Which CUDA Toolkit version for older NVIDIA Driver

17,937

Solution 1

I did not find an official NVIDIA document, but here is a list containing the minimum driver versions up to CUDA 6.5:

http://docs.roguewave.com/totalview/8.14.1/html/index.html#page/User_Guides/totalviewug-about-cuda.31.4.html

So for your driver version 304 this means CUDA 5 is supported.

However, there should be a newer driver (version 340.76) available which still supports your card according to NVIDIA (see "Supported Products"):

http://www.nvidia.com/Download/driverResults.aspx/81761/en-us

Using this driver you should be able to use CUDA 6.5.

Solution 2

Finally Nvidia decided to add a table to the release notes in which the driver versions are determined. Here is the link from Nvidia documentation for Cuda version 10 and here is the screenshot:

Also according to Cuda installation logs in Ubuntu and another stackoverflow post CUDA toolkits require the following minimum Nvidia driver versions in Linux OS:

at least 340.00 is required for CUDA 6.5 at least 331.00 is required for CUDA 6.0 at least 319.00 is required for CUDA 5.5

17,937

cdeterman

Data scientist interested in bioinformatics, machine learning, and multivariate analysis. Primarily R and Python programmer but familiar with many languages.

Updated on June 25, 2022

Comments

cdeterman about 2 years
I have been provided an older NVIDIA graphics card (GeForce 8400 GS) to begin exploring some GPU computing. I have tried to complete the installation successfully but have stumbled upon a problem. Here are my steps (on Ubuntu 14.04)

sudo apt-get install nvidia-current (this installs nvidia-304 in my case)

After rebooting, a quick query shows that my kernel is indeed using nvidia successfully
```
lspci -vnn | grep -i VGA -A 12

01:00.0 VGA compatible controller [0300]: NVIDIA Corporation GT218 [GeForce 8400 GS Rev. 3] [10de:10c3] (rev a2) (prog-if 00 [VGA controller])
...
Kernel driver in use: nvidia
```
Naturally, I thought I could then install cuda with:
```
sudo apt-get install cuda
```
But this tries to install nvidia-346 on my system causing my system to no longer display my desktop and the installation is incorrect. I have verified that the nvidia-346 is the problem by specifically installing it as opposed to nvidia-current. The Linux Getting Started Manual says I should just need to install CUDA with apt-get but I need an older driver for my graphics card.

How can I install CUDA to work correctly with my older nvidia driver so I can conduct some GPU computations? Is there a list someplace that lists the what CUDA toolkits go with each NVIDIA driver? I suspect I need an older toolkit, I just don't know which one.
cdeterman over 9 years

This was exactly what I needed. I installed the 340 driver and then the 6.5 toolkit and everything appears to be working normally. Thank you.
Mateusz Kacprzak over 8 years

I have the same problem, but 6.5 toolkit is forcing installation of driver which brakes X's. What should I do in that case? [Ubuntu 14.04, GeForce 8400M GS, working driver nvidia-340]
Greg K. about 8 years

In my case, i am working on a system with cuda driver 304 installed, but there are newer toolkits available in the system. In fact, I am able to compile with cuda-7.5 and execution looks ok. Can i trust the 7.5 toolkit with the older drivers or should i contact the system administrators?
crlb almost 8 years

So this [developer.nvidia.com/cuda-gpus](list) is not really relevant? on that list, GeForce 8400 GS is not even listed. So as long as you can install a certain driver version (e.g 340.21), you can use a specific version of CUDA (6.5)?
Andy almost 7 years

could you explain what logs exactly you would get those numbers from, and how? (OS?) Curious what the minimal required version for CUDA 9.0 is. 9.0 RC ships with 384.59, but I do not know the minimal required.
Saeed Mohtasham almost 7 years

They are part of installation logs using Ubuntu 16.04 as OS and using ".run" files for CUDA installation.
Joe over 5 years

Are these version numbers available programmatically somewhere in the CUDA API?
user27221 almost 4 years

not sure why anyone else would want to do this, but to use an older card (my 980 died and I had a spare gtx 580) with nvidia-docker in this way is possible by using an older version of nvidia-docker2 (eg. sudo apt-get install -y nvidia-docker2=2.2.2-1 --allow-downgrades see github.com/NVIDIA/nvidia-docker/issues/…) and an older cuda (say docker run --gpus all --rm -it nvidia/cuda:6.5 bash). working with ubuntu 18.04 with nvidia-driver version 340.108