As vfar as I understand this, yes, overclocking the GPUs will increase performance also for CUDA apps.
Since the E@H ABP2 CUDA App is currently doing only one part of the calculation (FFT, the most demanding operation in the CPU version) on the GPU and the rest still on the CPU, you will not see that much of an improvement. As always, overclocking has some risks: if it's overdone, you will get false results and possibly damage your hardware.
Most cards come from the factory already over-clocked ... with summer coming on, unless the machine is in a cool place you may want to think hard about OC ... as to the gains, because EaH does only about 30% of the work on the GPU, OC here is not going to buy all that much for you...
On the projects where the GPU is used more then OC buys you more ...
My 2 cents has always said that the instability and crashed tasks are not worth the minor increase in speed you are likely to obtain ...
Now the OC fans are going to jump in and tell you that I am all wet and that OC is the most wonderful thing in the world and cures zits as well ...
remember, it isn't their system they are risking ... tis yours ... YMMV ... :)
Hi,
I had my GPUs overclocked but see difference in runtime. As far as I can tell the gpu app is bandwidth limited. I see shorter runtimes, when I increase the PCIe clock. Going from 100->110MHz runtime goes down like 2-3%. Going from PCIe 16x to PCIe4x increase the runtime by 20%.
This is on a P35 Intel chipset with 2 9800GX2.
If the statement is still correct on a *modern* PCIe 2.0 system with no front side bus limitation, I can´t tell.
On my system I need more PCIe bandwidth.
Yeah, betrter PCI throughput will help almost every CUDA app. But I think what was meant by overclocking by the OP was the GPU clock , not the PCI frequency. Because CPU performance and, as you wrote, PCI throughput are dominant factors, the GPU clock rate should not have a dramatic impact on the current ABP2 CUDA app, at least that's my guesstimation.
What I was saying: As long as one is bandwidth limited, overclocking the core/shaders makes no sense.
For me it is the case. Overclocking by 10% resulted in no change in runtime. But overclocking the PCIe clock does something.
![[AF>EDLS] Polynesia](https://einsteinathome.org/sites/default/files/449945_avatar_2.jpg?1470080812 "[AF>EDLS] Polynesia"){kind=avatar}
increased frequency processor Nvidia GPU
)
As vfar as I understand this, yes, overclocking the GPUs will increase performance also for CUDA apps.
Since the E@H ABP2 CUDA App is currently doing only one part of the calculation (FFT, the most demanding operation in the CPU version) on the GPU and the rest still on the CPU, you will not see that much of an improvement. As always, overclocking has some risks: if it's overdone, you will get false results and possibly damage your hardware.
Personally I would not recommend it.
CU
HB
Most cards come from the
)
Most cards come from the factory already over-clocked ... with summer coming on, unless the machine is in a cool place you may want to think hard about OC ... as to the gains, because EaH does only about 30% of the work on the GPU, OC here is not going to buy all that much for you...
On the projects where the GPU is used more then OC buys you more ...
My 2 cents has always said that the instability and crashed tasks are not worth the minor increase in speed you are likely to obtain ...
Now the OC fans are going to jump in and tell you that I am all wet and that OC is the most wonderful thing in the world and cures zits as well ...
remember, it isn't their system they are risking ... tis yours ... YMMV ... :)
Hi, I had my GPUs overclocked
)
Hi,
I had my GPUs overclocked but see difference in runtime. As far as I can tell the gpu app is bandwidth limited. I see shorter runtimes, when I increase the PCIe clock. Going from 100->110MHz runtime goes down like 2-3%. Going from PCIe 16x to PCIe4x increase the runtime by 20%.
This is on a P35 Intel chipset with 2 9800GX2.
If the statement is still correct on a *modern* PCIe 2.0 system with no front side bus limitation, I can´t tell.
On my system I need more PCIe bandwidth.
Interesting. Yeah, betrter
)
Interesting.
Yeah, betrter PCI throughput will help almost every CUDA app. But I think what was meant by overclocking by the OP was the GPU clock , not the PCI frequency. Because CPU performance and, as you wrote, PCI throughput are dominant factors, the GPU clock rate should not have a dramatic impact on the current ABP2 CUDA app, at least that's my guesstimation.
HB
What I was saying: As long as
)
What I was saying: As long as one is bandwidth limited, overclocking the core/shaders makes no sense.
For me it is the case. Overclocking by 10% resulted in no change in runtime. But overclocking the PCIe clock does something.
Ah I see. You might have
)
Ah I see. You might have missed a "no" in your message, tho:
CU
HB
You´re right. I have even
)
You´re right.
I have even considered downclocking the gpu but haven´t done it so far.