📖 Ch.9 — GF vs FP16 BF16 MXFP4 baseline comparison (700w · 🔴 P0 · II — NUMERIC)

[gHashTag]

Ch.9 — GF vs FP16 BF16 MXFP4 baseline comparison

Master: trios#380 · Part: II — NUMERIC
Scope: software-only (Rust/Zig), no hardware/FPGA/KOSCHEI
Words: 700 · Priority: 🔴 P0 · Owner: bench-agent

🎯 Scope (FROZEN, v2.1 SOFTWARE-ONLY)

State-of-art baseline comparison: GF16 vs FP16 vs BF16 vs MXFP4 (arXiv:2510.01863 microscaling). Required by NeurIPS 2026 reproducibility checklist — must compare against current SOTA. Software roundtrip MSE on log-normal weights, ML-typical distributions.

🔑 Key claims / formulas

• MX/MXFP4 baseline (arXiv:2510.01863, arXiv:2510.14557 MX+)
• AdaptivFloat (Tambe 2019)
• BBFP NeurIPS 2020
• Roundtrip MSE per format
• Why GF16 wins φ-distance among ≤16-bit (in golden corridor)
• HONEST: does GF16 win on actual ML metric? OR only on φ-distance?

📦 Deliverables

• paper/sections/9_gf_vs_fp16_bf16_mxfp4_baseline_comparison.tex

✅ Definition of Done (v2.1)

• Word count 700 ±10%
• All claims cited (primary literature)
• No woody-shop / KOSCHEI / FPGA references
• Compiles in main.tex via tectonic
• L1 (no .sh) · L3 · L4 · L8

🤖 ONE-SHOT directive (when operator types ONE SHOT Ch.9)

A: take Markdown draft from §"Draft", convert to LaTeX, place in
paper/sections/9_<slug>.tex, ensure compile, open PR with Closes #<this>.

phi^2 + phi^-2 = 3 · TRINITY · NEVER STOP