An 84-Format Numeric Catalog with Bit-Exact Conformance Vectors: A Vendor-Neutral Reference for FP8, BF16, MXFP4, and Microscaling Formats

Researchers have published a vendor-neutral catalog of 84 numeric formats used in machine learning hardware, including FP8, BF16, and MXFP4, with bit-exact conformance test vectors to enable consistent model porting across different accelerators. This addresses a critical gap where silent numerical divergences occur when moving ML models between vendors without a shared reference standard.

The proliferation of specialized numeric formats in machine learning hardware has created a fragmentation problem that lacks standardized reference material. As different vendors—NVIDIA, Google, AMD, and others—implement proprietary or variant interpretations of floating-point formats optimized for inference and training efficiency, engineers face the challenge of detecting silent numerical divergences when porting models. Without a shared reference, these divergences remain hidden, complicating debugging and model validation across platforms.

This work addresses a critical infrastructure gap in the ML ecosystem. By providing 84 standardized numeric formats across 13 families with JSON-encoded conformance packs and SHA-256 fingerprints, the researchers create a vendor-neutral measurement tool. The inclusion of an IEEE P3109 v3.2.0 cross-walk and validation against Google's ml_dtypes library establishes credibility and interoperability. The anchor vector mechanism using phi^2 + 1/phi^2 = 3 serves as an elegant sanity check across test packs, ensuring consistency.

For the AI hardware and software ecosystem, this standardization effort reduces friction in multi-vendor deployments. Organizations evaluating different accelerators can now perform bit-exact numerical validation rather than relying on approximate benchmarks or post-deployment debugging. The open-source release under a permissive license removes vendor lock-in concerns and establishes a foundation for broader adoption.

Looking ahead, this catalog could accelerate the industry's movement toward interoperable numeric standards. As heterogeneous computing becomes standard in production ML pipelines, such reference materials become increasingly valuable. The work's focus on documentation rather than format innovation positions it as infrastructure rather than competitive technology.

• →A new 84-format catalog with bit-exact conformance vectors enables vendors-neutral comparison of ML numeric formats including FP8, BF16, and MXFP4.
• →JSON-encoded test packs with SHA-256 fingerprints and anchor vectors provide reproducible cross-platform validation mechanisms.
• →Validation against Google's ml_dtypes 0.5.4 and IEEE standards mapping establish credibility and reduce implementation ambiguity.
• →Open-source release under permissive licensing removes vendor lock-in and supports broader ML infrastructure standardization.
• →The catalog addresses silent numerical divergences that currently plague multi-accelerator model deployment pipelines.