Every run on a rented card: what it asked, what it found, and which
coefficients its predictions came from. Generated by python3 bench/runs.py
from the front matter of each report, so edit the report, never this page
(adding-a-run.md §4).
Every prediction the runs faced, drawn: predicted-vs-measured.svg. The sheet written before each run: runsheets/. The evidence, verbatim: raw/.
Coefficients in use: l40s-run1 — eff_mem 0.83, mfu 0.439: run 1's fit, faced by runs 2–3.
| Run | Date | Asked | Found | Coefficients |
|---|---|---|---|---|
| 1 | 2026-08-18 | What does the card really deliver against the spec-sheet coefficients? | eff_mem 0.83 and mfu 0.439 replace the guesses, and TPOT p99 allows 12 seats where the arithmetic said 32 | predicted with l40s; fitted l40s-run1 |
| 2 | 2026-08-23 | Do run 1's coefficients hold on another pod, under real arrivals, other chunk sizes and FP8 KV? | eff_mem 0.83 held to 0.01 %, FP8 KV doubles the pool exactly, and past the goodput peak a useful token costs 15× more | predicted with l40s-run1 |
| 3 | 2026-08-30 | How many seats is an 80 % prefix cache hit rate worth? | 3× the seats, 12–14 to 32–40, because the decode step itself fell where the derivation said it could not | predicted with l40s-run1 |
Coefficients in use: mi300x-run2 — eff_mem 0.57, mfu 0.247: run 2's fit, faced by runs 3–5.
| Run | Date | Asked | Found | Coefficients |
|---|---|---|---|---|
| 1 | 2026-09-27 | Does the arithmetic transfer to the card the SLO was derived for? | eff_mem 0.46 and mfu 0.166, a third and nearly two thirds below the priors, and latency rather than room ends the seat curve | predicted with mi300x; fitted mi300x-run1 |
| 2 | 2026-09-27 | Which server lever moves the card, one at a time, and does run 1's fit hold? | a second droplet ran the decode step 16–22 % faster, so run 1's fit did not hold, and the card was refitted at 0.57 / 0.247 | predicted with mi300x-run1; fitted mi300x-run2 |
| 3 | 2026-09-29 | Do two engines on one card, behind the prefix router, beat one engine? | the memory bill held and the pair out-produced one engine, but the router comparison was void and the two engines were unequal | predicted with mi300x-run2 |
| 4 | 2026-10-17 | Why was the pair unequal, and where does affinity routing beat round robin? | the engine launched first is the slow one, and with the instrument fixed the two routings part at 127 prefixes | predicted with mi300x-run2 |
| 5 | 2026-10-17 | What makes two engines on one card equal, and does a 70B obey run 2's coefficient? | one hardware queue per process ends the asymmetry, and the 70B's decode step held to run 2's eff_mem at TP = 1 | predicted with mi300x-run2 |