Cudnn 8.2.x for cuda 10.2 on Linux PPC not available

We see that cudnn 8.2.x ( in fact versions after 8.0.5) for cuda 10.2 haven’t been created for Linux PPC. cudnn 8.2.x for cuda 11.x is available for Linux Power architectures but not the ones for cuda 10.2.
Is there any plan to make that combination available in near future or NVIDIA is going to stop supporting newer cudnn for cuda 10.2 on Linux PPC?
Thanks in advance.

Our Linux PPC support is currently limited to the latest toolkit major version, with portability across minor versions (e.g. that means CUDA Toolkit 11.x for cuDNN 8.2.y). We don’t have any plans to change that, but we’d love to hear your feedback if that’s an issue for you.

1 Like

Thank you for your reply.

We work for an opensource project called Open-CE that contains everything needed to build conda packages of various machine learning and deep learning frameworks like Tensorflow, Pytorch, Xgboost, LightGBM, etc with gpu support on Linux x86_64 and Linux ppc64le architectures. Open-CE is used within various IBM products as a common platform of Data Science libraries for both x86 and Power.
And we have been supporting cudatoolkit 10.2 as well as 11.2 as of now. We would like to upgrade CUDA to 11.4, which means updated cuDNN too. But newer cuDNN isn’t available for CUDA 10.2 on PPC because of which we have to use cuDNN 8.0.5 for CUDA 10.2. It has worked in our last release of Open-CE 1.4.0 (which included TF 2.6, pytorch 1.9, etc.) but not sure if it would work in newer versions of TF and pytorch. We will try TF 2.7 (r2.7 branch) to see if cuDNN 8.0.5 still works with CUDA 10.2, and get back to you in case it doesn’t.

Thank you again!