How to upgrade from cuDNN 7.3.1 to 7.4 on Jetson Nano

Autonomous Machines Jetson & Embedded Systems Jetson Nano
1

Hi,

I currently have JetPack 4.2 installed on my Jetson Nano which has cuDNN 7.3.1.28-1+cuda10.0, however, on TensorFlow’s GPU build considerations page it states that TensorFlow 1.14 can only work with cuDNN 7.4 and cuda 10.0. How can I upgrade to cuDNN 7.4 on the Jetson Nano?

According to TensorFlow none of its versions will work with 7.3.

Thanks for the help!

2

Hi,

Please noticed that the official TensorFlow is for an x86 environment.
For Jetson, you will need to build it from source or use our prebuilt package here:

We do have a TF package for JetPack4.2:
https://developer.download.nvidia.com/compute/redist/jp/v42/tensorflow-gpu/

Due to GPU dependencies, it’s required to re-flash and install all the package from the same JetPack for compatibility.
This indicates that you need to upgrade the cuDNN package with newer JetPack version directly.

Thanks.

3

Hi,

Yes I built it from the source, however, running an object detection script using TensorFlow and a pi camera I run into an error stating that the version of TensorFlow requires cuDNN version 7.4 for TensorFlow 1.14, while the version I have on the Jetson Nano came with 7.3.1 on JetPack4.2. I checked the later versions of JetPack releases and between 4.2 and 4.2.1 cuDNN goes from 7.3.1 to 7.5.1 - is there anything with 7.4 available?

Thanks for your quick response!!

4

Hi,

The ARM package need to be compiled on the Jetson device under the same JetPack version directly.
If you build it on an x86 platform, it cannot be used on the ARM system.

We don’t have cuDNN 7.4 package. But you can install our prebuilt instead.

Thanks.