I need to use CUDA11.2 with tensorflow-gpu but it cannot match the driver version. What should I do?my OS:windows11 GPU:RTX3050
I have a similar problem and asked NVIDIA for support.
I have the newest driver for NVIDIA Studio installed. I would like to run tensorflow, which would require Cuda < 12.5 (preferably version 11.8).
So here :https://docs.nvidia.com/deploy/cuda-compatibility/forward-compatibility.html
I found that for the required cuda version I would need a nvidia driver version which is lower, correct ? But I can find it nowhere in the internet.
So at the moment nvidia-smi displays cuda version 12.9 and nvcc --version gives me cuda version 11.8. Tensorflow displays an error that it is not compatible with cuda 12.9
I got the following answer, which I interpret as ‘it is not possible to install the needed driver for this GPU’ , please find the answer attached below. I can’t really believe that I cannot use my GPU for tensorflow training… so if anyone has a solution please let me know. I am also willing to switch to Linux OS, but as far as I understood also Linux has only drivers which are too new, right ?
Hello,
Thank you for contacting NVIDIA Customer Care.
From the query description, I understand you are experiencing issues running TensorFlow due to a CUDA version mismatch.
Firstly, I apologize for the inconvenience caused due to this issue. Please be assured that I will do my best to help you further or point you in the right direction.
Based on the query described, I understand that you have the latest NVIDIA Studio driver installed, which supports CUDA 12.9, while you are trying to run TensorFlow with CUDA 11.8. Although you’ve installed CUDA Toolkit 11.8 (as confirmed by nvcc --version), the driver itself reports CUDA 12.9 via nvidia-smi, leading to a compatibility error within TensorFlow which is an expected behavior.
Since your system is equipped with the latest RTX 5070 Laptop GPU, please note that this RTX 50 Series laptop GPU is only supported from Studio Driver version 577.00 and above.
Older Studio drivers such as version 522.30 are only supported until the RTX 40 Series GPU’s and these align with earlier CUDA versions (such as 11.8) which are not compatible with the RTX 5070 Laptop GPUs, and attempting to roll back may result in hardware incompatibility or limited functionality.
If the issue persists even after the CUDA Toolkit installation, I recommend posting your query on the NVIDIA Developer Forums, where our engineers and the wider developer community can assist with advanced compatibility scenarios specific to TensorFlow.
Here are some information about my GPUs:
GPU 0: NVIDIA GeForce RTX 5070 Laptop GPU
Driver version: 32.0.15.7700
Driver date: 14/07/2025
DirectX version: 12 (FL 12.1)
Physical location: PCI bus 196, device 0, function 0
GPU 1: AMD Radeon™ 880M Graphics
Driver version: 32.0.13046.16
Driver date: 01/04/2025
DirectX version: 12 (FL 12.1)
Physical location: PCI bus 197, device 0, function 0