Skip to content

UE5M3 format #29

Description

@AntonOresten

Described in https://docs.nvidia.com/cuda/developer-preview/13.4/parallel-thread-execution/index.html#alternate-floating-point-data-formats. Seemingly a block-scale type for Rubin, leveraging the 8th bit that goes unused in the UE4M3 scale type of NVFP4.

The ue5m3 encoding does not support infinity. NaN value is limited to 0xff.

Trivially added with @microfloat Float8_E5M3FNU sign=0 exponent=5 significand=3 nonfinite=NanOnlyAllOnes, but also requires registration in the CUDA extension for it to be recognized by e.g. cuBLASLt.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions