How to overclock GPU
30 Nov 2011 8:42:12 UTC
Topic 196078
(moderation:
I just got a GTX580 and thought I'd return to crunch some E@H workunits. The cuda app seems to work ok but it's only running at 60% GPU utilization. I guess this is how the app is written but I thought I might speed it up by doing some overclocking.
There seem to be 4 parameters you can overclock on the GTX580: Core Voltage, Core Clock, Shader Clock and Memory Clock. I have fiddled around a little and gotten some speed increase but I'm curious about which parameter would benefit the E@H cuda app the most.
I'm not much interested in overclocking, say the memory, increasing the heat and noise on the card but then finding out that the E@H app don't benefit at all from the memory overclock :)
So, my question: Which overclock parameter(s) benefits the E@H app the most?
How to overclock GPU
)
I doubt that the E@H App would benefit from overclocking at all. The bottleneck of the computation is the part that's still done on the CPU. If you want to speed up E@H CUDA Apps, get a faster CPU. Besides expect the rate of invalid results to raise significantly with overclocking the GPU.
BM
BM
Thanks for the reply. Well
)
Thanks for the reply.
Well that's a little depressing. I realize the E@H calculations might be hard to run 100% on the GPU but when I crunched for Milkyway@Home the speed difference between CPU and GPU crunching was amazing. That's why I stopped crunching on CPU, felt like a complete waste of electricity.
I read something about running more than one GPU task simultaneously. Would that bring the GPU to 100% utilization? How is that done?
To run more GPU task you need
)
To run more GPU task you need to use a tweak with a file called app_info.xml..
See this thread:
http://einsteinathome.org/node/196075
RE: I read something about
)
For additional Information you also can read here http://einsteinathome.org/node/195643 (4 WUs on GTX480).
1st priority: run several WUs
)
1st priority: run several WUs simultaneously to make better use of your hardware as-is
Then:
- if you still want to overclock it's very probably the shader clock. Although this is tied to the chip clock on your card anyway. Leave the memory clock alone.
- don't increase voltage on a GF110 chip (yours). It's usually running hot enough already. You may want to try lowering the voltage to save power and reduce heat & noise. Careful, though, as lower voltages mean the chip reaches lower clocks reliably, i.e. you're working directly against overclocking.
MrS
Scanning for our furry friends since Jan 2002
If you run multiple work
)
If you run multiple work units at once on your 580, make sure to also have the card installed in a x16 slot. The extra PCI-E lanes do benefit performance with this application. As for overclocking, I run my 580s at 850 MHz with stock memory frequency. I am still able to undervolt my cards at this frequency without sacrificing stability. I run that in combination with my 920 @ 4.2 Ghz. The combination of the two has made it possible to complete three work units at once in around 3200-3300 seconds. GPU utilization is on average 88%.
Thanks for all the help guys.
)
Thanks for all the help guys.
RE: I doubt that the E@H
)
Another possebillity is running 2 per GPU, but I would not do this
unless it's a FERMI (400/500series) and are familiar with app_info.xml
file editting. One typo...............................