Huh. That's unexpected. What driver version do your machines have?

Matt

One of the problems may be that BOINC for Linux never seems to have been able to extract the version numbers from CUDA drivers and report it to the servers.

One of the problems may be that BOINC for Linux never seems to have been able to extract the version numbers from CUDA drivers and report it to the servers.

Well, that might account for it, I'll have to see what values the scheduler is working with.

In any case, Stoneageman: try version 821, to rule out a possible app build problem.

Matt

Huh. That's unexpected. What driver version do your machines have?

Matt

304.88 for my 500 & 600 series cards & 331.20 for the 780's. I'm now trying 331.49 on a 780 with cuda6 & it looks to be at least 5% slower than 331.20.

With the recent drivers, performance can be capped due to a lack of GPU utilization; I'm presently running a GERRARD Long run task on CUDA 6 (albeit on a GTX770 and W7) and the clocks are 1045MHz. Power is at 40%. It's normally OC'ed to ~1250MHz. Even with only 3 CPU threads in use for CPU projects and the GPU Utilization rising to 80% the GPU remains in that non boost clock rate (it drops but doesn't rise). It's a known issue, but it's hard to spot - might not happen for days if you only receive some WU types. So far I have only seen this on the first GPU, and rarely. This driver 'feature' might also be applicable to Linux.

I'm using 331.49 on Linux with a GTX670 and a GTX660Ti. I get the odd failure on that system, but so far I've not seen any issues with the few tasks that ran on CUDA6. Recently completed a Gianni WU using v8.20 (cuda6.0), presently running a NATHAN_KID using 8.20 (cuda6.0).
Going by NVidia X server the GTX670 GPU clock varies from 1071 to 1110MHz (with Prefer Maximum Performance selected).

I have not received any cuda6 WU's on my other Linux system with an earlier driver.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

I don't believe it! Just installed 331.49 on my 4x780 machine & now it's downloading 8x cuda5.5 tasks :-(

Was going to suggest using the Beta queue, but at least they won't all fail overnight!

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Looks like Matt has updated the cuda 6 app from 8.20 to 8.21

With the recent drivers, performance can be capped due to a lack of GPU utilization

Bring back swan_sync :-)
I've gone back to 331.20 & now just getting the occasional cuda6 task.

With the recent drivers, performance can be capped due to a lack of GPU utilization

Bring back swan_sync :-)
I've gone back to 331.20 & now just getting the occasional cuda6 task.

I don't think swan_sync has anything to do with maintaining max boost during situations where GPU utilization is low enough that it downclocks.

If you are looking at maintaining max boost, please have a look at my thread here, which describes a method to do it on Windows.
http://www.gpugrid.net/forum_thread.php?id=3647
It might be adapted to be usable for linux, somehow, if you can find a way to force the P2 power state.

Looks like Matt has updated the cuda 6 app from 8.20 to 8.21

pvh in number crunching is reporting that v8.21 is failing all tasks on GTX 550 Ti and GTX 660 Ti with driver 331.20 and 64-bit Linux (openSUSE 13.1).

8.21 also failed on my Ubuntu 13.10 system (331.49 drivers and either or both of a GTX670 and a GTX660Ti). Failures were quick:

I435-NATHAN_KIDc22_full3-4-10-RND0271_0 5588452 159186 6 Apr 2014 | 18:33:13 UTC 6 Apr 2014 | 18:36:13 UTC Error while computing 79.69 19.83 --- Long runs (8-12 hours on fastest card) v8.21 (cuda60)

I1R7-NATHAN_KIDc22_glu3-4-10-RND0934_0 5588270 159186 6 Apr 2014 | 17:24:38 UTC 6 Apr 2014 | 18:33:13 UTC Error while computing 11.17 2.92 --- Long runs (8-12 hours on fastest card) v8.21 (cuda60)

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help