PyTorch no longer supports this GPU because it is too old.

41

Thought I’d ask this here rather than starting a whole new thread. We’re shopping for a new GPU to use with PyTorch and I’d be very grateful for some recommendations regarding what models will work best. We don’t have unlimited budget, of course… :grin:

Any guidance would be much appreciated. (As you can imagine, I didn’t get very far searching the forums for “GPU” and “model”!)

1 Like
42

This blog post might be a good starting point:

43

Is there any hope for my GTX 670 (2gb) and 3.0 Cuda Capability?

1 Like
44

Is it still possible to install Pytorch with CUDA 5.0/5.2 support (without installing from source)?

45

I think CUDA5 was released in 2012 or 2013, so I doubt you could build PyTorch using these old versions at all.
What kind of GPU are you using that you would like to stick to CUDA5?

46

Nvidia Quadro M2000. Based on this thread Cuda 5 was only brought back for Pytorch 0.4, but that would be better than nothing.

47

Sorry, I might have misunderstood the question, but do you mean CUDA5 or a GPU with compute capability 5?
The latter should be supported in the current binaries, the former was never supported as far as I know.

48

My apologies, I think I was the one that was confused. I meant GPU with compute capability 5. I was able to get things working. Thank you

49

What about even older cards? There are many of us who cannot afford new ones. :frowning_face:
…dist-packages/torch/cuda/init.py:81: UserWarning:
Found GPU0 GeForce GTX 760 which is of cuda capability 3.0.
PyTorch no longer supports this GPU because it is too old.
The minimum cuda capability that we support is 3.5.

warnings.warn(old_gpu_warn % (d, name, major, capability[1]))
_CudaDeviceProperties(name=‘GeForce GTX 760’, major=3, minor=0, total_memory=1997MB, multi_processor_count=6)

50

I don’t see any support for CUDA3 cards. :frowning:

51

I suggest low-end cards should be supported for a certain period, it is better than nothing,especially for a beginner

52

Older compute capabilities are still supported in the current binary releases starting from sm_37, which means that the binaries support architectures from sm_37 (Kepler), through Maxwell, Pascal, Volta, Turing, to Ampere (sm_86). Your 650M was released in 2012 and its support was dropped a while ago while we moved to CUDA operators, which are not supported on these old compute capabilities anymore.