TensorFlow + RTX 5090 + WSL: CUDA 12 Installed in WSL but Windows Driver Uses CUDA 13 (original) (raw)

December 5, 2025, 10:48am 1

I’m trying to run TensorFlow with GPU on Windows 11 + WSL2 using an NVIDIA RTX 5090.

The issue

TensorFlow currently supports:

CUDA 12.3

cuDNN 9.x

Requires NVIDIA driver ≥ 560.94

But RTX 50-series GPUs ship with CUDA 13-based drivers only — and no 560.xx driver exists for these GPUs.

What we did

Installed CUDA 12.2 and cuDNN inside WSL

Verified that WSL CUDA toolkit is version 12.2

BUT WSL still gets the CUDA runtime from the Windows driver, which is CUDA 13

TensorFlow refuses to initialize the GPU because the driver version is too new and unsupported

Problem summary

Even in WSL, CUDA depends on the host Windows driver, so we cannot run CUDA 12 apps as long as the lowest CUDA version available for this graphics card is 12.8.

Question

Is there any workaround to run TensorFlow on an RTX 5090 today? For example:

A TF build that works with CUDA 13?

A compatibility layer between CUDA 12 and CUDA 13?

Any announcement from NVIDIA/Google about driver/toolkit support for 50-series?

Thanks in advance!

Please provide a version that allows RTX 5090 to be compatible with TensorFlow on Windows without installing a WSL environment!

This is expected behavior rather than a misconfiguration.

WSL always uses the host Windows CUDA driver, and RTX 5090 ships with CUDA 13–based drivers only. TensorFlow binaries are tightly coupled to specific CUDA majors and will reject newer drivers even if backward compatibility exists at the driver level.

There is currently no supported compatibility layer between CUDA 12 and CUDA 13 for TensorFlow.

Practically, your options are to wait for official TF CUDA 13 support, try nightly builds, build from source, or run on native Linux where driver/toolkit combinations are more flexible.