fig. 2.2.3 · the screen route, tile by tile: bounds computed, pairs kept, full dot products.⛶ TILE 11 OF 36 AT 0.90: THE BOUND KEPT 961 OF 1,048,576 PAIRS, SO 961 FULL DOTS; 1 FOUND THRESHOLD 0.80 KEEPS 21.5% 0.85 KEEPS 3.06% 0.90 KEEPS 0.087% 0.95 KEEPS 0.00024% ONE TILE: 1,024 × 1,024 ARXIV TITLE PAIRS. DOT: A PAIR THE RANK-64 BOUND COULD NOT RULE OUT (SHADE WHERE DENSE). RING: OVER THE THRESHOLD. 429 SPARSE 898 SPARSE 563 SPARSE 902 SPARSE 1,011 SPARSE 510 SPARSE 860 SPARSE 973 SPARSE 998 SPARSE 425 SPARSE 961 SPARSE SCREENED 11 OF 36 TILES PAIRS BOUNDED 9,435,136 KEPT BY BOUND 8,530 0.090% FULL DOTS 8,530 AT OR OVER 0.90 3 TILE 11 · ROWS 0–1,023 × 4,096–5,119 OVER 0.90 BOUNDS COMPUTED 1,048,576 KEPT BY BOUND 961 0.092% FULL DOTS 961 AT OR OVER 0.90 1 0.092% KEPT, UNDER THE 1% LINE: ONLY THE SURVIVORS GET A FULL 768-D DOT. STEP THE TILES BELOW, OR CLICK A THRESHOLD. STEP · RUN · RESET 0 36 TILES TILE 11 OF 36 TILE 11 OF 36 AT 0.90: THE BOUND KEPT 961 OF 1,048,576 PAIRS, SO 961 FULL DOTS; 1 FOUND
source: gpu-threshold-join arxiv-100000.npy rows 0-49,999, CPU replay of the screen route (_screen_math, _screen_policy, _workspace) · screening.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 .