Yesterday I upgraded my GPU to a factory over clocked GTX 660Ti.
I've run about 15 units, two at a time, and observed a speedup of about 700s compared to the BRP4 v1.25 on Einstein. At the same time the CPU-time per result has decreased by about 400s, thats more than half on my system.
v1.28 on Albert, x2, run time ~2170s and CPU time ~360s.
v1.25 on Einstein, x2, run time ~2900 and CPU time ~770s.
I ran the new CUDA 1.28 app via one of my Windows systems today. I have not been able to get much work today but the two tasks that ran via my GTX 580, completed at 834 seconds each. This is with one task running at a time. GPU load was at approximately 90-91% while running one task.
If memory serves me right, the previous application ran at around 1360 seconds per task with the 1.25 app via this system. This is a very decent improvement in performance. Thanks for the work put into optimizing the BRP4 applications.
Thanks for the feedback. That's actually a bit more of a speedup than I had expected based on some tests on slower hardware. Definitely in relative terms, the speedup is more pronounced on faster cards.
I will now install the Linux CUDA app on Albert, stand by for more tests. I'm eager to see those GTX 680 .... ;-)
How bout a single work unit?I want to see if a 660Ti is faster than a 580gtx/
Got hold of 2 BRP4-tasks and ran them one at a time on my over clocked GTX660Ti (Core-[at]-1201.9 MHz).
Run time: 1183.69 and 1175.39 seconds for an average of 1179.54 s.
GPU load ~84%.
This on Win7 x64, PCI-E 3.0x16.
Here are some preliminary numbers for the GTX 680.
One task per GPU
System #1 - Single GPU
x16 3.0 - 721 seconds
System #2 - Multi GPU
x16 3.0 - 785 seconds
x8 3.0 - 901 seconds
Overall, the performance looks great so far. I want to do some more testing with multiple tasks running at once, different PCI-E configurations, and with the CPU dedicated for BRP4 GPU only. The above tests were done with ~50% CPU load from running other CPU tasks at the same time.
I have a possible finding--not even a little bit sure--but am posting in case others might spot such a thing.
I've got two different hosts with the same GPU, a GTX 460. Neither has had downclocking problems for some weeks, but I found both downclocked severely today, with the problem persisting through system reboot.
It might just barely be possible that running the current Albert BRP1.28 CUDA ap, or on an even less likely note the Albert 0.29 Gamma Ray Pulsar application--or switching back and forth from those to the current Einstein applications was involved.
More likely something else in my system's history was the problem, but I thought I'd post the suspicion in case someone else sees something.
I'm not even sure what the true downclock frequency was, as different sources reported different numbers, but it was either 405 MHz or less--the reduction in power consumption and GPU temperature, while reporting exceedingly high GPU utilization, but making very slow progress on the WU was persuasive that downclocking was at hand.
I have sometimes the same issue.It only happens when i overclock the gpu +130Mhz from baseline.After several hours the screen goes blank and with gpu-z it shows core speed of 400Mhz then i proceed with reboot and everything goes to normal.
I think it some fail safe method that Nvidia uses.Also I noticed when boinc is not running the core speed drops to 50mhz and goes to 800 instantaneous if a demanding gpu work is needed
Yesterday I upgraded my GPU
)
Yesterday I upgraded my GPU to a factory over clocked GTX 660Ti.
I've run about 15 units, two at a time, and observed a speedup of about 700s compared to the BRP4 v1.25 on Einstein. At the same time the CPU-time per result has decreased by about 400s, thats more than half on my system.
v1.28 on Albert, x2, run time ~2170s and CPU time ~360s.
v1.25 on Einstein, x2, run time ~2900 and CPU time ~770s.
GPU-load also increased from ~80% to 95+%.
Great work!
How bout a single work unit?I
)
How bout a single work unit?I want to see if a 660Ti is faster than a 580gtx/
RE: How bout a single work
)
I'd like to test that but as of right now I can't get any more work for BRP4, probably because the server status pages says 0 tasks to send...
BRP3Cuda32 GTX550Ti 1793 /
)
BRP3Cuda32
GTX550Ti 1793 / 303 single wu
i3 win7/64 7.0.31 x16 slot
I ran the new CUDA 1.28 app
)
I ran the new CUDA 1.28 app via one of my Windows systems today. I have not been able to get much work today but the two tasks that ran via my GTX 580, completed at 834 seconds each. This is with one task running at a time. GPU load was at approximately 90-91% while running one task.
If memory serves me right, the previous application ran at around 1360 seconds per task with the 1.25 app via this system. This is a very decent improvement in performance. Thanks for the work put into optimizing the BRP4 applications.
Thanks for the feedback.
)
Thanks for the feedback. That's actually a bit more of a speedup than I had expected based on some tests on slower hardware. Definitely in relative terms, the speedup is more pronounced on faster cards.
I will now install the Linux CUDA app on Albert, stand by for more tests. I'm eager to see those GTX 680 .... ;-)
Cheers
HB
RE: How bout a single work
)
Got hold of 2 BRP4-tasks and ran them one at a time on my over clocked GTX660Ti (Core-[at]-1201.9 MHz).
Run time: 1183.69 and 1175.39 seconds for an average of 1179.54 s.
GPU load ~84%.
This on Win7 x64, PCI-E 3.0x16.
Here are some preliminary
)
Here are some preliminary numbers for the GTX 680.
One task per GPU
System #1 - Single GPU
x16 3.0 - 721 seconds
System #2 - Multi GPU
x16 3.0 - 785 seconds
x8 3.0 - 901 seconds
Overall, the performance looks great so far. I want to do some more testing with multiple tasks running at once, different PCI-E configurations, and with the CPU dedicated for BRP4 GPU only. The above tests were done with ~50% CPU load from running other CPU tasks at the same time.
I have a possible
)
I have a possible finding--not even a little bit sure--but am posting in case others might spot such a thing.
I've got two different hosts with the same GPU, a GTX 460. Neither has had downclocking problems for some weeks, but I found both downclocked severely today, with the problem persisting through system reboot.
It might just barely be possible that running the current Albert BRP1.28 CUDA ap, or on an even less likely note the Albert 0.29 Gamma Ray Pulsar application--or switching back and forth from those to the current Einstein applications was involved.
More likely something else in my system's history was the problem, but I thought I'd post the suspicion in case someone else sees something.
I'm not even sure what the true downclock frequency was, as different sources reported different numbers, but it was either 405 MHz or less--the reduction in power consumption and GPU temperature, while reporting exceedingly high GPU utilization, but making very slow progress on the WU was persuasive that downclocking was at hand.
I have sometimes the same
)
I have sometimes the same issue.It only happens when i overclock the gpu +130Mhz from baseline.After several hours the screen goes blank and with gpu-z it shows core speed of 400Mhz then i proceed with reboot and everything goes to normal.
I think it some fail safe method that Nvidia uses.Also I noticed when boinc is not running the core speed drops to 50mhz and goes to 800 instantaneous if a demanding gpu work is needed