Topic
Benchmarking
Timing, harnesses and the gap between a reading and a result.
Lab
- 17 Sept 2026The second instrument
- 21 Aug 20261.024×, and not rounded up
- A 16×16 window at 9.92 ns against 4.85 ns by hand
- Two frames in three seconds, and a mechanism for the rest
- 4.93× against a 5.11× ceiling
- A flat result, reported flat
- Rows nobody retired, stale by 5.76×
- Struck as UNMATCHED, then falsified at 1.24× slower
- A bake-off invalid by construction: N_prop 80 vs 16
- 3.754× behind ggml, on weaker footing
- 24 bytes inside the timed region