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.

You could always check the Nvidia site for the latest version of CUDA and it's document, as well as the new features it provides. — kangshiyin, Jan 23 '13 at 02:33

score 4 · Accepted Answer · edited Jan 23 '13 at 04:13

4

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.

edited Jan 23 '13 at 04:13

Greg Smith

11,007
2
36
37

answered Jan 23 '13 at 00:45

Eugene

9,242
2
30
29

Could you tell me which version support that? I didn't find it from the manual. – Hailiang Zhang Jan 23 '13 at 02:14
Simply download the latest toolkit. I believe 4.1/4.2 should support this as well but I'm not sure. – Eugene Jan 23 '13 at 17:09

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

1 Answers1