GPU SM Frequency

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi everyone,
I use RTX2080Ti and A10 to run my program where includes a kernel function. I have found something interesting. The GPU clock speed given from Nvidia is shown as follows: RTX 2080Ti 1.35GHz,and A10 0.885GHz. The actual SM frequency values are given from Profiling tool Nsight Compute.
When I ran different kernel functions on the same GPU, I have found that the SM Frequency are also different. I take RTX 2080 Ti as an example, the frequency is ±0.03GHz. Does different code design and the stall on the warps have effect on the SM frequency?
Another one is that, when I compute same kernel but on different GPUs, it looks like the GPU only execute the program at the frequency given by Nvidia. It won’t have a much higher frequency than the given frequency. It just floats in the range I have mentioned in the last problem below.
And the last question is that, I have also found a frequency change. Because the GPU I am using is on a server. There might be other process at the same time on the server. But when there is another process on the GPU I currently use and the power, memory usage do not exceed the limit. I actually find a rise in SM frequency and my profiling result for my kernel function can have a 2-3µs difference. The GPU will have better performance when it is more loaded/ occupied by processes?

Thanks for any help!

2

nsight compute has default clock-control behavior. You might want to investigate that. You might want to give more information about how exactly you are profiling with nsight compute. And if you need to drill into the nsight compute clock control system, you might wish to ask such questions on the nsight compute forum.

1 Like
3

Thanks a lot. I have chosen the clock control to Base in Nsight Compute. I would compare the option Base and None.