is there anyway to avoid this serialization behavior in cuda profiling?

Question

According to CUDA streams not overlapping , "the profiler will serialize streaming to get accurate timing data". Now the question is, is there anyway to avoid this serialization behavior in cuda profiling (say nvvp)? I am using Fermin M2090 and cuda-4.0.

Answer 1

The Visual Profiler 5.0 (including nvprof and CUPTI) and Nsight Visual Studio Edition 2.0 and greater (>2 years old) support concurrent kernel trace for Fermi and Kepler devices.