I’m running YOLOV3 TensorRT stand-alone app in my Jetson TX2. Without optimizing the TX2 I got 50~60ms inference for image classification with 60~99% GPU usage.
The interesting thing is, after I executed
sudo nvpmodel -m 0
I got very fast inference time around 17~18ms, but the GPU usage was very low, down to around 5%. Why is that?
Note:
I checked the GPU usage using tegrastats and gpuGraphTX python program.
I can’t use jetson_clocks.sh due to the absence of fan
FYI, the model only changes the available range of clocks and voltages (the DVFS table). It has no bearing on the actual currently selected clock. If the clocks are such that they have idled back, and on-demand use can trigger a higher clock, then the higher clock won’t occur until time under load has passed. If GPU requires data to feed it, then I would expect idled back speed to also reduce GPU use until clocks ramp up and have something to feed the GPU. Try without jetson_clocks, but put the system under a load and keep it under load for some time. Watch the load over time.