Upcoming changes to precompiled GPU architectures in TensorFlow pip package.

[Goldie Gadde]

TL;DR: Changes to precompiled GPU Architectures in TensorFlow pip packages,  if you use GPUs please continue reading.

TensorFlow pip packages currently contain SASS for CUDA® architectures 3.5, 3.7, 5.2, 6.0, 6.1, 7.0 and PTX for 7.0. 

In the next few days TensorFlow nightly builds will be changed to contain SASS for CUDA® architectures "3.5, 5.0, 6.0, 7.0, 7.5, 8.0 and PTX for 8.0".

Here are some user-visible implications of this change:

1. You may observe minor performance degradations for GPUs with compute capability 3.7 (e.g. K80), 5.2 (e.g. TitanX and M40) and 6.1 (e.g. P4).  If you need peak performance on these GPUs then please build TensorFlow from source or use TensorFlow Enterprise DL Containers published through the Google Cloud AI platform.

2. Currently TensorFlow fails with a “device kernel image is invalid” error on GPUs with compute capability 5.0 (e.g. some GeForce GTX GPUs).  After this change TensorFlow should run successfully on these GPUs.

These changes will be reflected in the tf-nightly pip packages this week and we will send a follow up email once the changes are submitted.

TensorFlow 2.4 will be the official release reflecting these changes.

Feel free to reach out to us on the TensorFlow developers mailing list if you have any questions

Thanks!