pytorch - 💡(How to fix) Fix Question: Why can't PyTorch be compiled for older GPUs (sm_30) if the hardware is capable? [2 comments, 2 participants]

I'm a non-programmer trying to understand something.

I have a Quadro K5000 (Compute Capability 3.0) from 2012. It runs games, design software, 4K video — everything works perfectly.

But when I try to use PyTorch for AI training, I get: "CUDA error: no kernel image is available for execution on the device"

I understand the card only supports sm_30, and PyTorch binaries are compiled for sm_35+. But my question is:

Why can't I compile PyTorch myself to support sm_30?

I've read that the issue is in NumericLimits.cuh and a few other files. Some values like 0x7C00 and 0x7F80 don't compile for older architectures. But these seem like software limits, not hardware limits.

Is there an actual physical limitation that prevents a 4GB GPU from doing AI training? Or is this purely a software/compilation decision?

I'm not accusing anyone. I just want to understand if my hardware is truly incapable, or if the software stack has simply moved on and left older but functional hardware behind.

Thank you for any explanation.

See also my detailed question: https://gist.github.com/lynex001/7ee3a87b16ac2f7bc404b4581d91d70e