I’m running TensorFlow inside the NVIDIA NGC TensorFlow 25.02 container.
Setup:
- GPU: NVIDIA GeForce RTX 5080 (compute capability 12.0, Blackwell)
- TensorFlow: 2.17.0
- CUDA (inside container): 12.8.0.38
- Host driver: 576.88 (reports CUDA API 12.9)
- OS: Ubuntu 22.04
When I run my TensorFlow code, I get the following message many times:
'+ptx85' is not a recognized feature for this target (ignoring feature)
From what I understand, this happens because the TensorFlow build in 25.02 wasn’t compiled with PTX for compute_12.0, so it falls back to older architecture kernels. The code runs fine, but it’s not fully optimized for the Blackwell GPU architecture.
Has anyone managed to get this container running with proper compute 12.0 PTX support? And does anyone know when official TensorFlow builds will add full Blackwell support?