 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.