Running the latest Jetpack on my Jetson Xavier AGX development system, in MAXN power mode. Timing a kernel using both Cuda events and std::chrono, which are consistent with each other. When I run my app directly on the Jetson using an Ubuntu bash window I get timing results that are reasonable and also consistent with what the profiler reports. When I run the exact same thing using a MobaXterm connection, the times are always longer by about 10x. That’s the really odd part. I also notice when I run the profiler that the GPU clock rate, and therefore the reported duration, can be quite variable.
Why would the exact same run from MobaXterm be 10x slower?
Is there a way to force the clock to be the max 1.377 GHz for the run?