strawmANN benchmark report

strawmANN against Qdrant · AMD Ryzen AI 9 HX PRO 370 w/ Radeon 890M (24 logical) · generated 2026-10-08 08:54:50 CEST

Dataset: dbpedia-openai-1m
Vectors: 1,000,000 × 1,536 · 10,000 held-out queries
Metric: cosine
Storage datatype: strawmANN float32 · Qdrant float32 (default)
Quantization: none (fp32) · separate rows measure binary (1 bit/dim), PQ (product), SQ8 scalar, turboquant

strawmANN: 55/55 ok Qdrant: 55/55 ok

Summary

At equal recall, strawmANN serves 1.06x to 1.36x Qdrant's throughput. That range spans the recall levels measured; allowing for the uncertainty in each recall figure widens it to 0.85x to 1.46x.

throughput at equal recall

1.06 to 1.36x

strawmANN over Qdrant, recall@10 0.894 to 0.989

p99 latency at 90% load

strawmANN2.76 ms

Qdrant14.05 ms

W4-sat90, open loop, one offered rate

upload and index build

strawmANN259 s

Qdrant350 s

W1 + W2

peak memory (RSS)

strawmANN36.2 GiB

Qdrant43.6 GiB

largest before the concurrent-write rows (W11); of it, anonymous 4.0 GiB / 3.4 GiB

written to disk

strawmANN63.5 GiB

Qdrant330.8 GiB

before the concurrent-write rows (W11); they wrote 1.4 GiB / 22.0 GiB more

Throughput against recall

Read it vertically: at any recall both engines reach, the higher curve is faster. Throughput from bfb W10, recall from the conformance sweep, joined on ef. ef is the same unit here: Qdrant held 2 segments for this run, one populated and 1 empty (read back per segment from the engine), and strawmANN held one graph, so equal ef is equal traversal width and the curves may be read at equal x. That is a property of this run, not of the engines.

Every figure is the median of 3 passes per engine, run alternately (A/B/A/B). What this does not establish.

Glossary

recall@10: Of the ten nearest neighbours a query really has, the share the engine returned. 1.0 is a perfect answer. "Really has" is settled by an exhaustive fp64 search, not by the other engine.
ef: The candidate-list size: how many nodes the index keeps in play while searching. Larger is slower and more accurate. It does not mean the same amount of work in both engines, which is why the headline holds recall fixed instead.
MRDE: Mean relative distance error: when the engine returns a wrong neighbour, how much further away it is than the right one. Small numbers mean the misses were near-misses.
qps: Queries per second. On a batched row one request carries several queries, and the request rate is shown under it.
p50 / p99 / p99.9: The latency half of requests beat, that 99% beat, that 99.9% beat. The tail is what a user notices.
closed / open loop: A closed loop sends the next request only when the last one comes back, so a slow server receives less work and its tail looks better than it is. An open loop sends at a fixed rate regardless.
noise floor: How much a number moves between identical runs on this machine. A ratio inside it is shown grey with ≈: no measured difference. Hover a ratio to see its band.
conformance tier: What the two engines were shown to agree on before any speed was quoted. T1 licenses a single engine's own numbers; T3 licenses comparing the two, and requires their recall to be statistically indistinguishable.
§ numbers: Sections of docs/spec.md, the written rule each claim in the appendix is measured against. §7.1 is the host gate, §7.4 the comparison rules, §8 what may be published.

Throughput

Queries per second; higher is better. The ratio is strawmANN over Qdrant: green is faster, red slower, grey ≈ inside the noise floor (hover a ratio for its band). A dash means the pair is not compared, and the note says why. An ef sweep is one row showing its range; every measurement is under All throughput rows.

workload		strawmANN	Qdrant	ratio	notes
W3	search, fp32, single query	1,059	627	1.69x
W4	search, saturating (closed loop)	3,572	3,384	parity
W5	search batched (16 distinct dataset queries per request)	3,542250 requests/s	3,441237 requests/s	parity
W6	quantized: scalar	8,112	7,160	1.13x
W6 ef 32 to 512	SQ8 recall control (latency only)	2,457 to 24,808	2,411 to 17,155	1.06 to 1.45x	1 of 5 points at parity
W7	quantized: binary + oversampling	23,264	13,558	1.72x
W7-2bit	quantized: binary quantization, 2 bits + oversampling	20,459	12,314	1.66x
W7-1p5bit	quantized: binary quantization, 1.5 bits + oversampling	21,077	12,685	1.66x
W8	quantized: PQ	6,115	5,067	1.21x
W14-1bit	quantized: TurboQuant, 1 bit + oversampling	14,063	12,605	1.12x
W14-1p5bit	quantized: TurboQuant, 1.5 bits + oversampling	11,493	9,944	1.16x
W14-2bit	quantized: TurboQuant, 2 bits + oversampling	15,832	11,533	1.37x
W14-4bit	quantized: TurboQuant, 4 bits + oversampling	12,937	10,006	1.29x
W9	exact / brute force	13	10	1.27x
W10 ef 32 to 512	recall control (latency only)	1,170 to 10,076	1,051 to 7,343	1.14 to 1.37x	1 of 5 points at parity
W12-sel1	filtered search, one keyword (~1% of bench12)	3,958	1,395	2.84x
W12-sel10	filtered search, any of 10 keywords (~10% of bench12)	2,536	835	3.04x
W12-sel1 ef 32 to 512	filtered recall control, one keyword (latency only)	3,941 to 3,966	552 to 3,802	1.70 to 7.16x	1 of 5 points not compared recall differs
W12-sel10 ef 32 to 512	filtered recall control, any of 10 keywords (latency only)	443 to 7,481	338 to 2,160	1.31 to 3.03x	2 of 5 points not compared recall differs
W13	scroll / pagination	50,411	46,314	1.09x	strawmANN: 92% is client and socket, not server Qdrant: 82% is client and socket, not server
W11-steady	mixed read/write below the rebuild threshold: search bench2 while 49,500 synthetic points append	541	494	-	one pass apart: strawmANN pass 1 +15%
W11	mixed read/write: search bench2 while 198,000 synthetic points append (runs last)	121	72	-	drifted: strawmANN -11% across passes

Throughput by workload

Higher is better. Hatched bars are rows the table does not compare (W11): each bar is that engine's own rate, and the pair is not a result.

Recall

recall@10 against an exact fp64 search, not against the other engine. Equal ef is not equal work in the two engines, so the comparison that counts holds recall fixed and compares throughput there.

At matched recall

recall@10	measured at	strawmANN q/s	Qdrant q/s	strawmANN / Qdrant	range over recall CI
0.8936	Qdrant at ef=32, strawmANN interpolated	9,986	7,343	1.36x	1.26 – 1.46x
0.9400	strawmANN at ef=64, Qdrant interpolated	6,082	4,931	1.23x	1.18 – 1.29x
0.9434	Qdrant at ef=64, strawmANN interpolated	5,688	4,790	1.19x	1.08 – 1.31x
0.9667	strawmANN at ef=128, Qdrant interpolated	3,581	3,198	1.12x	1.05 – 1.20x
0.9692	Qdrant at ef=128, strawmANN interpolated	3,255	3,062	1.06x	0.92 – 1.23x
0.9811	strawmANN at ef=256, Qdrant interpolated	2,058	1,891	1.09x	0.97 – 1.23x
0.9822	Qdrant at ef=256, strawmANN interpolated	1,912	1,811	1.06x	0.85 – 1.31x
0.9894	strawmANN at ef=512, Qdrant interpolated	1,170	1,102	1.06x	0.91 – 1.24x

How this is computed

Interpolated linear in log(q/s) between the two bracketing measurements, nothing extrapolated past the measured range. Throughput is bfb's W10 sweep, recall the conformance sweep, joined on ef — same m and ef_construct, but separate builds, so the pairing assumes two builds with identical parameters are equivalent.

The last column is why the ratio is not a result on its own. The ratio treats the anchor's recall as exact; it is an estimate, and moving it across its 95% interval moves the interpolated rate with it — near the top of the sweep 0.003 of recall spans a factor of 1.8, so two decimals there quote the interpolation. Even that is the narrow reading: it moves one recall and not the other, adjacent anchors share bracketing segments, and the throughputs behind it are medians of the run's passes.

At matched recall, SQ8

The same reading over the scalar-quantized collection. Under the `pool` policy both engines rescore `ef` candidates, so W6 is compared row by row as well; held at equal recall, the reading does not depend on the rows landing in one recall band. Note where each engine's curve stops, because a recall only one of them reaches is the more useful fact about an encoding than any ratio.

recall@10	measured at	strawmANN q/s	Qdrant q/s	strawmANN / Qdrant	range over recall CI
0.8916	strawmANN at ef=32, Qdrant interpolated	24,808	16,709	1.48x	1.41 – 1.56x
0.9395	strawmANN at ef=64, Qdrant interpolated	14,496	11,599	1.25x	1.20 – 1.30x
0.9396	Qdrant at ef=64, strawmANN interpolated	14,471	11,592	1.25x	1.08 – 1.44x
0.9666	strawmANN at ef=128, Qdrant interpolated	8,120	7,329	1.11x	1.04 – 1.18x
0.9676	Qdrant at ef=128, strawmANN interpolated	7,797	7,209	1.08x	0.90 – 1.30x
0.9808	strawmANN at ef=256, Qdrant interpolated	4,485	4,416	1.02x	0.91 – 1.13x
0.9818	Qdrant at ef=256, strawmANN interpolated	4,172	4,251	0.98x	0.80 – 1.21x
0.9893	strawmANN at ef=512, Qdrant interpolated	2,457	2,496	0.98x	0.85 – 1.14x

How this is computed

Interpolated linear in log(q/s) between the two bracketing measurements, nothing extrapolated past the measured range. Throughput is bfb's W6 sweep, recall the conformance sweep, joined on ef — same m and ef_construct, but separate builds, so the pairing assumes two builds with identical parameters are equivalent.

