Message boards :
Number crunching :
Experience with opt CUDA versus stock CUDA on my ION
Message board moderation
Author | Message |
---|---|
SciManStev Send message Joined: 20 Jun 99 Posts: 6652 Credit: 121,090,076 RAC: 0 |
Except for the hard drive issues, that is exactly what happened to me with my GTX 295. The screen would blank and recover. Hardly any GPU WU's made it without generating an error. I am draining my cache now, and haven't had a GPU WU in about 6 days, and my RAC is just about where it was when I started draining the cache. I figured that my total RAC, which is totally CPU, would be about the same by using standard apps and my GPU. I will have the twin GTX 480's installed this weekend. Steve Warning, addicted to SETI crunching! Crunching as a member of GPU Users Group. GPUUG Website |
SciManStev Send message Joined: 20 Jun 99 Posts: 6652 Credit: 121,090,076 RAC: 0 |
Except for the hard drive issues, that is exactly what happened to me with my GTX 295. The screen would blank and recover. Hardly any GPU WU's made it without generating an error. I am draining my cache now, and haven't had a GPU WU in about 6 days, and my RAC is just about where it was when I started draining the cache. I figured that my total RAC, which is totally CPU, would be about the same by using standard apps and my GPU. I will have the twin GTX 480's installed this weekend. I would have tried it out, as the instructions were clear, but with the 480's pending, I just gave up and put all my hopes in the 480's. I think that this has been a very useful bit of knowedge learned, on how to run stock CUDA, and optimized MB and AP. My thanks to Sutaru as well. I think many will benefit from that information. As you stated, it worked perfectly with stock apps. Steve Warning, addicted to SETI crunching! Crunching as a member of GPU Users Group. GPUUG Website |
BilBg Send message Joined: 27 May 07 Posts: 3720 Credit: 9,385,827 RAC: 0 |
. ... I will have the twin GTX 480's installed this weekend. AFAIK GTX 480 can't be used for CUDA on SETI now - they need to change the app! http://setiathome.berkeley.edu/forum_thread.php?id=59894 . Â - ALF - "Find out what you don't do well ..... then don't do it!" :) Â |
Sutaru Tsureku Send message Joined: 6 Apr 07 Posts: 7105 Credit: 147,663,825 RAC: 5 |
Sten-Arne, SciManStev, thanks for the flowers! :-) BilBg, I guess SciManStev know this. ;-) |
Jack Shaftoe Send message Joined: 19 Aug 04 Posts: 44 Credit: 2,343,242 RAC: 0 |
It probably depends on the card, driver and BOINC version you are using. I'm using a GTX 260, and optimized CUDA works quite well for me. The only crashing tasks for me are VLAR's. However, I have experienced a few "your display driver stopped working" black screens here and there. That said - it doesn't seem to directly correlate with a crashed task - sometimes when I check BOINC after this happens I have no failed CUDA wu's in the list. Using 197.45 driver. 6.10.43 BOINC |
SciManStev Send message Joined: 20 Jun 99 Posts: 6652 Credit: 121,090,076 RAC: 0 |
My GTX 295 would crash and send back errors regardless of which driver I used, or which version of BOINC. I tried evey combination I could, multiple times, and nothing worked as long as I used the optimized apps. Steve Warning, addicted to SETI crunching! Crunching as a member of GPU Users Group. GPUUG Website |
Fred J. Verster Send message Joined: 21 Apr 04 Posts: 3252 Credit: 31,903,643 RAC: 0 |
[OFF TOPIC] Just started trying the ATI cards (HD5770/HD4850) at MW and desperatly trying one the ATI_AP app.'s at Bêta. [ON TOPIC] It ain't easy, programming a GPU, specially when too little info regarding these card's, is provided by the manufacturer. NVidia has been helpfull, so I heard, but ATI, is a different matter, so it seems. One should think, ATI has something to gain here also. Use BOINC 6.10.43 on 3 QUAD's and 1 LT, used the Unified Installer v.035 (X64). On the Q6600's I run optimized MB (CUDA too). Never changed the driver version, (only back)190.38 NVidida. CAL 1.41 Since the heat problems were solved, they run without error's, if I don't forget to check the REScheduler (1.9) on my XP64(X9650+2xGTS250;4GiG), that is :-) |
