Tensor Core not following power limitation

Hi All,

I am using the Nvidia tensor-core SGEMM sample (CUDASamples/cudaTensorCoreGemm.cu at master · BigNerd95/CUDASamples · GitHub) on Titan V100 and trying to monitor the effect on performance when we limit the total power to GPU. To limit the total power available to GPU, I am using this command ‘nvidia-smi -i 0 -pl 150’ . But I am observing that when running kernels on the Tensor core, the GPU does not follow the power limitation and runs at 250W. Why is this happening? Does GPU automatically switch to maximum power consumption when running WMMA instructions, and is it possible to limit power to GPU when using tensor core?