Binary Vector Search at 350GB/S Using ARM Neon (opens in new tab)

(topk.io)

7 pointsEquiet8mo ago2 comments

2 comments

re: optimization for 1024b vectors — do you pad shorter ones, or fallback to a more general kernel?

We do a projection of the original vectors so that it matches one of our optimized kernel. This generally gives us better recall vs. simple padding since all bits are utilized.

j / k navigate · click thread line to collapse

Binary Vector Search at 350GB/S Using ARM Neon (opens in new tab)

(topk.io)

7 pointsEquiet8mo ago2 comments

2 comments

MarekDlugos8mo ago

re: optimization for 1024b vectors — do you pad shorter ones, or fallback to a more general kernel?

marekgalovic8mo ago

We do a projection of the original vectors so that it matches one of our optimized kernel. This generally gives us better recall vs. simple padding since all bits are utilized.

j / k navigate · click thread line to collapse