Jack Shaftoe Send message Joined: 19 Aug 04 Posts: 44 Credit: 2,343,242 RAC: 0 |
The following are only two examples of many of not VLAR killed task from your CUDA. Typical "opt" crashes: I don't blame the app for those, I was playing with EVGA Precision several times yesterday when overclocking my card. Last week Mon > Thu would be a good time to analyze - was travelling for work and it ran the whole time. Right now, I am looking at my BOINC Manager, 7 Cuda tasks finished. 6 successful, one killed after 4 seconds - most likely VLAR, and one running - 68% complete after 28min. |
Jack Shaftoe Send message Joined: 19 Aug 04 Posts: 44 Credit: 2,343,242 RAC: 0 |
I just set my Network Activity to suspended so I could look at my results tomorrow. My card completes 1 MB every 25-40 minutes so I should have about 40 or 50 complete by midday tomorrow and can update you on number of failures. |
Jack Shaftoe Send message Joined: 19 Aug 04 Posts: 44 Credit: 2,343,242 RAC: 0 |
|
perryjay Send message Joined: 20 Aug 02 Posts: 3377 Credit: 20,676,751 RAC: 0 |
Morning Jack, I like the to completion times on those you still have waiting to run. :-) Tasks that are taking you a half hour to 45 minutes to run are showing a guesstimate of three hours. Have you tried putting a flops count in your app_info? It should bring your DCF way down and you will get a lot more work once the boys from Berkeley get this upload problem sorted out. PROUD MEMBER OF Team Starfire World BOINC |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
However, to each his own. I'm not trying to make everyone revert to stock. For me it worked perfect and I reduced my errors to zero %. V12 is possible to summon driver restarts on lowest-end GPUs more often than stock, cause it has one of thresholds changed that governs choice between 2 different kernels (big and faster one and small and slower one) to do pulse search. small kernel was introduced (though it's slower) to fight with too long kernel execution that causes driver restarts. So definitely your experience can be quite different than mid-range GPU owner. |
Jack Shaftoe Send message Joined: 19 Aug 04 Posts: 44 Credit: 2,343,242 RAC: 0 |
Raistmer, can you confirm - stock app is just as fast as optimized app for cuda? |
Jack Shaftoe Send message Joined: 19 Aug 04 Posts: 44 Credit: 2,343,242 RAC: 0 |
Morning Jack, No, not yet. Can you share details on how to do this PJ? |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
Raistmer, can you confirm - stock app is just as fast as optimized app for cuda? It depends from your hardware & software perhaps. On my own hosts V12 runs faster. Also, it has few changes to save some time here and there that missing in stock app. So I don't see why it should be slower. But for some configs it's possible. As one could know binaries of even slightly different rebuilds can different in speed. Different memory buffers placing and you will get slightly different speed. Newly rebooted system will give different (sometimes pretty noticeable one, not diminishing results) performance when one that runs for few days already with many processes started stopped due run. All these factors adds to error in any benchmarking. So it's question of user preference, if he perceives/measures that one binary faster for his PC, why not use what is faster. Another thing is error rate. As I said already, for very low-end GPUs and ION is such example driver restart rate can be higher than with stock app. When discussed mod was in beta-testing there was no ION card involved. So better participation in testing is required, btw... |
perryjay Send message Joined: 20 Aug 02 Posts: 3377 Credit: 20,676,751 RAC: 0 |
Jack, The instructions are in this thread http://setiathome.berkeley.edu/forum_thread.php?id=53593 once you get the math taken care of, post back if you have trouble figuring out where it goes. PROUD MEMBER OF Team Starfire World BOINC |
©2024 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.