Arithmetic Series of Block Formats W9(n) = 32n for Exact Affine vec4 * mat4 Transformations: Closure Theorems, Hardware Profiles, Performance Model, and Microarchitectural Analysis

Bulk evaluation of affine transformations y = M3x3 * v + w * t (vec4 * mat4) on modern graphics processing units (GPUs) is severely memory-bandwidth bound at low vector-to-matrix reuse ratios. Standard IEEE 754 formats (FP16, FP32, FP64) store a homogeneous matrix as 16 isolated scalars, allocating memory to the redundant fourth row [0, 0, 0, 1] and independent exponents per component, while introducing intermediate rounding errors and catastrophic cancellation risks during dot-product evaluation—even in mixed-precision modes (FP16 → FP32, FP32 → FP64). This work formally defines a hardware-independent arithmetic series of shared-exponent block formats governed by W9(n) = 9(3n) + 5n = 32n bits for positive integers n, yielding periodic effective bit-widths of W1(n) = 32n/9 bits per scalar (starting at 3.555... bits) and W3(n) = 32n/3 bits per triplet (starting at 10.666... bits). For inputs residing on the dyadic grid D_n, two exact closure theorems are proved, guaranteeing zero arithmetic and output error (e_arith = e_out = 0): a general level promotion n → 2n + 1 for arbitrary linear matrices, and a tightened promotion n → 2n for orthogonal rotations derived via the Cauchy–Schwarz inequality. Based on exact bit-width thresholds of contemporary instruction set architectures (n ≤ 5, n ≤ 10, n ≤ 21), four operating profiles (Profile 1 through Profile 4) are established, delineating current GPU/CPU capabilities from new hardware requirements for n ≥ 24. Quantitative comparison demonstrates that at identical matrix memory footprints, the proposed series delivers a systematic precision advantage of +6 bits over FP16 (64× lower error), +11 bits over FP32 (2,048× lower error, ±14.9 nm at 1 km), and +18 bits over FP64 (262,144× lower error). Finally, a Roofline-based derivation of performance inequalities, five rules of computational behavior, and a microarchitectural analysis of current GPUs precisely define the conditions under which execution outperforms IEEE 754, while explaining the physical causes of slowdown in L1-resident loops.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24419663
Primary Topic
Numerical Methods and Algorithms
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Arithmetic Series of Block Formats W9(n) = 32n for Exact Affine vec4 * mat4 Transformations: Closure Theorems, Hardware Profiles, Performance Model, and Microarchitectural Analysis

Michal Mazgal
Zenodo (CERN European Organization for Nuclear Research)
Numerical Methods and Algorithms
preprint

Arithmetic Series of Block Formats W9(n) = 32n for Exact Affine vec4 * mat4 Transformations: Closure Theorems, Hardware Profiles, Performance Model, and Microarchitectural Analysis

Michal Mazgal
preprint en

Abstract

Bulk evaluation of affine transformations y = M3x3 * v + w * t (vec4 * mat4) on modern graphics processing units (GPUs) is severely memory-bandwidth bound at low vector-to-matrix reuse ratios. Standard IEEE 754 formats (FP16, FP32, FP64) store a homogeneous matrix as 16 isolated scalars, allocating memory to the redundant fourth row [0, 0, 0, 1] and independent exponents per component, while introducing intermediate rounding errors and catastrophic cancellation risks during dot-product evaluation—even in mixed-precision modes (FP16 → FP32, FP32 → FP64). This work formally defines a hardware-independent arithmetic series of shared-exponent block formats governed by W9(n) = 9(3n) + 5n = 32n bits for positive integers n, yielding periodic effective bit-widths of W1(n) = 32n/9 bits per scalar (starting at 3.555... bits) and W3(n) = 32n/3 bits per triplet (starting at 10.666... bits). For inputs residing on the dyadic grid D_n, two exact closure theorems are proved, guaranteeing zero arithmetic and output error (e_arith = e_out = 0): a general level promotion n → 2n + 1 for arbitrary linear matrices, and a tightened promotion n → 2n for orthogonal rotations derived via the Cauchy–Schwarz inequality. Based on exact bit-width thresholds of contemporary instruction set architectures (n ≤ 5, n ≤ 10, n ≤ 21), four operating profiles (Profile 1 through Profile 4) are established, delineating current GPU/CPU capabilities from new hardware requirements for n ≥ 24. Quantitative comparison demonstrates that at identical matrix memory footprints, the proposed series delivers a systematic precision advantage of +6 bits over FP16 (64× lower error), +11 bits over FP32 (2,048× lower error, ±14.9 nm at 1 km), and +18 bits over FP64 (262,144× lower error). Finally, a Roofline-based derivation of performance inequalities, five rules of computational behavior, and a microarchitectural analysis of current GPUs precisely define the conditions under which execution outperforms IEEE 754, while explaining the physical causes of slowdown in L1-resident loops.

Zenodo (CERN European Organization for Nuclear Research)
Numerical Methods and Algorithms
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon
Arithmetic Series of Block Formats W9(n) = 32n for Exact Affine vec4 * mat4 Transformations: Closure Theorems, Hardware Profiles, Performance Model, and Microarchitectural Analysis — Michal Mazgal · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS