Why is the cuda runtime library not part of the driver? Now it seems really messy trying to pass programs written against the runtime api to other people since I would need to include copies of the runtime when giving the program or have them install cuda-toolkit, which just seems silly.
I’d think it’s not that much of a problem to maintain binary compatibility, or am I just missing something here?
Now writing against the driver api has the advantage that the program will just work as long as there is new enough driver present. Other alternative could be to provide statically the runtime library, but it is not present in the toolkit (or the sdk it seems)