4 ms·
An odd thing to me about the GPU space is how rapidly many ML libraries that rely on CUDA deprecate earlier versions. Two years ago I bought a used workstation
by anigbrowl 3y ago
An odd thing to me about the GPU space is how rapidly many ML libraries that rely on CUDA deprecate earlier versions. Two years ago I bought a used workstation laptop with (iirc) a Kepler GPU, and I was surprised to find that virtually none of the popular Python libraries would run on it. I spent a whole day stepping back through earlier library versions which of course created other kinds of dependency problems before giving up in favor of an external GPU.
I wasn't trying to do anything difficult, I was literally at the 'hello GPU world' stage and wanting to try out some very basic exercises from books. Can't help feeling that Nvidia is doing a lot of the groundwork with CUDA and some library maintainers are just surfing on it.
- dist-epoch 3y agoThat's because newer GPUs support newer instructions sets. Like AVX, AVX2, AVX512, but basically a new set every year instead of every 10 years like in CPU land. And it's not just about speed, they provide new functionality. So at some point supporting older GPUs significantly slows down those who have the newer GPUs. The next architecture after Kepler, Pascal, added huge new capabilities, that's why nobody supports Kepler anymore, it's too limited. > wanting to try out some very basic exercises Should be possible to do that with an older Python (3.5?) and associated libraries. Versions around 2016 should work.
- rahimnathwani 3y agoI seem to recall ~6 years ago I had to recompile an ML library because my CPU didn’t support AVX2. So even though the release binaries required AVX2, the source code was written in a way that it could work without it. I wonder whether this required specific effort by the developers of that library, or whether it was taken care of by some underlying tooling? Can you write CUDA code today that uses the latest features but can be compiled for older cards as well?
- google234123 3y agoDeprecating support for a 10-12 year old gpu doesn’t seem rapid to me