The last column reads as it does in the fp32 table above.

At matched recall, filtered to 10%

The same reading under a keyword filter over the points the condition matched, 19,810 to 20,087 across 3 builds in strawmANN and 19,841 to 20,038 across 3 builds in Qdrant (bfb draws the keyword payloads unseeded at each upload). The per-row table compares W12-sel10 at the one ef it measures; held at equal recall, the comparison extends to every recall both engines reach. Note the shape rather than any single number: one engine's curve is flat in ef and the other's is steep, so where you match decides the ratio.

recall@10	measured at	strawmANN q/s	Qdrant q/s	strawmANN / Qdrant	range over recall CI
0.9699	Qdrant at ef=32, strawmANN interpolated	4,630	2,160	2.14x	2.01 – 2.29x
0.9728	strawmANN at ef=64, Qdrant interpolated	4,433	2,019	2.20x	2.02 – 2.39x
0.9893	strawmANN at ef=128, Qdrant interpolated	2,517	1,380	1.82x	1.73 – 1.92x
0.9901	Qdrant at ef=64, strawmANN interpolated	2,355	1,354	1.74x	1.46 – 2.07x
0.9959	strawmANN at ef=256, Qdrant interpolated	1,468	909	1.62x	1.47 – 1.77x
0.9972	Qdrant at ef=128, strawmANN interpolated	1,012	832	1.22x	0.88 – 1.68x
0.9993	Qdrant at ef=256, strawmANN interpolated	546	522	1.05x	0.87 – 1.26x
0.9998	Qdrant at ef=512, strawmANN interpolated	475	338	1.41x	1.26 – 1.57x

How this is computed

Interpolated linear in log(q/s) between the two bracketing measurements, nothing extrapolated past the measured range. Throughput is bfb's W12-sel10 sweep, recall the conformance sweep, joined on ef — same m and ef_construct, but separate builds, so the pairing assumes two builds with identical parameters are equivalent.

The last column reads as it does in the fp32 table above.

Recall by `ef`

ef	strawmANN				Qdrant
ef	recall@1	recall@10	recall@100	MRDE	recall@1	recall@10	recall@100	MRDE
ef=32	0.8966	0.8928	-	1.78e-03	0.8994	0.8936	-	1.84e-03
ef=64	0.9371	0.9400	-	8.94e-04	0.9412	0.9434	-	8.80e-04
ef=128	0.9663	0.9667	0.9290	4.64e-04	0.9665	0.9692	0.9309	4.44e-04
ef=256	0.9803	0.9811	0.9657	2.61e-04	0.9816	0.9822	0.9678	2.48e-04
ef=512	0.9890	0.9894	0.9833	1.40e-04	0.9907	0.9900	0.9847	1.33e-04

How recall was measured

10,000 held-out queries, limit 10, ε=9.537e-07, against the fp64 oracle rather than against the other engine. Base checksum 71f1df815731cdbf: the same corpus the latency rows were measured on. Each pass rebuilds the collection. Across 3 independent builds of it, recall@10 at ef=512 spread by: strawmANN 0.00040; Qdrant 0.00049. Over the same builds the engine reported unreachable nodes: strawmANN 0 of 200,000 and 0 of 990,000 — a node nothing points at is invisible at any ef, so that is where a graph-quality difference shows rather than being inferred from the recall beside it. Level seed: strawmANN 0x57ea3111. Recorded as provenance: under this engine's level draw, six builds at four seeds sit within 0.00003 of recall@10 at ef 512, well inside the spread above.

What each quantization costs in recall

Colour is the encoding here, not the engine — the engines are the line style. Each sweep is joined on the quantization parameters it actually sent, so a curve speaks only for the search its row ran. Higher is better, and the encodings are not free: read this against the throughput their rows bought.

Latency

Client-side round trip. A closed loop understates the tail (a stalled server stops receiving requests), so this keeps the fixed-rate, open-loop rows, which offer the same load to both engines, plus single-client W3 and batched W5. Every row is under All latency rows.

workload	strawmANN			Qdrant
workload	p50	p99	p99.9	p50	p99	p99.9
W3 search, fp32, single query	948 µs	1.38 ms	1.90 ms	1.59 ms	2.27 ms	2.51 ms
W4-sat50 search, fixed rate at 50% of saturation (open loop)	1.02 ms	1.62 ms	1.84 ms	1.74 ms	2.84 ms	3.71 ms
W4-sat70 search, fixed rate at 70% of saturation (open loop)	1.11 ms	1.93 ms	2.28 ms	2.00 ms	3.79 ms	4.76 ms
W4-sat90 search, fixed rate at 90% of saturation (open loop)	1.38 ms	2.76 ms	3.44 ms	3.13 ms	14.05 ms	23.35 ms
W5 search batched (16 distinct dataset queries per request)	288.12 ms	294.42 ms	303.98 ms	295.78 ms	339.82 ms	402.78 ms

Ingest and index build

Qdrant indexes while it ingests, so its upload time already contains most of the indexing; strawmANN uploads raw and builds afterwards. Compare the sum, not the upload line.

Where the load time goes

Solid is upload, hatched is the wait for Green. The bar's whole length is the sum this section asks you to compare; the split is why the upload line alone is not comparable between these engines.

workload		strawmANN	Qdrant
W0-upload	d=4 floor: graph traversal with the distance taken out	17.05 s	22.05 s
W1	ingest throughput (no index wait)	5.46 s	34.33 s
W2	index build, time to Green	253.88 s	315.80 s
W6-upload	search, SQ8 scalar quantization	144.46 s	188.44 s
W7-upload	search, binary quantization	50.14 s	171.48 s
W7-2bit-upload	search, binary quantization	68.18 s	134.32 s
W7-1p5bit-upload	search, binary quantization	66.18 s	205.53 s
W8-upload	search, product quantization	375.31 s	886.32 s
W14-1bit-upload	upload	259.93 s	115.27 s
W14-1p5bit-upload	upload	262.94 s	159.39 s
W14-2bit-upload	upload	260.86 s	177.44 s
W14-4bit-upload	upload	262.91 s	204.51 s
W12-upload	upload	39.12 s	74.19 s
W1 + W2	upload and index, together	259.3 s	350.1 s	1.35x

Memory and disk

What each engine held in memory and moved to and from disk over the whole run. Per-row figures are under Storage and I/O.

	strawmANN	Qdrant
storage on disk	35.5 GiBallocated at --capacity	26.5 GiB
peak memory (RSS)	36.2 GiB	43.6 GiB
read from disk	74.1 MiB	2.6 GiB
written to disk	63.5 GiB	330.8 GiB

Peak memory and storage on disk are read before the concurrent-write rows (W11): an engine rewriting segments maps old and new files at once, and RSS counts each mapping. Bytes written are summed over the same rows, since W11's volume is set by the harness's write rate; the other disk figures are totals over every row.

Appendix

How the run was set up, every row of every table, and the diagnostics behind them. Charts on a logarithmic axis say so on the axis: read the positions there, not the distances.

Run conditions and conformance

Measured as-deployed: the scheduler was left as it ships, so a difference is what a user would see rather than the engine in isolation, and ambient load is part of the measurement. Qdrant ran equal-work: asked for one populated graph (--segments 1), as strawmANN serves, so ef means the same thing on both sides. This is the configuration §8's comparative licensing is built around, and it is a control rather than a deployment. Its optimizer treats the count as a target; what it actually held is read back from the engine and stated under Collection config.

Conformance tier reached: T4 quantization fidelity. T1 passed, so single-engine performance rows are licensed (§8). T3 passed, so a strawmANN-vs-Qdrant throughput comparison is licensed: the engines are at equal recall (§7.4). Conformance hash cf021f3ef3d16b39.

On exact search the two engines' scores differ by at most 3.576e-07 (p99 1.788e-07), against a calibrated absolute ε of 9.537e-07. §8.1: bit-exactness is unachievable between two different summation orders, so this — equal values within a measured tolerance — is the claim the speed rests on.

The conformance harness's single-client rate does not order the engines consistently (0.48x to 2.59x); a different instrument, so no ratio is formed from it.

