2 pointsby AkaiNa7 hours ago2 comments
  • p1esk7 hours ago
    That’s just good old block floating point, right?

    https://en.wikipedia.org/wiki/Block_floating_point

    • AkaiNa6 hours ago
      Yes. But standard block floating point uses a linear grid scaled by a shared exponent. Whereas AXS-6 uses a NormalFloat grid scaled by a shared exponent to maximize information density for bell-curve distributed weights. Essentially a Block Scaled Normalfloat-5.
  • 7 hours ago
    undefined