Data¶
apple-m2 (NEON) indicative slice — one ISA¶
Grade: indicative bench run, NOT a ledger publication. Direct Google Benchmark run of
bench_compare on apple-m2-mba (NEON), --benchmark_min_time=0.5s, machine not in the pinned-
quiet state the ledger requires (this was an active dev session). Numbers are real and stable
across repeats (see below), but they are not CV-gated, empty-deviations ledger entries, so
they carry no qle: id. The publishable per-µarch entries are pending a quiet registered run
(R-06). Reported to characterize one point of H2, not to conclude the study.
n = 65536 (DRAM-resident), i64, kGt, wall-clock per call:
| selectivity | bitmap (ns) | selvec (ns) | selvec / bitmap |
|---|---|---|---|
| 1 % | 20205 | 22402 | 1.11 |
| 10 % | 20295 | 21982 | 1.08 |
| 50 % | 20161 | 22293 | 1.11 |
| 90 % | 20125 | 22011 | 1.09 |
| 99 % | 20227 | 22024 | 1.09 |
Stability check (sel=50, two extra repeats): bitmap 20034 / 21313 ns, selvec 22283 / 22018 ns —
the ~7–11 % selvec penalty is robust to run-to-run noise on this host.
Publishable ledger entries — PENDING (R-06)¶
The CV-gated, entry-referenced data across ≥3 ISAs and ≥5 µarchs is not yet collected — only
apple-m2-mba is registered. When x86 (AVX2, AVX-512) and a Graviton-class ARM part are
registered, quiver_ledger.py run --machine <id> --filter compare produces the entries per
method.md, committed under ledger/results/ and referenced here as qle:<entry_id>.
No cross-µarch numbers are recorded until then (Charter T2 — no invented data).
Traceability: bench_compare (selvec_gt/bitmap_gt); ledger machine apple-m2-mba; gate M9.