Every figure is the median of 3 passes per row per engine, alternated A/B/A/B (§7.4), and the spread is this run's own over 42 of 55 rows. The ingest and index-build rows (W0-upload, W1, W2, W6-upload, W7-upload, W7-2bit-upload, W7-1p5bit-upload, W8-upload, W14-1bit-upload, W14-1p5bit-upload, W14-2bit-upload, W14-4bit-upload, W12-upload) have none, so no build time carries a verdict. Each pass rebuilds the index, so the bands include build variance and are wider than a floor measured against one standing graph.

Units are queries per second. bfb reports rps, which counts batch requests: at --search-batch-size 16 the two differ by 16x, and reading one as the other once turned a 2.2x speedup into an apparent 7x regression. The tables show the request rate wherever it diverges.

The open-loop rows are not a speed. W4-sat50/70/90 offer a fixed fraction of measured saturation; serving it means the engine kept up, not that it was faster, so no ratio is printed. They exist because --parallel is a closed loop, where a stalled server stops receiving requests and understates its own tail.

ef is the same unit here: Qdrant held 2 segments for this run, one populated and 1 empty (read back per segment from the engine), and strawmANN held one graph, so equal ef is equal traversal width and the curves may be read at equal x. That is a property of this run, not of the engines.

What was measured builds, dataset, host

	strawmANN `sm-dbp1m-perf-1008`	Qdrant `qd-dbp1m-perf-1008`
engine	commit `7be7ce5844f5` `ReleaseFast`, native build, vnni on binary sha256 `a785c2b080c5d068` built 2026-10-07T15:45:29Z with zig 0.16.0	version 1.19.2-dev binary `~/.cache/strawmann/qdrant-dbeb0f73/qdrant` sha256 `dbeb0f73dea2d371` build `878843e6` (from the server's banner, not a checkout) native binary outside a checkout: sha256 is the identity (§8.9) network `native` no container in the path, like strawmann
measured	2026-10-07T19:45:36Z to 2026-10-08T03:57:10Z (3 passes)	2026-10-07T21:07:30Z to 2026-10-08T06:01:39Z (3 passes)
profile	`as-deployed` · both engines
§7.1 gate	pass · both engines
environment hash	`373e8531fd3d1bf6` · both engines
load generator	bfb dev @ caaeeef (qdrant/bfb#185, BM25 text requests; carries #183 and #172, findings 32's --rps reaping fix) · both engines
client	qdrant-client 1.16.1-dev (git dev branch) · both engines
launched as	`~/Workspace/strawmann/zig-out/bin/strawmann --port 6334 --capacity 1237500 --connections 64 --workers 7 --io-threads 1 --pin --cpus 4-11 --data-dir ~/.cache/strawmann/strawmann-storage --default-placement cached`	`~/.cache/strawmann/qdrant-dbeb0f73/qdrant`
cores the engine could use	`4-11` (8 cores) requested by the harness `4-11` (8 cores) observed on 8 threads `0-23` (24 cores) observed on 1 thread	`4-11` (8 cores) requested by the harness `4-11` (8 cores) observed on 42 threads

Dataset

dbpedia-openai-1m, 1000000 × 1536, cosine, 10000 held-out queries

ground truth: recomputed by our fp64 oracle

26 files, each pinned by sha256 in datasets.json

Host

AMD Ryzen AI 9 HX PRO 370 w/ Radeon 890M

24 logical cores · 58.6 GB · kernel 7.0.0-38-generic

memory bandwidth: 76.2 GB/s aggregate (24 threads) · 43.1 GB/s single core (57% of bus)

strawmann startup banner, 2026-10-07T19:45:36Z; the Qdrant run agrees within 10%

All throughput rows every measurement, with its notes

workload		strawmANN	Qdrant	ratio	notes
W0	d=4 floor: graph traversal with the distance taken out	4,564	3,615	1.26x	1.26x is 1.74x less work per query x 0.67x cores busy during the row (0.67 against 1.01) x 1.00x clock x 1.09x counted on-CPU share: the middle term is occupancy, not search speed d=4 floor: this measures graph traversal and heaps with the distance taken out, not request plumbing (findings 28: 76% of the server's CPU was HNSW search, 24% kernel; the RPC path is a small remainder)
W3	search, fp32, single query	1,059	627	1.69x
W4	search, saturating (closed loop)	3,572	3,384	parity	within the ±5.8% band this dataset's noise floor puts on W4: no measured difference, not a small one
W4-sat50	search, fixed rate at 50% of saturation (open loop)	1,732	1,732	offered
W4-sat70	search, fixed rate at 70% of saturation (open loop)	2,424	2,424	offered
W4-sat90	search, fixed rate at 90% of saturation (open loop)	3,117	3,117	offered
W5	search batched (16 distinct dataset queries per request)	3,542250 requests/s	3,441237 requests/s	parity	per-batch latency (16 queries/request); queries: dataset, random-sample within the ±5.9% band this dataset's noise floor puts on W5: no measured difference, not a small one batched 16, so qps and rps differ by 16x; the ratio uses qps
W6	quantized: scalar	8,112	7,160	1.13x
W6-ef32	SQ8 recall control, ef=32 (latency only)	24,808	17,155	1.45x
W6-ef64	SQ8 recall control, ef=64 (latency only)	14,496	11,592	1.25x
W6-ef128	SQ8 recall control, ef=128 (latency only)	8,120	7,209	1.13x
W6-ef256	SQ8 recall control, ef=256 (latency only)	4,485	4,251	1.06x
W6-ef512	SQ8 recall control, ef=512 (latency only)	2,457	2,411	parity	within the ±3.1% band this dataset's noise floor puts on W6-ef512: no measured difference, not a small one
W7	quantized: binary + oversampling	23,264	13,558	1.72x
W7-2bit	quantized: binary quantization, 2 bits + oversampling	20,459	12,314	1.66x
W7-1p5bit	quantized: binary quantization, 1.5 bits + oversampling	21,077	12,685	1.66x
W8	quantized: PQ	6,115	5,067	1.21x
W14-1bit	quantized: TurboQuant, 1 bit + oversampling	14,063	12,605	1.12x
W14-1p5bit	quantized: TurboQuant, 1.5 bits + oversampling	11,493	9,944	1.16x
W14-2bit	quantized: TurboQuant, 2 bits + oversampling	15,832	11,533	1.37x
W14-4bit	quantized: TurboQuant, 4 bits + oversampling	12,937	10,006	1.29x
W9	exact / brute force	13	10	1.27x	1.27x is 2.53x less work per query x 0.50x cores busy during the row (4.00 against 7.98) x 1.00x clock: the middle term is occupancy, not search speed brute force over the whole collection, no index involved
W10-ef32	recall control, ef=32 (latency only)	10,076	7,343	1.37x
W10-ef64	recall control, ef=64 (latency only)	6,082	4,790	1.27x
W10-ef128	recall control, ef=128 (latency only)	3,581	3,062	1.17x
W10-ef256	recall control, ef=256 (latency only)	2,058	1,811	1.14x
W10-ef512	recall control, ef=512 (latency only)	1,170	1,051	parity	within the ±11.6% band this dataset's noise floor puts on W10-ef512: no measured difference, not a small one
W12-sel1	filtered search, one keyword (~1% of bench12)	3,958	1,395	2.84x	filtered search over a keyword payload index both engines were asked to build; whether each did is read back from the engine and refused beside the row when it did not. Recall is measured under the filter, against ground truth restricted to the matching set
W12-sel10	filtered search, any of 10 keywords (~10% of bench12)	2,536	835	3.04x	filtered search over a keyword payload index both engines were asked to build; whether each did is read back from the engine and refused beside the row when it did not. Recall is measured under the filter, against ground truth restricted to the matching set
W12-sel1-ef32	filtered recall control, one keyword, ef=32 (latency only)	3,959	3,802	-	recall unequal: 1.0000 vs 0.9829; §7.4 compares at equal recall
W12-sel1-ef64	filtered recall control, one keyword, ef=64 (latency only)	3,958	2,329	1.70x
W12-sel1-ef128	filtered recall control, one keyword, ef=128 (latency only)	3,941	1,398	2.82x
W12-sel1-ef256	filtered recall control, one keyword, ef=256 (latency only)	3,966	857	4.63x
W12-sel1-ef512	filtered recall control, one keyword, ef=512 (latency only)	3,956	552	7.16x
W12-sel10-ef32	filtered recall control, any of 10 keywords, ef=32 (latency only)	7,481	2,160	-	recall unequal: 0.9378 vs 0.9699; §7.4 compares at equal recall
W12-sel10-ef64	filtered recall control, any of 10 keywords, ef=64 (latency only)	4,433	1,354	-	recall unequal: 0.9728 vs 0.9901; §7.4 compares at equal recall
W12-sel10-ef128	filtered recall control, any of 10 keywords, ef=128 (latency only)	2,517	832	3.03x
W12-sel10-ef256	filtered recall control, any of 10 keywords, ef=256 (latency only)	1,468	522	2.81x
W12-sel10-ef512	filtered recall control, any of 10 keywords, ef=512 (latency only)	443	338	1.31x
W13	scroll / pagination	50,411	46,314	1.09x	strawmANN: -n 200000 (4x the table's 50000) Qdrant: -n 100000 (2x the table's 50000) strawmANN: 0.13 ms of the client's 0.14 ms p50 is not server time (13x), so 92% of this row is the load generator and the socket Qdrant: 0.13 ms of the client's 0.16 ms p50 is not server time (6x), so 82% of this row is the load generator and the socket 1.09x is 3.26x less work per query x 0.33x cores busy during the row (1.38 against 4.21) x 1.00x clock x 1.02x counted on-CPU share: the middle term is occupancy, not search speed
W11-steady	mixed read/write below the rebuild threshold: search bench2 while 49,500 synthetic points append	541	494	-	strawmANN: qps over the 26.1 s of the append the search saw (505 over the whole search); append 1,900 points/s Qdrant: qps over the 26.1 s of the append the search saw (427 over the whole search); append 1,900 points/s strawmANN pass 1 +15% against two passes that agree: one pass set the spread on this row, so a band built from it is not a noise band search-during-write; no recall join
W11	mixed read/write: search bench2 while 198,000 synthetic points append (runs last)	121	72	-	strawmANN: qps over the 60.0 s of the append the search saw (106 over the whole search); append 3,300 points/s Qdrant: qps over the 186.7 s of the append the search saw (68 over the whole search); append 1,322 points/s strawmANN -11% across its passes, monotonically: the spread on this row is drift rather than noise, so the median is a trend's midpoint and not a repeatable measurement search-during-write; no recall join

