Nsight Compute: The frequency is not fixed

[kawano.junpei]

I am using Nsight Compute for performance measurement, and I have a question about how to read the item.
It uses the GPU of A100 and used Nsight Compute with the following settings.

[GPU Clock (A100)]
Base: 765 MHz
Boost: 1410 MHz

[Clock Control (Nsight Compute)]
None

If you check the information for a specific kernel in this state, the value of “SM Frquency” is about 1200 - 1350.
(As an example, we use “histogram” for the cuda-sample. GitHub - NVIDIA/cuda-samples: Samples for CUDA Developers which demonstrates features in CUDA Toolkit )

If “Clock Control” is “None”, the frequency is fixed at 1410 Mhz. Why is this value not constant?

In the first place, I think “SM Frequency” refers to the frequency, but what does “SM Frequency” refer to?
Regarding the question, if I set “Clock Control” to “Base”, will this value be 765 Mhz?

[rs277]

@kawano.junpei

If “Clock Control” is “None”, the frequency is fixed at 1410 Mhz. Why is this value not constant?

“None” means there is no clock control, therefore the clock starts out at the boost frequency and during kernel execution, the frequency will vary, depending on workload, temperature etc.

“Base” means the clock is locked at the base frequency and will not vary, enabling repeatability between profile runs.

[skps23]

Another similar discussion here:

Sum of kernel time is different in ncu and nsys

For the best single kernel duration (full GPU) I would use NCU with --cache-control none --clock-control none and metrics collection limited to only gpu__time_duration.sum. nvidia-smiis only required if you want to manually manipulate the GPU clock frequencies.

Would you recommend this?

[rs277]

I think “single” is the operative word here.

For measurement over “several thousand of times in a loop”, as outlined in this thread, instrumentation and average/median of the result is probably more realistic.