fig. 2.2.5 · where screening spends its kernel time, on selective and unselective data.⛶ GPU KERNEL TIME BY KERNEL · FORCED SCREEN ROUTE · RTX A6000 · 50,000 ROWS ONE BAR IS ONE KERNEL’S SHARE OF ITS QUERY’S GPU KERNEL TIME; THE ACCENT IS THE KERNEL THAT DOMINATES. BOUND COUNT AND SCATTER PROJECTION GEMM SPARSE FULL DOTS CLASSIFICATION SCANS ARXIV TITLES · THRESHOLD 0.9 SCREENING PAYS · 47.27 MS OF KERNEL TIME 0% 50% 100% 31.32 MS · 66% 12.01 MS · 25% 3.29 MS · 7.0% 0.17 MS · 0.4% 0.48 MS · 1.0% THE BOUND WORK, NOT THE DOTS, IS THE BILL ISOTROPIC CONTROL · THRESHOLD 0.8 SCREENING DOES NOT PAY · 5,320.64 MS OF KERNEL TIME 0% 50% 100% 69.78 MS · 1.3% 33.44 MS · 0.6% 5,171.23 MS · 97% 43.68 MS · 0.8% 2.51 MS · <0.1% FULL DOTS 1,574× ARXIV’S · TOTAL 113×
Nsight Systems kernel sums inside the query’s own NVTX range, preparation and warmup excluded: instrumented kernel components, not wall-clock timings.
source: gpu-threshold-join benchmarks/preserved/engineering-finish-2026-09-18/profiles arxiv-09 and isotropic-08 nvtx_kern_sum.csv (RTX A6000, d1a7f1e) · bottleneck.data.ts
View in context: Knuth's Root of All Evil Has a Break-Even Point → · All figures
Andrew Fribush , software and computer engineer in Austin. I build consequential systems where software meets institutions: finance, healthcare, commerce, government data, and the mechanisms that govern AI agents. Taking new engagements: hello@andrewfribush.com .