jetson_clocks vs nvpmodel


I have a program that runs multiple kernels, and it’s also quite heavy on the cpu0 (reading a lot of data from usb).

I have been using it for a while with nvpmodel -m 3 succesfully. Recently I did some heavy optimizations, and the kernels execute in less time, but now it doesn’t work with model 3 or any other model. It only works if I run jetson_clocks (otherwise the load in cpu0 goes to 100% and everything pretty much freezes, I can’t even run the profiler remotely). Running gtop it seems that cpu0 is not going to the max frequency, but it’s hard to say because the system is not responsive.

I wonder if somebody has an explanation for that. Thanks.

Hi cpg_To, the jetson_clocks script disables the DVFS governor (Dynamic Voltage & Frequency Scaling) and locks the CPU/GPU/memory clocks to their maximum values defined by the active nvpmodel. jetson_clocks will only lock the clocks up to the limits set by the current nvpmodel (so if your active nvpmodel is profile #3, it will not go above the limits defined by profile #3).

To fully set the board to it’s max limits, run nvpmodel -m 0 (MAX-N) followed by jetson_clocks. The processor will then be locked as high as it can go. Hope that helps clear it up.