Andrew Fribush

fig. 2.20.1: sized on real records, 16 boxes finish in the window; sized on synthetic strings, 8 boxes take twice as long

fig. 2.20.1 · sized on real records, 16 boxes finish in the window; sized on synthetic strings, 8 boxes take twice as long.
REAL RECORDS: 16 BOXES IN THE WINDOW; SYNTHETIC: 8 TAKE 13.6 HONE SQUARE IS ONE L40S; BAR IS HOURS; GREY BAND IS 6 TO 10 H.DEV BOX · JIT FALLBACK44 REC/S · 140 BOXES0.8 HSYNTHETIC STRINGS858 REC/S · 8 BOXES13.6 H · 2.0X THE WINDOWREAL RECORDS383 REC/S · 16 BOXES6.8 H · IN THE WINDOW0 H6121824WINDOW 6.8 H

Boxes by the runbook’s formula, box-hours = 150,000,000 / rate / 3,600, rounded up to fill the window. The run used 16 boxes at 383 records a second on one L40S each, 6.8 planned hours a box; the last vector landed 5.98 hours after the first.

source: credit_150m_embedding_metrics.csv lines 3, 13, 14, 24, 26, 43; ec2_qwen3_768_embed_calibration.md lines 5, 108-110 · fleetSizer.data.ts

View in context: Embedding 150 million records for about $230 → · 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.