An 83-Format Numeric Catalog with Bit-Exact Conformance Vectors: A Vendor-Neutral Reference for FP8, BF16, MXFP4, and Microscaling Formats

The paper presents a comprehensive catalog of 83 numeric formats relevant to machine learning hardware, including FP8 (E4M3, E5M2), BF16, MXFP4, and microscaling formats. It includes six bit-exact conformance packs that provide a standardized reference for these formats, ensuring consistency across different accelerators and mitigating silent divergences in model portability. This resource is essential for practitioners to ensure accurate cross-platform model deployment and to facilitate better understanding and debugging of format-related discrepancies.