A Tuned Hand-Written GEMM Reached 0.99 of cuBLAS, and the Launch-Overhead Rule Failed: Five Pre-registered Tests of Textbook GPU Rules on One Consumer Laptop GPU

Textbook GPU performance rules (coalescing multiplies bandwidth, occupancy theory predicts the best block size, kernel launches cost under 10 microseconds, shared-memory tiling speeds up matrix multiply, a tuned hand-written GEMM still trails the vendor library) are usually taught as facts and rarely measured under frozen rules on named hardware. This study pre-registered all five as testable claims with numeric verdict lines, froze the protocols and rules in the repository's scaffold commit before any harness code or measurement existed, and ran them on one named device (NVIDIA GeForce RTX 5050 Laptop GPU, 8 GB, on the host BlackBox). Of the five verdicts, two are SUPPORTED, two are PARTIALLY SUPPORTED, and one is REFUTED. A shared-memory tiled GEMM beat a naive GEMM by 8.62 times at the verdict size 1024 (bootstrap 95 percent interval 8.35 to 8.68), far above the frozen 1.5 times line. A tuned tiled GEMM reached 0.990 of cuBLAS SGEMM throughput at size 2048 (interval 0.98 to 1.02), against a frozen support line of 0.25. Coalesced access reached 4.22 times the bandwidth of stride-32 access (305.2 against 72.3 GB/s, interval 4.07 to 4.31), below the frozen 5 times line, so that hypothesis is only partially supported. Achieved occupancy at the fastest block size sat 28.0 percentage points below the theoretical curve, outside the frozen 10 point band, although the fastest block size was exactly the plateau onset the theory predicts, so that hypothesis is also partially supported. The launch-overhead rule failed outright: the median per-launch cost of an empty kernel was 12.28 microseconds against a frozen line of below 10 microseconds, and the hypothesis is REFUTED. Every figure in this paper was recomputed from the committed raw timing arrays for this preprint, and the study's own interval analysis regenerates byte-identically. The claims are deliberately narrow: one consumer laptop GPU, one session per study, clocks never locked, and no device peak stated anywhere in the record, so no figure here is a fraction of peak.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-18
DOI
https://doi.org/10.5281/zenodo.24329059
Primary Topic
Parallel Computing and Optimization Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

A Tuned Hand-Written GEMM Reached 0.99 of cuBLAS, and the Launch-Overhead Rule Failed: Five Pre-registered Tests of Textbook GPU Rules on One Consumer Laptop GPU

Jesun Ahmad Ushno
1 links
Zenodo (CERN European Organization for Nuclear Research)
Parallel Computing and Optimization Techniques
preprint

A Tuned Hand-Written GEMM Reached 0.99 of cuBLAS, and the Launch-Overhead Rule Failed: Five Pre-registered Tests of Textbook GPU Rules on One Consumer Laptop GPU

Jesun Ahmad Ushno
preprint en

Abstract

Textbook GPU performance rules (coalescing multiplies bandwidth, occupancy theory predicts the best block size, kernel launches cost under 10 microseconds, shared-memory tiling speeds up matrix multiply, a tuned hand-written GEMM still trails the vendor library) are usually taught as facts and rarely measured under frozen rules on named hardware. This study pre-registered all five as testable claims with numeric verdict lines, froze the protocols and rules in the repository's scaffold commit before any harness code or measurement existed, and ran them on one named device (NVIDIA GeForce RTX 5050 Laptop GPU, 8 GB, on the host BlackBox). Of the five verdicts, two are SUPPORTED, two are PARTIALLY SUPPORTED, and one is REFUTED. A shared-memory tiled GEMM beat a naive GEMM by 8.62 times at the verdict size 1024 (bootstrap 95 percent interval 8.35 to 8.68), far above the frozen 1.5 times line. A tuned tiled GEMM reached 0.990 of cuBLAS SGEMM throughput at size 2048 (interval 0.98 to 1.02), against a frozen support line of 0.25. Coalesced access reached 4.22 times the bandwidth of stride-32 access (305.2 against 72.3 GB/s, interval 4.07 to 4.31), below the frozen 5 times line, so that hypothesis is only partially supported. Achieved occupancy at the fastest block size sat 28.0 percentage points below the theoretical curve, outside the frozen 10 point band, although the fastest block size was exactly the plateau onset the theory predicts, so that hypothesis is also partially supported. The launch-overhead rule failed outright: the median per-launch cost of an empty kernel was 12.28 microseconds against a frozen line of below 10 microseconds, and the hypothesis is REFUTED. Every figure in this paper was recomputed from the committed raw timing arrays for this preprint, and the study's own interval analysis regenerates byte-identically. The claims are deliberately narrow: one consumer laptop GPU, one session per study, clocks never locked, and no device peak stated anywhere in the record, so no figure here is a fraction of peak.

Zenodo (CERN European Organization for Nuclear Research)
Parallel Computing and Optimization Techniques
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