“For older cards there seems to be a bit of a mix, some are better and
others not. We may change those to use arrays too, but more testing is needed,
only Titan and Tesla K20 (sm_35) is changed for now.”
I guess that “some are better” means the high end keplers (gtx)?
Wouldn’t it be faster to use arrays on sm30 680gtx also ? There is an if macro “<350” in the kernel files for the titan.
But 680 is sm30.
Has anyone made some tests ?