Performance of Jetson TX2 when working with TensorFlow

Hello Community,

I’m new to DeepLearning and want to get into this topic, especially Reinforcement Learning.

I freshly bought a JetsonTX2 and flashed it with the newest Jetpack3.3.
I also installed TensorFlow according to this instruction:

I want to do a Reinforcement Learning project with openAI gym but face major problems when it comes to performance of the Jetson TX2. I have some example code for the implementation of a DDPG Agent for the Pendulum-v0 environment:

However the framerate drops as soon as the agent tries to train. Even without TensorFlow running, examples like the BipedalWalker-v2 Environment have an framerate around 15-20 FPS.
Checking the system monitor i can see that all 6 CPU cores are activated and working between 0 and 60 %. Also the GPU memory is around 33% (2.2 / 7.1 GB usage). According to other forums I switch the nvpmodel to mode 0 which should be the high performance one.

Output from “~/tegrastats”:

RAM 2521/7853MB (lfb 632x4MB) CPU [50%@806,0%@345,0%@345,49%@806,54%@806,50%@806] BCPU@43.5C MCPU@43.5C GPU@42C PLL@43.5C Tboard@41C Tdiode@41C PMIC@100C thermal@42.9C VDD_IN 2309/2386 VDD_CPU 462/442 VDD_GPU 230/230 VDD_SOC 461/461 VDD_WIFI 0/43 VDD_DDR 578/588

I never before worked on an NVIDIA Board before and wanted to know if this is a known problem or simply some user errors from my side.

Thx for your support


You can use this script to maximize the Jetson performance first:

sudo ./

More, we also have a tutorial for reinforcement learning with GYM.
It’s recommended to take it a look first: