I measured flat scalar binary search and bounded SIMD pivots over identical 32-, 64- and 128-bit keys. The SIMD path wins some short windows, but a single size cutoff is a poor default. In particular, scalar bit lifting needs only four comparisons for 15 keys and avoids range clamps at that complete window.

These are key-array lookup measurements. They do not include fractional-index routing, value access, page faults or database publication. The .ff and .fv formats do not yet have a complete production query reader.

Measured choices

The table gives the range of scalar/SIMD latency ratios across eight workloads for each width and count. A ratio above one favors SIMD. These are medians of three process medians, not confidence intervals.

Key bits	Entries	M2 Max NEON	i9-12900K AVX2
32	4	1.26–2.36	1.21–6.25
64	4	1.43–2.71	1.50–4.99
128	4	1.46–3.21	1.34–1.87
32	15	0.58–0.98	0.87–1.47
64	15	0.90–1.14	0.32–0.86
128	15	1.07–1.63	0.43–0.87
128	32	1.14–2.19	0.49–0.98

Automatic dispatch uses conservative measured choices:

NEON: four entries at every width; also two, three, fifteen, sixteen and thirty-two entries for 128-bit keys.
AVX2: four entries for 32-bit keys; two through four for 64-bit keys; two or four for 128-bit keys.
Other targets: scalar binary search. Explicit SIMD remains available.

A separate M2 Max run checks the resulting automatic path. For 128-bit keys, its scalar/automatic ratios are 1.10–1.76 at fifteen entries, 1.32–1.95 at sixteen, and 1.12–2.41 at thirty-two. Every selected automatic case has a favorable median in that run. Dispatch is not free: unselected cases can differ from explicit binary search because of the wrapper and generated code; 32-bit fifteen-entry ratios span 0.90–1.04. I keep both explicit methods available. The AVX2 run measures the selected primitives, not a second timing run of the final dispatch wrapper. Upper bounds have correctness coverage but these timing runs measure lower bounds only.

Method

fixed_search.cc tests every count from one through thirty-two, plus a sixty-three-entry binary fallback control. Keys have either random words or shared leading words. Each case uses one hot window or 8192 windows, varied byte alignment, and independent or result-dependent requests. The allocated window footprint ranges up to about 8.4 MiB; this is resident-memory traffic, not a cold-storage measurement.

Each collection has three fresh processes, each with three rotated repeats of 65536 lookups per method: 21,384 timed rows covering 792 workloads. Independent std::lower_bound results validate warmups and every timed checksum, including the dependent recurrence. Fixture creation and validation are outside the timed region. No timing rows were discarded.

The M2 Max uses Apple Clang 21 at -O3, with local worker builds and timings paused. The AVX2 run uses Ubuntu Clang 20, pinned to P-core CPU 4 on an i9-12900K, with -march=x86-64 -mavx2 -mpopcnt -mno-avx512f. This does not establish AVX-512 performance or a universal crossover on other processors.

Protected-page tests separately cover empty spans, short tails, unaligned starts and valid page crossings. NEON and portable sanitizer runs passed; SSE2 and AVX2 also ran on x86. AVX-512 has compile coverage only.

Reproduction and evidence

c++ -std=c++20 -O3 -DNDEBUG -Iinclude bench/fixed_search.cc -o fixed-search
mkdir fixed-results
./fixed-search 3 65536 > fixed-results/process-0.csv
./fixed-search 3 65536 > fixed-results/process-1.csv
./fixed-search 3 65536 > fixed-results/process-2.csv
python3 bench/fixed_search_analyze.py fixed-results

The retained collections identify the source and binary hashes for each run. Each collection contains raw CSVs in raw.tar.gz, all per-case medians and process ranges in cases.csv, and per-width/count summaries in summary.csv. Extract the raw archive into its directory and run the analyzer to reproduce its derived tables. The selection runs used header revision 5263d76 and benchmark revision cb2b029; the M2 dispatch run records its changed header hash separately. Current source reproduction makes a new measurement rather than replacing those observations.