跳转至

Golden Ruler：包含FP8、BF16、MXFP4及微标度格式位精确一致性向量的数字格式目录

文章背景与核心概要

机器学习硬件的快速发展引入了大量的数字格式，包括 FP8（E4M3 和 E5M2）、BF16、MXFP4、微标度块格式以及众多研究变体。这种碎片化超出了供应商中立、位精确参考资料的发展速度，导致工程师在不同硬件加速器之间移植模型时经常遇到静默分歧（silent divergences）。

为了应对这一挑战，“Golden Ruler”作为数字格式的综合注册表和验证套件应运而生。它提供了 109 种格式的目录、六个位精确一致性包、IEEE P3109 v3.2.0 交叉映射，以及用于跨包完整性检查的严格验证与锚定机制。本文严格定位为注册表填充工作——不提出新格式，不作模型准确性声明，也不断言优于任何供应商的实现。

📌 Summary

The rapid evolution of machine learning hardware has introduced a proliferation of numeric formats—including FP8 (E4M3 and E5M2), BF16, MXFP4, microscaling block formats, and numerous research variants. This fragmentation has outpaced the availability of vendor-neutral, bit-exact reference material, often leading engineers to encounter silent divergences when porting models across different hardware accelerators.

To address this challenge, Golden Ruler serves as a comprehensive registry and validation suite for numerical formats. It provides: * A Catalog of 109 Formats: Spanning 12 distinct clusters, including the TNF, BNF, and GF-T ladders, with decimal representations folded into IEEE. * Six Bit-Exact Conformance Packs: Covering GF16, MXFP4 elements, BF16, FP8 E4M3, FP8 E5M2, and E8M0 block scales. * IEEE P3109 v3.2.0 Cross-Walk: Mapping each conformance pack directly to its corresponding standards-track configured format. * Rigorous Verification & Anchors: Each pack is a self-contained JSON document featuring a SHA-256 fingerprint, a shared row schema, and an anchor vector encoding the mathematical identity \(\phi^2 + 1/\phi^2 = 3\) as a cross-pack sanity check. Cross-validation is performed against ml_dtypes 0.5.4 (Google/JAX).

Note: This work is framed strictly as registry filling—it does not propose new formats, make model-accuracy claims, or assert superiority over any vendor's implementation.

📂 Access and Resources

📋 Additional Metadata

  • Submission History:
  • [v1] Mon, 8 Jun 2026
  • [v2] Mon, 22 Jun 2026
  • [v3] Fri, 4 Sep 2026 (Retitled to remove format count from the title; updated catalog size to 109 formats across 12 clusters; corrected Section 6 regarding tt-trinity-corona post-silicon status).
  • Subjects: Hardware Architecture (cs.AR), Artificial Intelligence (cs.AI), Mathematical Software (cs.MS), Performance (cs.PF), Numerical Analysis (math.NA)
  • Classifications:
  • MSC: 65Y04, 68N20
  • ACM: G.1.0; D.3.4; B.2.4