W10: throughput against ef

All latency rows closed loop included, p50 to max

A row flagged “… not server time” is one where the client saw far more than the server reported: below the rate at which a queue can form, what is left is the load generator's own scheduling.

workload		strawmANN p50	strawmANN p95	strawmANN p99	strawmANN p99.9	strawmANN max	Qdrant p50	Qdrant p95	Qdrant p99	Qdrant p99.9	Qdrant max
W0	d=4 floor: graph traversal with the distance taken out closed loop	215 µs	242 µs	262 µs	307 µs	1.36 ms	277 µs	319 µs	348 µs	411 µs	1.68 ms
W3	search, fp32, single query closed loop	948 µs	1.21 ms	1.38 ms	1.90 ms	2.61 ms	1.59 ms	2.13 ms	2.27 ms	2.51 ms	4.63 ms
W4	search, saturating (closed loop) closed loop	17.94 ms	19.53 ms	20.22 ms	24.43 ms	37.60 ms	18.77 ms	24.35 ms	33.05 ms	44.66 ms	75.65 ms
W4-sat50	search, fixed rate at 50% of saturation (open loop) open loop	1.02 ms	1.37 ms	1.62 ms	1.84 ms	6.31 ms	1.74 ms	2.38 ms	2.84 ms	3.71 ms	8.84 ms
W4-sat70	search, fixed rate at 70% of saturation (open loop) open loop	1.11 ms	1.55 ms	1.93 ms	2.28 ms	6.39 ms	2.00 ms	3.11 ms	3.79 ms	4.76 ms	13.06 ms
W4-sat90	search, fixed rate at 90% of saturation (open loop) open loop	1.38 ms	2.34 ms	2.76 ms	3.44 ms	7.24 ms	3.13 ms	6.40 ms	14.05 ms	23.35 ms	40.71 ms
W5	search batched (16 distinct dataset queries per request) closed loop	288.12 ms	292.48 ms	294.42 ms	303.98 ms	318.36 ms	295.78 ms	312.32 ms	339.82 ms	402.78 ms	446.41 ms
W6	quantized: scalar closed loop	7.90 ms	8.48 ms	8.72 ms	26.73 ms	32.37 ms	8.80 ms	11.14 ms	13.85 ms	24.76 ms	33.84 ms
W6-ef32	SQ8 recall control, ef=32 (latency only) closed loop	2.57 ms	2.77 ms	2.87 ms	22.28 ms	26.80 ms	3.54 ms	5.81 ms	7.42 ms	24.74 ms	31.29 ms
W6-ef64	SQ8 recall control, ef=64 (latency only) closed loop	4.42 ms	4.75 ms	4.89 ms	19.24 ms	23.85 ms	5.32 ms	7.98 ms	9.83 ms	23.49 ms	30.74 ms
W6-ef128	SQ8 recall control, ef=128 (latency only) closed loop	7.90 ms	8.48 ms	8.71 ms	24.79 ms	30.67 ms	8.76 ms	11.15 ms	13.74 ms	23.13 ms	30.92 ms
W6-ef256	SQ8 recall control, ef=256 (latency only) closed loop	14.29 ms	15.39 ms	15.84 ms	23.82 ms	34.13 ms	15.00 ms	16.72 ms	17.98 ms	29.11 ms	40.12 ms
W6-ef512	SQ8 recall control, ef=512 (latency only) closed loop	26.07 ms	28.11 ms	28.96 ms	29.95 ms	42.91 ms	26.50 ms	29.00 ms	30.34 ms	41.10 ms	51.20 ms
W7	quantized: binary + oversampling closed loop	2.72 ms	2.86 ms	2.94 ms	18.45 ms	23.64 ms	4.16 ms	9.36 ms	13.00 ms	20.16 ms	30.03 ms
W7-2bit	quantized: binary quantization, 2 bits + oversampling closed loop	3.11 ms	3.26 ms	3.35 ms	14.11 ms	17.04 ms	4.61 ms	10.14 ms	13.92 ms	26.40 ms	43.02 ms
W7-1p5bit	quantized: binary quantization, 1.5 bits + oversampling closed loop	3.00 ms	3.15 ms	3.24 ms	25.79 ms	29.79 ms	4.48 ms	9.82 ms	13.51 ms	23.95 ms	33.60 ms
W8	quantized: PQ closed loop	10.44 ms	11.01 ms	11.27 ms	26.49 ms	34.46 ms	12.55 ms	13.87 ms	15.65 ms	24.60 ms	35.31 ms
W14-1bit	quantized: TurboQuant, 1 bit + oversampling closed loop	4.53 ms	4.80 ms	4.92 ms	15.31 ms	20.61 ms	4.48 ms	10.06 ms	14.03 ms	25.03 ms	35.70 ms
W14-1p5bit	quantized: TurboQuant, 1.5 bits + oversampling closed loop	5.55 ms	5.87 ms	6.03 ms	19.45 ms	24.95 ms	5.88 ms	11.74 ms	16.09 ms	24.04 ms	36.46 ms
W14-2bit	quantized: TurboQuant, 2 bits + oversampling closed loop	4.02 ms	4.23 ms	4.34 ms	19.09 ms	24.08 ms	4.93 ms	10.70 ms	14.78 ms	29.95 ms	39.38 ms
W14-4bit	quantized: TurboQuant, 4 bits + oversampling closed loop	4.92 ms	5.23 ms	5.37 ms	27.84 ms	33.86 ms	5.79 ms	11.84 ms	16.51 ms	26.39 ms	43.46 ms
W9	exact / brute force closed loop	620.80 ms	798.63 ms	825.19 ms	836.42 ms	836.70 ms	810.76 ms	836.96 ms	848.33 ms	878.97 ms	926.20 ms
W10-ef32	recall control, ef=32 (latency only) closed loop	783 µs	1.07 ms	1.26 ms	1.57 ms	5.85 ms	1.01 ms	1.70 ms	2.00 ms	2.53 ms	5.65 ms
W10-ef64	recall control, ef=64 (latency only) closed loop	1.31 ms	1.78 ms	2.03 ms	2.48 ms	8.12 ms	1.56 ms	2.59 ms	3.09 ms	3.86 ms	8.20 ms
W10-ef128	recall control, ef=128 (latency only) closed loop	2.23 ms	3.09 ms	3.47 ms	3.89 ms	9.45 ms	2.54 ms	3.93 ms	4.62 ms	5.73 ms	9.55 ms
W10-ef256	recall control, ef=256 (latency only) closed loop	3.88 ms	5.46 ms	6.13 ms	6.98 ms	12.46 ms	4.32 ms	6.49 ms	7.57 ms	9.01 ms	13.73 ms
W10-ef512	recall control, ef=512 (latency only) closed loop	6.79 ms	9.74 ms	10.95 ms	12.34 ms	16.47 ms	7.53 ms	10.92 ms	12.35 ms	14.29 ms	18.26 ms
W12-sel1	filtered search, one keyword (~1% of bench12) closed loop	16.08 ms	17.63 ms	17.90 ms	26.35 ms	35.83 ms	45.57 ms	48.09 ms	50.14 ms	58.15 ms	88.57 ms
W12-sel10	filtered search, any of 10 keywords (~10% of bench12) closed loop	25.30 ms	26.86 ms	27.51 ms	32.76 ms	50.35 ms	76.84 ms	81.90 ms	84.25 ms	98.42 ms	143.19 ms
W12-sel1-ef32	filtered recall control, one keyword, ef=32 (latency only) closed loop	16.09 ms	17.60 ms	17.88 ms	30.97 ms	43.16 ms	16.52 ms	21.14 ms	30.08 ms	44.02 ms	62.90 ms
W12-sel1-ef64	filtered recall control, one keyword, ef=64 (latency only) closed loop	16.10 ms	17.59 ms	17.85 ms	23.13 ms	35.14 ms	27.30 ms	29.06 ms	30.88 ms	38.66 ms	56.92 ms
W12-sel1-ef128	filtered recall control, one keyword, ef=128 (latency only) closed loop	16.10 ms	17.58 ms	17.86 ms	52.45 ms	60.30 ms	45.80 ms	48.01 ms	49.57 ms	56.59 ms	83.87 ms
W12-sel1-ef256	filtered recall control, one keyword, ef=256 (latency only) closed loop	16.08 ms	17.65 ms	17.93 ms	37.02 ms	49.38 ms	74.30 ms	78.04 ms	80.77 ms	88.55 ms	139.58 ms
W12-sel1-ef512	filtered recall control, one keyword, ef=512 (latency only) closed loop	16.13 ms	17.59 ms	17.84 ms	30.91 ms	43.62 ms	115.60 ms	120.15 ms	124.34 ms	134.45 ms	207.08 ms
W12-sel10-ef32	filtered recall control, any of 10 keywords, ef=32 (latency only) closed loop	8.51 ms	9.11 ms	9.45 ms	20.99 ms	27.49 ms	29.69 ms	31.92 ms	32.99 ms	43.04 ms	60.45 ms
W12-sel10-ef64	filtered recall control, any of 10 keywords, ef=64 (latency only) closed loop	14.46 ms	15.39 ms	15.77 ms	29.24 ms	39.57 ms	47.13 ms	50.72 ms	52.55 ms	64.03 ms	84.94 ms
W12-sel10-ef128	filtered recall control, any of 10 keywords, ef=128 (latency only) closed loop	25.52 ms	27.09 ms	27.65 ms	28.54 ms	43.96 ms	76.99 ms	81.98 ms	84.22 ms	92.99 ms	116.41 ms
W12-sel10-ef256	filtered recall control, any of 10 keywords, ef=256 (latency only) closed loop	43.72 ms	46.10 ms	47.08 ms	48.66 ms	61.75 ms	122.27 ms	130.02 ms	133.88 ms	143.82 ms	219.01 ms
W12-sel10-ef512	filtered recall control, any of 10 keywords, ef=512 (latency only) closed loop	147.64 ms	157.26 ms	160.17 ms	163.18 ms	169.77 ms	189.02 ms	199.45 ms	205.24 ms	225.01 ms	344.10 ms
W13	scroll / pagination closed loop	139 µs	207 µs	233 µs	271 µs	1.04 ms	159 µs	222 µs	292 µs	446 µs	2.17 ms
W11-steady	mixed read/write below the rebuild threshold: search bench2 while 49,500 synthetic points append closed loop	13.96 ms	30.03 ms	40.01 ms	153.28 ms	209.70 ms	12.14 ms	53.99 ms	68.87 ms	87.99 ms	171.05 ms
W11	mixed read/write: search bench2 while 198,000 synthetic points append (runs last) closed loop	58.76 ms	168.97 ms	196.61 ms	225.00 ms	252.88 ms	115.38 ms	196.24 ms	230.21 ms	371.97 ms	836.47 ms

Storage and I/O block layer and syscalls, per row

The syscall rows count every descriptor, sockets included, so on a search row they measure the network rather than the disk.

	strawmANN	Qdrant
storage on disk	35.5 GiBbefore W11-steady, W11allocated at --capacity	26.5 GiBbefore W11-steady, W11
peak RSS	36.2 GiB	43.6 GiB
of which anonymous	4.0 GiB	3.4 GiB
of which file-backed	29.8 GiB	34.9 GiB
disk read bytes	74.1 MiB	2.6 GiB
disk write bytes	64.9 GiB	352.8 GiB
disk read ops	56,907	69,420
disk write ops	1,767,581	6,074,116
syscall reads (all fds)	12,643,194	57,660
syscall writes (all fds)	5,276,430	39,372,080

measured via strawmANN: proc+cgroup / Qdrant: proc+cgroup. proc supplies syscall counts and block-layer bytes, cgroup supplies block-layer operations and bytes, so a row one interface does not carry reads n/a via …. unknown means it was not measured, and 0 means it was: an engine started with no --data-dir has no store, which is the row this section exists for. storage on disk is the level before the rows with a concurrent writer: during those an engine that rewrites segments is caught mid-rewrite, and the same row has read 3.47 and 10.11 GiB on two runs of one binary. peak RSS does include them, being a peak.

Per workload

A search row doing block-layer reads is an engine going to disk to answer a query.

workload	strawmANN read	strawmANN written	strawmANN read ops	strawmANN write ops	Qdrant read	Qdrant written	Qdrant read ops	Qdrant write ops
W0-upload	4.0 KiB	60.5 MiB	3	977	114.2 MiB	383.8 MiB	3,057	8,306
W0	0	0	0	3	0	0	0	6
W1	180.0 KiB	5.7 GiB	135	10,389	4.1 MiB	15.7 GiB	300	261,940
W2	0	5.7 GiB	0	93,561	172.7 MiB	25.5 GiB	4,524	436,561
W3	0	0	0	3	0	0	0	0
W4	0	0	0	0	0	0	0	0
W4-sat50	0	0	0	0	0	0	0	0
W4-sat70	0	0	0	0	0	0	0	0
W4-sat90	0	0	0	0	0	0	0	0
W5	0	0	0	0	0	0	0	0
W6-upload	0	5.7 GiB	0	93,552	225.8 MiB	34.4 GiB	6,180	581,960
W6	0	0	0	0	0	0	0	0
W6-ef32	0	0	0	0	0	0	0	0
W6-ef64	0	0	0	0	0	0	0	0
W6-ef128	0	0	0	0	0	0	0	0
W6-ef256	0	0	0	0	0	0	0	15
W6-ef512	0	0	0	0	0	0	0	54
W7-upload	12.0 KiB	5.7 GiB	9	93,820	317.9 MiB	37.2 GiB	8,214	640,505
W7	0	0	0	0	0	0	0	0
W7-2bit-upload	380.0 KiB	5.7 GiB	285	93,550	218.6 MiB	30.3 GiB	5,670	521,555
W7-2bit	0	0	0	0	0	0	0	0
W7-1p5bit-upload	344.0 KiB	5.7 GiB	258	93,550	262.7 MiB	32.9 GiB	6,759	568,461
W7-1p5bit	0	0	0	0	0	0	0	0
W8-upload	92.0 KiB	5.7 GiB	69	94,031	175.9 MiB	25.3 GiB	4,500	433,868
W8	0	0	0	0	0	0	0	0
W14-1bit-upload	40.0 KiB	5.7 GiB	30	93,548	233.0 MiB	30.2 GiB	5,982	518,948
W14-1bit	0	0	0	0	0	0	0	0
W14-1p5bit-upload	20.0 KiB	5.7 GiB	15	93,549	235.1 MiB	30.9 GiB	5,991	538,858
W14-1p5bit	0	0	0	0	0	0	0	0
W14-2bit-upload	332.0 KiB	5.7 GiB	249	93,568	213.0 MiB	30.7 GiB	5,394	535,263
W14-2bit	0	0	0	0	0	0	0	0
W14-4bit-upload	24.0 KiB	5.7 GiB	18	93,554	216.3 MiB	32.1 GiB	5,460	552,407
W14-4bit	0	0	0	0	0	0	0	0
W9	0	0	0	0	0	0	0	15
W10-ef32	0	0	0	0	0	0	0	0
W10-ef64	0	0	0	0	0	0	0	0
W10-ef128	0	0	0	0	0	0	0	0
W10-ef256	0	0	0	0	0	0	0	0
W10-ef512	0	0	0	0	0	0	0	0
W12-upload	768.0 KiB	1.1 GiB	576	796,306	53.5 MiB	5.3 GiB	1,494	92,031
W12-sel1	0	0	0	3	0	0	0	438
W12-sel10	12.0 KiB	0	9	45	0	0	0	69
W12-sel1-ef32	0	0	0	33	0	0	0	0
W12-sel1-ef64	0	0	0	0	0	0	0	48
W12-sel1-ef128	88.0 KiB	0	66	0	0	0	0	81
W12-sel1-ef256	0	0	0	0	0	0	0	66
W12-sel1-ef512	0	0	0	0	0	0	0	0
W12-sel10-ef32	0	0	0	0	0	0	0	0
W12-sel10-ef64	0	0	0	24	0	0	0	0
W12-sel10-ef128	0	0	0	45	0	0	0	0
W12-sel10-ef256	0	0	0	54	0	0	0	0
W12-sel10-ef512	0	0	0	0	0	0	0	0
W13	0	0	0	0	0	0	0	0
W11-steady	64.4 MiB	290.1 MiB	49,452	4,689	73.6 MiB	12.2 GiB	1,809	210,007
W11	7.5 MiB	1.1 GiB	5,733	18,727	167.0 MiB	9.8 GiB	4,086	172,654

Scheduler and memory CPU use, run-queue wait, migrations

Whether the engine was running while it ran. of wall is CPU over elapsed, waiting is runnable-but-not-scheduled, migrations checks the pinning claim, and switches gives voluntary over involuntary — the scheduler taking the core away against the engine choosing to sleep, which per unit of work is the cheapest signal of lock contention there is. A dash is not a zero: an index build's threads exit before the row does, and threads shows what fraction of the row the survivors account for.

