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

(github.com)

8 points | by mxfeinberg 12 hours ago ago

3 comments

  • 0-_-0 4 hours ago

    32 bits vs 4 bits it looks like

    [-]
    • mxfeinberg an hour 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.

  • Slovian 12 hours ago

    [dead]