Notes from Speed v Accuracy BoF
55 Attendees

(notes do not duplicate slide content)

Ben Woodard question on exp() speedup. Ran some tests on old and new math libraries
	* 10,000 speedup affects HPC users because nodes fall out of lockstep, reducing overall speedup
	* found accuracy was better (2.5x) not worse
	* climate modelers seem to be really sensitive to last bit variation. Affects stability of algorithms

Comment: Many academics have accepted implicit inaccuracy in vendor libraries and adjusted their algorithms to accommodate that

Comment that quick rounding would help those where algorithms do not converge.

Could this be a tunable library

Ideas from Ben Woodard
	1. Vector units are much deeper now - could use those
	1. Compute one above/below to compute derivative to give rounding direction

Patick McGehearty notes that Intel compilers take advantage of 80 bit accuracy, which Arm does not have