Time spent waiting for a core

Lower is better; the bar between a pair is the gap.

Peak memory per row

Peak RSS while the row ran, from the engine's own process. Unlike the disk counters this is not refused across the two engines: residency decides where bytes live, and this is what the process held either way.

workload	strawmANN							Qdrant
workload	cpu	of wall	waiting	switches vol/invol	migrations	faults min/maj	threads	cpu	of wall	waiting	switches vol/invol	migrations	faults min/maj	threads
W0-upload	111.0 s	541%	43.2 ms	51,993 / 577	2	113,520 / 0	9 1%	158.5 s	495%	15,024.7 ms	155,135 / 34,839	12,858	424,958 / 193,878	42 7%
W0	7.7 s	70%	3.0 ms	164,728 / 39	0	0 / 0	9	14.2 s	102%	3.7 ms	561,018 / 18	197	464 / 0	36 0%
W1	11.5 s	137%	107.4 ms	587,441 / 24	0	1,576,572 / 0	9	260.1 s	711%	108,197.3 ms	158,131 / 88,923	31,954	630,876 / 1,597,619	53 68%
W2	2,022.1 s	771%	1,260.2 ms	596,502 / 9,180	0	1,640,235 / 0	9 1%	2,115.7 s	596%	108,969.3 ms	197,060 / 102,055	32,701	1,414,315 / 1,626,490	38 3%
W3	43.6 s	92%	4.9 ms	192,393 / 55	0	0 / 0	9	79.5 s	100%	5.9 ms	562,653 / 242	228	1,385 / 0	38 0%
W4	100.3 s	715%	7.0 ms	89,883 / 97	0	119 / 0	9	117.4 s	792%	96,311.0 ms	42,815 / 43,120	5,047	30,332 / 0	44 0%
W4-sat50	190.5 s	165%	17.1 ms	546,259 / 124	0	3 / 0	9	351.1 s	304%	11,591.5 ms	1,936,534 / 9,832	469,270	4,655 / 0	52 0%
W4-sat70	206.3 s	250%	21.2 ms	481,356 / 122	0	12 / 0	9	381.6 s	462%	74,808.3 ms	1,660,735 / 50,690	648,154	3,791 / 0	68
W4-sat90	270.9 s	421%	13.1 ms	444,623 / 156	0	9 / 0	9	444.9 s	692%	217,282.7 ms	743,500 / 186,940	286,321	25,140 / 0	44 0%
W5	99.7 s	706%	9.7 ms	6,586 / 204	0	752 / 0	9	115.8 s	796%	86,320.1 ms	10,060 / 16,525	1,552	22,143 / 0	44 0%
W6-upload	1,106.8 s	714%	467.4 ms	94,346 / 4,940	0	2,205,991 / 0	9 1%	1,048.0 s	434%	65,433.9 ms	222,909 / 63,211	30,253	2,981,965 / 1,686,609	41 0%
W6	44.9 s	723%	6.8 ms	80,702 / 77	0	18 / 0	9	55.0 s	784%	15,840.5 ms	36,076 / 25,029	2,607	2,779 / 0	46 65%
W6-ef32	15.5 s	753%	1.6 ms	34,950 / 17	0	6 / 0	9	22.2 s	750%	12,726.4 ms	27,488 / 10,720	5,947	438 / 0	46
W6-ef64	25.8 s	737%	2.0 ms	60,525 / 30	0	0 / 0	9	33.7 s	772%	14,418.4 ms	27,201 / 14,284	4,624	205 / 0	46
W6-ef128	44.9 s	725%	2.7 ms	80,921 / 49	0	0 / 0	9	55.1 s	790%	15,464.4 ms	31,248 / 23,776	2,507	721 / 0	46
W6-ef256	80.0 s	715%	7.6 ms	96,652 / 110	0	50 / 0	9	93.9 s	796%	18,336.9 ms	33,558 / 29,702	482	1,416 / 0	46
W6-ef512	144.6 s	709%	8.1 ms	108,115 / 184	0	5 / 0	9	165.7 s	797%	27,282.8 ms	41,911 / 39,644	424	2,482 / 0	46
W7-upload	288.9 s	487%	148.7 ms	96,206 / 1,445	1	1,698,301 / 0	9 3%	608.6 s	165%	39,248.8 ms	306,147 / 35,273	27,924	2,997,835 / 1,861,540	41 0%
W7	16.2 s	741%	1.5 ms	55,285 / 20	0	45 / 0	9	28.6 s	767%	93,182.8 ms	77,152 / 53,631	19,523	4,456 / 0	81
W7-2bit-upload	312.3 s	406%	170.2 ms	92,859 / 1,545	3	1,756,298 / 11	9 2%	636.3 s	363%	44,457.0 ms	234,126 / 40,379	29,749	2,162,663 / 1,824,782	47 7%
W7-2bit	18.4 s	740%	2.3 ms	59,034 / 25	0	0 / 0	9	31.4 s	766%	93,687.6 ms	74,566 / 54,149	19,425	3,632 / 0	87
W7-1p5bit-upload	298.1 s	397%	143.3 ms	95,767 / 1,542	6	1,736,859 / 2	9 3%	627.1 s	245%	46,578.9 ms	252,830 / 41,923	27,665	2,437,715 / 1,816,141	46 20%
W7-1p5bit	17.8 s	736%	2.0 ms	58,181 / 23	0	0 / 0	9	30.6 s	768%	93,526.9 ms	74,150 / 55,118	19,700	2,705 / 0	86
W8-upload	2,968.7 s	769%	1,353.4 ms	105,755 / 13,536	0	1,859,672 / 0	9 0%	5,981.8 s	652%	27,983.8 ms	657,619 / 156,783	100,504	1,707,263 / 1,611,187	43 0%
W8	58.8 s	713%	3.4 ms	108,957 / 74	0	0 / 0	9	78.7 s	794%	15,653.6 ms	35,300 / 28,647	650	3,577 / 0	49
W14-1bit-upload	2,059.2 s	766%	1,109.5 ms	93,003 / 8,926	1	1,708,495 / 1	9 0%	650.8 s	368%	54,863.8 ms	255,078 / 78,545	35,230	2,151,305 / 1,777,602	47 12%
W14-1bit	26.2 s	730%	2.0 ms	82,975 / 35	0	0 / 0	9	30.8 s	769%	94,570.4 ms	71,672 / 51,017	18,635	5,125 / 0	85
W14-1p5bit-upload	2,082.3 s	766%	1,126.5 ms	94,130 / 9,238	1	1,739,075 / 0	9 0%	849.4 s	354%	53,287.1 ms	258,560 / 87,859	29,537	2,294,444 / 1,803,171	46 18%
W14-1p5bit	31.9 s	726%	2.9 ms	90,188 / 41	0	0 / 0	9	39.1 s	771%	97,130.4 ms	61,299 / 45,926	17,066	6,949 / 0	84
W14-2bit-upload	2,059.1 s	764%	1,112.1 ms	96,300 / 9,161	0	1,772,660 / 0	9 0%	989.6 s	390%	72,137.4 ms	241,792 / 115,524	33,208	2,223,495 / 1,827,480	47 10%
W14-2bit	23.5 s	734%	2.5 ms	75,301 / 36	0	0 / 0	9	33.6 s	767%	94,664.6 ms	68,675 / 51,312	19,444	4,576 / 0	85 13%
W14-4bit-upload	2,071.0 s	762%	1,135.9 ms	95,546 / 9,158	1	1,900,954 / 2	9 0%	1,050.4 s	327%	62,528.4 ms	306,032 / 211,676	33,566	2,695,357 / 1,919,687	47 18%
W14-4bit	28.4 s	728%	2.2 ms	81,154 / 37	0	0 / 0	9	38.9 s	772%	97,214.2 ms	62,007 / 47,275	16,353	4,632 / 0	85
W9	636.2 s	400%	71.2 ms	10,366 / 1,824	0	0 / 0	9	1,612.5 s	798%	14,216.7 ms	26,586 / 22,802	6,641	1,513 / 0	49
W10-ef32	34.7 s	694%	4.0 ms	97,500 / 50	0	0 / 0	9	40.3 s	589%	14,951.6 ms	247,892 / 24,101	84,216	1,035 / 0	49
W10-ef64	58.1 s	704%	5.9 ms	103,016 / 75	0	0 / 0	9	66.3 s	633%	41,212.9 ms	307,642 / 46,407	157,577	535 / 0	66 0%
W10-ef128	98.9 s	706%	9.6 ms	108,882 / 140	0	0 / 0	9	108.4 s	662%	57,485.6 ms	338,258 / 55,003	177,203	472 / 0	66
W10-ef256	171.9 s	706%	17.3 ms	113,368 / 260	0	0 / 0	9	190.4 s	689%	83,088.3 ms	370,980 / 83,785	196,685	904 / 0	66
W10-ef512	301.4 s	704%	24.7 ms	116,256 / 508	0	0 / 0	9	338.6 s	711%	92,337.9 ms	340,069 / 97,355	146,149	2,196 / 0	49 0%
W12-upload	299.7 s	614%	181.1 ms	32,443 / 1,498	0	453,100 / 41	9 2%	473.9 s	571%	17,205.5 ms	52,358 / 17,842	6,853	346,436 / 446,599	51 0%
W12-sel1	90.0 s	710%	6.6 ms	36,245 / 144	0	48 / 0	9	286.5 s	799%	39,910.4 ms	57,181 / 57,674	471	4,266 / 0	58
W12-sel10	139.8 s	708%	8.3 ms	122,727 / 185	0	6 / 2	9	478.6 s	798%	57,403.3 ms	76,468 / 75,612	1,105	6,255 / 0	58
W12-sel1-ef32	89.8 s	709%	7.2 ms	33,099 / 119	0	0 / 0	9	104.7 s	793%	68,961.4 ms	36,826 / 37,131	2,672	19,528 / 0	58 86%
W12-sel1-ef64	90.1 s	710%	6.1 ms	46,408 / 124	0	0 / 0	9	171.6 s	798%	27,271.3 ms	42,807 / 40,870	303	2,525 / 0	58
W12-sel1-ef128	90.0 s	707%	6.1 ms	41,290 / 144	0	13 / 10	9	285.9 s	798%	39,808.4 ms	57,043 / 55,693	550	3,145 / 0	58
W12-sel1-ef256	89.9 s	709%	6.2 ms	41,961 / 138	0	54 / 0	9	466.5 s	799%	54,864.6 ms	75,336 / 74,196	1,177	6,221 / 0	58
W12-sel1-ef512	90.2 s	711%	5.4 ms	43,063 / 114	0	0 / 0	9	723.8 s	799%	68,963.1 ms	92,603 / 92,525	2,374	8,405 / 0	58
W12-sel10-ef32	48.2 s	717%	2.3 ms	100,250 / 49	0	0 / 0	9	185.2 s	799%	29,128.7 ms	44,729 / 42,773	303	3,330 / 0	58
W12-sel10-ef64	80.7 s	713%	3.7 ms	113,741 / 89	0	0 / 0	9	294.3 s	796%	41,081.1 ms	58,550 / 56,951	545	4,735 / 0	58
W12-sel10-ef128	141.0 s	708%	5.0 ms	124,440 / 135	0	0 / 0	9	479.6 s	797%	57,549.0 ms	75,509 / 75,109	1,042	5,123 / 0	58
W12-sel10-ef256	240.2 s	704%	11.3 ms	132,489 / 286	0	3 / 0	9	765.9 s	800%	74,054.5 ms	97,165 / 97,026	2,862	10,499 / 0	58
W12-sel10-ef512	796.9 s	706%	70.1 ms	68,835 / 1,377	0	0 / 0	9	1,183.0 s	799%	87,860.9 ms	115,762 / 116,908	5,854	15,545 / 0	58
W13	5.6 s	137%	4.6 ms	245,377 / 6	0	0 / 0	9	9.2 s	411%	812.8 ms	383,390 / 4,583	113,205	677 / 0	73
W11-steady	256.2 s	784%	195,850.3 ms	52,523 / 87,295	779	485,661 / 15,720	18 44%	252.7 s	773%	212,991.1 ms	117,319 / 94,771	44,937	652,509 / 127,904	65 90%
W11	603.3 s	802%	506,643.3 ms	24,151 / 125,288	1,011	769,326 / 1,133	19 71%	1,488.1 s	794%	1,211,680.4 ms	200,009 / 358,593	37,046	1,053,039 / 352,436	59 68%

Waiting and stalls where the time went when the engine was not running

waiting is runnable and not scheduled, blocked on disk is not runnable at all, and pressure is PSI for the engine's cgroup (some = at least one task stalled, full = every runnable task).

Time stalled on I/O

Time every task in the engine's cgroup was stalled on I/O. Lower is better. 38 rows never stalled on I/O and are not drawn.

Time stalled on memory

Time every task in the engine's cgroup was stalled on memory. Lower is better. 52 rows never stalled on memory and are not drawn.

workload	strawmANN					Qdrant
workload	waiting	blocked on disk	cpu some/full	io some/full	memory some/full	waiting	blocked on disk	cpu some/full	io some/full	memory some/full
W0-upload	43.2 ms	0 ms	14 / 9 ms	1 / 1 ms	0 / 0 ms	15,024.7 ms	0 ms	1,372 / 16 ms	65 / 63 ms	0 / 0 ms
W0	3.0 ms	0 ms	47 / 47 ms	0 / 0 ms	0 / 0 ms	3.7 ms	0 ms	78 / 78 ms	0 / 0 ms	0 / 0 ms
W1	107.4 ms	0 ms	53 / 53 ms	42 / 42 ms	0 / 0 ms	108,197.3 ms	0 ms	9,400 / 26 ms	716 / 224 ms	0 / 0 ms
W2	1,260.2 ms	0 ms	194 / 79 ms	10 / 10 ms	0 / 0 ms	108,969.3 ms	0 ms	9,440 / 85 ms	24,405 / 23,623 ms	0 / 0 ms
W3	4.9 ms	0 ms	40 / 40 ms	0 / 0 ms	0 / 0 ms	5.9 ms	0 ms	102 / 102 ms	0 / 0 ms	0 / 0 ms
W4	7.0 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	96,311.0 ms	0 ms	5,124 / 2 ms	0 / 0 ms	0 / 0 ms
W4-sat50	17.1 ms	0 ms	156 / 156 ms	0 / 0 ms	0 / 0 ms	11,591.5 ms	0 ms	1,634 / 271 ms	0 / 0 ms	0 / 0 ms
W4-sat70	21.2 ms	0 ms	156 / 156 ms	0 / 0 ms	0 / 0 ms	74,808.3 ms	0 ms	8,194 / 219 ms	0 / 0 ms	0 / 0 ms
W4-sat90	13.1 ms	0 ms	151 / 151 ms	0 / 0 ms	0 / 0 ms	217,282.7 ms	0 ms	18,988 / 93 ms	0 / 0 ms	0 / 0 ms
W5	9.7 ms	0 ms	2 / 2 ms	0 / 0 ms	0 / 0 ms	86,320.1 ms	0 ms	3,976 / 1 ms	0 / 0 ms	0 / 0 ms
W6-upload	467.4 ms	0 ms	76 / 31 ms	1 / 1 ms	0 / 0 ms	65,433.9 ms	0 ms	5,991 / 79 ms	36,586 / 34,486 ms	0 / 0 ms
W6	6.8 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	15,840.5 ms	0 ms	1,759 / 2 ms	0 / 0 ms	0 / 0 ms
W6-ef32	1.6 ms	0 ms	2 / 2 ms	0 / 0 ms	0 / 0 ms	12,726.4 ms	0 ms	1,387 / 2 ms	0 / 0 ms	0 / 0 ms
W6-ef64	2.0 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	14,418.4 ms	0 ms	1,509 / 2 ms	0 / 0 ms	0 / 0 ms
W6-ef128	2.7 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	15,464.4 ms	0 ms	1,686 / 1 ms	0 / 0 ms	0 / 0 ms
W6-ef256	7.6 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	18,336.9 ms	0 ms	2,176 / 1 ms	0 / 0 ms	0 / 0 ms
W6-ef512	8.1 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	27,282.8 ms	0 ms	3,372 / 1 ms	0 / 0 ms	0 / 0 ms
W7-upload	148.7 ms	0 ms	38 / 28 ms	38 / 38 ms	104 / 104 ms	39,248.8 ms	0 ms	3,633 / 148 ms	181,334 / 176,519 ms	0 / 0 ms
W7	1.5 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	93,182.8 ms	0 ms	3,310 / 3 ms	0 / 0 ms	0 / 0 ms
W7-2bit-upload	170.2 ms	0 ms	43 / 28 ms	53 / 53 ms	0 / 0 ms	44,457.0 ms	0 ms	4,123 / 85 ms	14,994 / 13,827 ms	0 / 0 ms
W7-2bit	2.3 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	93,687.6 ms	0 ms	3,440 / 3 ms	0 / 0 ms	0 / 0 ms
W7-1p5bit-upload	143.3 ms	0 ms	51 / 29 ms	72 / 72 ms	0 / 0 ms	46,578.9 ms	0 ms	4,157 / 104 ms	79,114 / 77,860 ms	0 / 0 ms
W7-1p5bit	2.0 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	93,526.9 ms	0 ms	3,395 / 3 ms	0 / 0 ms	0 / 0 ms
W8-upload	1,353.4 ms	0 ms	183 / 45 ms	61 / 61 ms	753 / 753 ms	27,983.8 ms	0 ms	6,799 / 314 ms	18,380 / 14,853 ms	0 / 0 ms
W8	3.4 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	15,653.6 ms	0 ms	1,873 / 1 ms	0 / 0 ms	0 / 0 ms
W14-1bit-upload	1,109.5 ms	0 ms	152 / 41 ms	69 / 69 ms	0 / 0 ms	54,863.8 ms	0 ms	5,102 / 97 ms	16,034 / 15,307 ms	0 / 0 ms
W14-1bit	2.0 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	94,570.4 ms	0 ms	3,416 / 3 ms	0 / 0 ms	0 / 0 ms
W14-1p5bit-upload	1,126.5 ms	0 ms	153 / 40 ms	41 / 41 ms	0 / 0 ms	53,287.1 ms	0 ms	5,201 / 108 ms	26,373 / 25,216 ms	0 / 0 ms
W14-1p5bit	2.9 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	97,130.4 ms	0 ms	3,715 / 3 ms	0 / 0 ms	0 / 0 ms
W14-2bit-upload	1,112.1 ms	0 ms	152 / 42 ms	43 / 43 ms	0 / 0 ms	72,137.4 ms	0 ms	6,745 / 93 ms	18,667 / 17,541 ms	0 / 0 ms
W14-2bit	2.5 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	94,664.6 ms	0 ms	3,524 / 3 ms	0 / 0 ms	0 / 0 ms
W14-4bit-upload	1,135.9 ms	0 ms	155 / 40 ms	54 / 54 ms	0 / 0 ms	62,528.4 ms	0 ms	6,106 / 131 ms	65,932 / 63,445 ms	0 / 0 ms
W14-4bit	2.2 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	97,214.2 ms	0 ms	3,737 / 3 ms	0 / 0 ms	0 / 0 ms
W9	71.2 ms	0 ms	18 / 18 ms	0 / 0 ms	0 / 0 ms	14,216.7 ms	0 ms	1,639 / 37 ms	0 / 0 ms	0 / 0 ms
W10-ef32	4.0 ms	0 ms	11 / 11 ms	0 / 0 ms	0 / 0 ms	14,951.6 ms	0 ms	1,759 / 23 ms	0 / 0 ms	0 / 0 ms
W10-ef64	5.9 ms	0 ms	10 / 10 ms	0 / 0 ms	0 / 0 ms	41,212.9 ms	0 ms	3,902 / 31 ms	0 / 0 ms	0 / 0 ms
W10-ef128	9.6 ms	0 ms	9 / 9 ms	0 / 0 ms	0 / 0 ms	57,485.6 ms	0 ms	5,598 / 33 ms	0 / 0 ms	0 / 0 ms
