GPU support for Tensorflow 2.6.0 on Jetson nano

I got this error when I was converting a model using TensorRT.:
Registered Device: CPU
Registered Kernel: GPU

I tested it with tf.test.is_gpu_available() and it returned False, so I guess I got this error since TensorFlow failed to run on GPU. Then I found Tensorflow 2,6,0 requires a CUDA version of 11.2. While the latest CUDA version supported by Jetpack 4.6 is 10.2. Is there a way to build GPU support for 2.6.0 on Jetson nano? I saw many people using 2.6.0 so I wonder if I did anything wrong.

hey @zceelii
Can you once try with Previous tensorflow version Installing TensorFlow For Jetson Platform :: NVIDIA Deep Learning Frameworks Documentation please refer this docs and let us know whether you can leverage the power of gpu, as you are using tensorflow framework and trying to convert model using Tensorrt you can optimize the model using TF-TRT Installation Guide :: NVIDIA Deep Learning TensorRT Documentation follow this docs to install uff-converter to use tensorrt with tensorflow.

Thank you


It’s known that our TensorFlow v2.6.0 package has some issue on GPU mode.
Our internal team is actively working on this.

Currently, please use v2.5.0 instead.

$ sudo pip3 install --pre --extra-index-url tensorflow==2.5.0+nv21.8


1 Like

I tried to use Tensorflow 2.3.1 before but I got this error before getting into the GPU problem. I got libnvinfer 8 installed and 2.6.0 doesn’t has this bug. How can I solve this problem?
Many thanks