Hello, i want to start working with PyTorch, but my graphic card is to old (compute capability 3.0).
So i decided to buy a newer one.
The problem is that i can’t find an info what compute capability i need at minimum?
I read a lot of threads where members said that 5.0 is also to old.
So maybe 6.1 is to old too or may be deprecated soon.
What graphic card i have to buy so that i can work with pytorch?
4.000$ for a TITAN V is a little bit to much for me.
Regards
Hi,
5.2 should keep being supported for a while as lots of people use Titan X cards.
The 10xx cards should be as well.
The reason for removing the 2.x support is that so many features are missing that GPU implementation would need to be so different.
Finally for 3.0+ cards, you can still compile from source and use them. It is just not in the binary versions because the number of people using them is low and it makes the binary huge (500MB+).
Hi @albanD,
I have similar question - where could I find information about CUDA version vs compute capability range?
In my understanding upper versions of CUDA drivers increase demand to compute capability. In my specific case I have card GeForce GTX 860M, which compute capability is 3.0/5.0. I installed the latest CUDA toolkit (11.2) and v460.67 driver. I want to verify that this version of CUDA toolkit fit my hardware as I also noticed pytorch and CUDA mismatch toolkit is also possible
You could check the CUDA - Wikipedia page, where the CUDA toolkits with the supported compute capabilities are listed or e.g the CUDA11 release notes, which also mention the deprecation:
Support for the following compute capabilities are deprecated in the CUDA Toolkit:
- sm_35 (Kepler)
- sm_37 (Kepler)
- sm_50 (Maxwell)
Thank you for extra information. My card is fit in Maxwell group. CUDA 11.0 and newer do not support it.
The CUDA11.0 release notes claim that sm_50 is deprecated, but not removed, so you should still be able to build CUDA applications for it with the mentioned CUDA toolkit.