How reliable is the duration reported by nvprof?

I am trying to profile a run of AlexNet for caffe and I saw some strange results for the duration depending on how many metrics I collected. For one case I only collected the Floating point operations (Single Precision) and for the other case I collected all supported metrics. The results were vastly different for the two cases: 17 milli-seconds versus 69 micro-seconds.
My instinct was that the 68 micro-seconds number was more accurate as it was probably computed as an average of various runs for the second case. But then I ran the first convolution layer separately, (using the cudnn API as does caffe-compiled with cudnn support), collecting all supported metrics, and this time the duration reported was 58 micro-seconds! The question is: Is the duration reported by the profiler reliable at all?

My experience is that it is not reliable at all.