Trying to figure out what the last version of numba is that still supports CUDA 10.x and Compute Capability 3.0. The changes for 0.55.1 read:
Support for CUDA toolkits < 10.2 is deprecated and will be removed in Numba 0.56.
Is it < or <= ? There are CUDA SDK 10.0, 10.1 and 10.2.
I am attempting to prepare a demo on an old laptop with an early Kepler Chip inside (Quadro K1000M).
It’s <. 0.56 requires CUDA 10.2: Overview — Numba 0.56.2+0.gd6731f6d2.dirty-py3.7-linux-x86_64.egg documentation
0.57 will require CUDA 11.0.
You can still use the Kepler chip with newer toolkits - the last driver to support it was r470, so you should be able to use up to CUDA toolkit 11.4 on that laptop for the demo.
Thanks a lot. Then the linked wikipedia article is not right (?). Kepler chips can have compute capabilities 3.0 to 3.7. CUDA SDK 11.0 to 11.4 support down to compute capability 3.5. My chip has compute capability 3.0. You’re sure this works?
Oops, I’m probably wrong on some things - I had not realised the K1000M was 3.0 and not 3.5 - let me re-check everything I wrote above.
You’re right, it will have to be CUDA toolkit 10.2, and may also need Numba 0.55.2 or maybe an older version - I don’t think Numba has been tested on anything with Compute Capability 3.0 for quite some time, so I don’t know what will work on that laptop for sure.
numba 0.53 used to work on CUDA SDK 10.1 on a similar machine just fine, i.e. compute capability 3.0. Just awfully slow, but that’s down to the limits of this GPU and its cooling, I presume. Thanks for confirming my suspicion. CUDA SDK 10.2 will land me in 2018 software land … considering moving this onto some server instead.