← Back to context

Comment by lifthrasiir

21 hours ago

There are too many fp8/fp4 formats out there, many with different trade-offs. At least fp16 is standardized...

2 comments

lifthrasiir

Reply
elvircrn  21 hours ago

Yes, all of them would be great.

  • lifthrasiir  8 hours ago

    Hmm, you successfully nerd-sniped me. AFAIK there are at least:

    - 8-bit floating point types: (S1)E4M3 (finite only, has no-negative-zero variant), (S1)E5M2 (has no-negative-zero variant), E8M0

    - 6-bit floating point types: (S1)E2M3, (S1)E3M2

    - 4-bit floating point types: (S1)E2M1, NF4

    - Block floating point types: MXFP8 (E5M2/E4M3*32+E8M0), MXFP6 (E3M2/E2M3*32+E8M0), MXFP4 (E2M1*32+E8M0), NXFP [1]

    [1] https://arxiv.org/pdf/2412.19821

    Do you have anything to add on?