TurboQuant can reduce vector index size by 10x at 100M Row Scale

8 pointsposted 11 hours ago
by mxfeinberg
(github.com)

3 Comments

0-_-0

3 hours ago

32 bits vs 4 bits it looks like

mxfeinberg

29 minutes ago

Yup, and unlike the original turboquant paper, my implementation is pinned to using a 4 bit code book so I could use SIMD kernels for performance.