Does the crunching time decrease (for similar WUs) at the higher GPU utilization?

MrS
____________
Scanning for our furry friends since Jan 2002

I meant in his specific case, where GPU utilization increases if he's crunching CPU projects along GPU-Grid.

MrS
____________
Scanning for our furry friends since Jan 2002

This is interesting, as 1.6 GHz is not exactly slow.. and the CPU quickly increases its clock if performance is demanded. Anyway, you might want to either

- crunch in the CPU as well
- create an environment variable "Swan_Sync", set its value to 0 and reboot. Thereby you dedicate a logical core to GPU-Grid, GPU performance should increase.

MrS
____________
Scanning for our furry friends since Jan 2002

Some time ago I speculated that this would be the case. CPU behavior could vary substantially by model. It would be difficult to spot the CPU being used, increasing the clock and then downclocking again, and it might not always occur. I can open large apps without the CPU moving from 1.6GHz to 3.8GHz, as it uses threads from different cores, but sometimes the clock does rise, so it's maybe not even predictable. Using SWAN_SYNC does keep the clocks high. Another thing is that as well as seeing performance reduction from CPU thread saturation, you would see the clocks drop on many models; the various i7 models all have turbo boost steps, which tend to get reduced in increments of usually 100MHz each time. So using one thread might keep full turbo (say 3.8GHz), but this would reduce to 3.4GHz when all the threads are in use (11% drop before even considering thread saturation). Different tasks require different amounts of CPU use too, so the impact of freeing more or less CPU cores, or using SWAN_SYNC varies from task to task.
____________
FAQ's

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

Well.. yes. However, SWAN_SYNC=0 would only use one (logical) core and might increase your GPU speed. Anyway, you could also easily combine both: 7 other CPU threads and one to hardcore-service your GPU. Although I'm not sure how much this stil helps with the current client. And a new one is being beta tested since a few weeks..

MrS
____________
Scanning for our furry friends since Jan 2002

SWAN_SYNC=0 only made ~3% difference on Windows the last time I checked, CUDA 3.1app.
These days it's impact on performance would probably vary more from task to task, but overall be less. I think it was initially ~9%.
SWAN_SYNC=0 doesn't presently work with the CUDA4.2app; Michael said he disabled it, and the task results suggest it was re-enabled.
If it was available it might help keep the GPU clocks high on the Kepler's and the CPU clock high too.
It would certainly need to be tested again, for both Fermi and Kepler cards and for Windows and Linux (when there's a CUDA 4.2 Linux app).

BTW the performance improvement I'm seeing on my CC2.0 Fermi between running tasks on the CUDA4.2 and CUDA3.1 app varies from ~23% to ~56%

IBUCH_3_affTRYP ~23% faster
PAOLA_3EKO ~36% faster
MJHARVEY_MJHXA1 ~56% faster

____________
FAQ's

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