Considering CUDA version bump

[Stanley Seibert]

Hi Numba Users,

Numba currently works with both CUDA 7.5 and CUDA 8.0, but we could simplify the internals of our CUDA target if we raise our minimum requirement to CUDA 8.0.  Those of you using conda will be able to easily install the required CUDA 8 toolkit libraries, but it would still require that the NVIDIA driver on your system (which conda cannot install or upgrade) be version 367 or later.

We're curious if there's anyone out there who needs to stay on CUDA 7.5, especially those of your working on clusters where you don't have administrative privileges.   We don't want to get ahead of people's CUDA upgrade cycle.

Thanks!

[rl...@lbl.gov]

Sorry for the late reply! It seems that on the Titan cluster for instance (which our team is using), CUDA 8.0 is not yet installed.

However, I definitely understand the need or the numba team to drop support for multiple CUDA versions. That is fine with us: we can always stick to an earlier version of numba, which still supports CUDA 7.5.

[Stanley Seibert]

OK, that's good to know.  I suspected that large institutional clusters would not have upgraded to CUDA 8 just yet.  I was leaning toward punting this change to at least the January release (0.31), which will hopefully give places more time to upgrade.

We haven't seen any evidence that moving our code gen up to CUDA 8 would result in any noticeable performance improvements.

--
You received this message because you are subscribed to the Google Groups "Numba Public Discussion - Public" group.
To unsubscribe from this group and stop receiving emails from it, send an email to numba-users+unsubscribe@continuum.io.
To post to this group, send email to numba...@continuum.io.
To view this discussion on the web visit https://groups.google.com/a/continuum.io/d/msgid/numba-users/82327a66-cad7-45b2-bc72-970207710233%40continuum.io.
For more options, visit https://groups.google.com/a/continuum.io/d/optout.