Comment by physicsguy
25 days ago
I don’t know that that is fair.
A number of years ago I worked on a POWER9 GPU cluster. This was quite painful - Python had started moving to use wheels and so most projects had started to build these automatically in CI pipelines but pretty much none of these even supported ARM let alone POWER9 architecture. So you were on your own for pretty much anything that wasn’t Numpy. The reason for this of course is just that there was little demand and as a result even fewer people willing to support it.
Not just little demand, also expensive and uncommon hardware. If the maintainers don't have the hardware to test on they can't guarantee support for that hardware. Not having hardware available often happens because there's little demand for it, but the difficulty of maintaining software for rare hardware further reduces the demand for that hardware.
At least it's been fine for four years of research software on a POWER9 cluster I support (with nodes like the Summit system's).