Copy engine vs SM mode

[MARD1NO]

I find in SM Mode(for example, write a copy kernel), memcpy's bandwidth is higher than copy engine(cudaMemcpy) in H20. 3.3T/s vs 2T/s, is this reasonable?

[jnbntz]

The CUDA Samples are intended for illustration and functional use, but are NOT intended for serious performance testing.

We'd recommend you use the NVBandwidth code to do performance testing.

[MARD1NO]

The CUDA Samples are intended for illustration and functional use, but are NOT intended for serious performance testing.

We'd recommend you use the NVBandwidth code to do performance testing.

I just observe it on my nsight system profile, and found cuda D2D elapsed time is much higher than my handwritter kernels, is this reasonable? I only know cudaMemcpy do not use any SM resource