Profiler time do not match NVIDIA Nsight Computer

I profiled the fc layer with torch.profiler, got kernel time 1.9 ms ,

with profile(activities=[ProfilerActivity.CUDA, ProfilerActivity.CPU], record_shapes = True, profile_memory = True, with_modules=True, )

but I got kernel time 2.2ms from nsight computer.
there is 10% miss, where is the error?
It is important for me