Get different metric results using nsight compute and nvprof

I’m profiling a kernel using nvprof and ncu.

The gld_efficiency metric using nvprof shows this:

But the corresponding metric in nsight comput show this:

I see in the manual that they are the same metric, both shows if there are any waste in bandwidth. But why they differ so much?

which is right ?? thank you !

Thanks for reporting this. We’re aware of this issue and are working to determine the correct way to map this. I will try and provide some more information after our investigation.

thank you. If I want to see the gld_efficiency metric, which is the right way by now? Using nvprof is the right result ?
thank you!

It is true that these 2 metrics should match, but in your case the don’t. We’re investigating that bug. To evaluate in Nsight Compute whether the value is correct, look at the Excessive Sector Accesses on the Source Page. Source lines with high values would lead to a lower efficiency.