I am using NVML library (nvmlDeviceGetPowerUsage) to get the power consumption of Tesla K40c GPU.
The results ranges from 21 Watts to 69 Watts, according to different frequency settings. And whether running a kernel does not influence the results.
However, the maximum power consumption is 235 Watts in the specification.
Why there is such a large gap between the measured results and the specification?
It may be a combination of workload and sampling frequency.
It’s not clear how often you are sampling the power usage using NVML. Furthermore it’s not clear what workloads you are running.
I can detect a large power difference between idle and when I am running a large CUBLAS SGEMM operation on a K40c using nvidia-smi, which is also built on top of NVML. So I’m quite sure it’s possible to witness a meaningful difference.