W10-ef256	17.3 ms	0 ms	8 / 8 ms	0 / 0 ms	0 / 0 ms	83,088.3 ms	0 ms	8,517 / 37 ms	0 / 0 ms	0 / 0 ms
W10-ef512	24.7 ms	0 ms	7 / 7 ms	0 / 0 ms	0 / 0 ms	92,337.9 ms	0 ms	10,244 / 67 ms	0 / 0 ms	0 / 0 ms
W12-upload	181.1 ms	0 ms	38 / 24 ms	63 / 63 ms	4,928 / 4,928 ms	17,205.5 ms	0 ms	1,673 / 22 ms	212 / 136 ms	0 / 0 ms
W12-sel1	6.6 ms	0 ms	2 / 2 ms	0 / 0 ms	0 / 0 ms	39,910.4 ms	0 ms	4,975 / 2 ms	0 / 0 ms	0 / 0 ms
W12-sel10	8.3 ms	0 ms	2 / 2 ms	3 / 3 ms	0 / 0 ms	57,403.3 ms	0 ms	7,160 / 3 ms	0 / 0 ms	0 / 0 ms
W12-sel1-ef32	7.2 ms	0 ms	2 / 2 ms	0 / 0 ms	0 / 0 ms	68,961.4 ms	0 ms	3,902 / 1 ms	0 / 0 ms	0 / 0 ms
W12-sel1-ef64	6.1 ms	0 ms	2 / 2 ms	0 / 0 ms	0 / 0 ms	27,271.3 ms	0 ms	3,364 / 1 ms	0 / 0 ms	0 / 0 ms
W12-sel1-ef128	6.1 ms	0 ms	2 / 2 ms	2 / 2 ms	0 / 0 ms	39,808.4 ms	0 ms	4,957 / 2 ms	0 / 0 ms	0 / 0 ms
W12-sel1-ef256	6.2 ms	0 ms	2 / 2 ms	0 / 0 ms	0 / 0 ms	54,864.6 ms	0 ms	6,844 / 3 ms	0 / 0 ms	0 / 0 ms
W12-sel1-ef512	5.4 ms	0 ms	2 / 2 ms	0 / 0 ms	0 / 0 ms	68,963.1 ms	0 ms	8,609 / 5 ms	0 / 0 ms	0 / 0 ms
W12-sel10-ef32	2.3 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	29,128.7 ms	0 ms	3,616 / 1 ms	0 / 0 ms	0 / 0 ms
W12-sel10-ef64	3.7 ms	0 ms	3 / 3 ms	0 / 0 ms	0 / 0 ms	41,081.1 ms	0 ms	5,115 / 2 ms	0 / 0 ms	0 / 0 ms
W12-sel10-ef128	5.0 ms	0 ms	2 / 2 ms	0 / 0 ms	0 / 0 ms	57,549.0 ms	0 ms	7,182 / 3 ms	0 / 0 ms	0 / 0 ms
W12-sel10-ef256	11.3 ms	0 ms	2 / 2 ms	0 / 0 ms	0 / 0 ms	74,054.5 ms	0 ms	9,245 / 5 ms	0 / 0 ms	0 / 0 ms
W12-sel10-ef512	70.1 ms	0 ms	18 / 18 ms	0 / 0 ms	0 / 0 ms	87,860.9 ms	0 ms	10,967 / 8 ms	0 / 0 ms	0 / 0 ms
W13	4.6 ms	0 ms	29 / 29 ms	0 / 0 ms	0 / 0 ms	812.8 ms	0 ms	118 / 27 ms	0 / 0 ms	0 / 0 ms
W11-steady	195,850.3 ms	0 ms	25,581 / 7 ms	691 / 147 ms	0 / 0 ms	212,991.1 ms	0 ms	20,494 / 34 ms	141 / 28 ms	0 / 0 ms
W11	506,643.3 ms	0 ms	62,796 / 12 ms	91 / 24 ms	0 / 0 ms	1,211,680.4 ms	0 ms	129,757 / 65 ms	1,819 / 203 ms	0 / 0 ms

A dash under blocked on disk is not a zero: most hosts ship with kernel.task_delayacct=0, which reports the field as a permanent zero and would manufacture the strongest claim here — that the memory-resident engine never waits for a device — out of a sysctl. bench/setup.py apply turns it on.

Hardware counters cycles, DRAM loads, TLB walks, IPC per query

What the core did per query, from perf stat attached to the engine for the same window as the /proc counters. §5's cost model is a set of claims about instructions per cycle, DRAM traffic and TLB reach; these are those quantities, measured on the row the headline quotes rather than on a microbenchmark.

What one query cost, in cycles

Lower is better; the bar between a pair is the gap. Load rows are absent: they have no queries to divide by, and an absolute count under a per-query axis would be a different quantity wearing this one's label.

Demand loads from DRAM per query

Demand loads served from DRAM, times the cache line this host reports. Lower is better. Demand only: the hardware prefetcher's fills are a separate counter and are not in this number, so a row that streams — an exact scan above all — moved far more than this says. On the graph rows, where there is little for a prefetcher to predict, it is most of the traffic and is the outstanding-miss quantity §5.2 argues the engine is limited by.

TLB walks per query

Data TLB misses that reached a page walk, per query. Lower is better. §5.5 argues a memory-resident index needs hugepage care; this is what not taking it costs, and `--no-huge-pages` is the A/B arm that prices it.

Instructions per cycle

How well the core was fed while it ran. Not a score: a low IPC on a memory-bound row is what §5.2 predicts, and a high one on a row that does more work per query is not a win. Read it beside the two charts above.

Branch mispredictions per 1k instructions

A graph traversal is a chain of data-dependent branches and none of them is predictable from the last query, so this is the column that separates a slow kernel from an unpredictable walk. Per thousand instructions rather than per branch: `branches` costs a PMU counter that `ref-cycles` needs more, and MPKI is the comparable form anyway.

workload	strawmANN						Qdrant
workload	IPC	GHz	branch MPKI	cycles/query	demand DRAM/query	TLB walks/query	IPC	GHz	branch MPKI	cycles/query	demand DRAM/query	TLB walks/query
W0-upload	1.50	2.011.006x nominal	5.34	-	-	-	1.29	2.011.006x nominal	4.53	-	-	-
W0	1.38	2.011.005x nominal	5.31	256,241	6.8 KiB	984.0	1.22	2.011.006x nominal	3.12	445,649	37.7 KiB	1,139.6
W1	2.12	2.011.006x nominal	0.70	-	-	-	1.07	2.011.008x nominal	1.24	-	-	-
W2	0.41	2.011.008x nominal	2.37	-	-	-	0.60	2.011.007x nominal	1.71	-	-	-
W3	0.71	2.011.006x nominal	3.92	1,676,546	5,752.4 KiB	4,252.4	0.69	2.011.007x nominal	1.93	3,042,929	4,543.4 KiB	6,071.3
W4	0.30	2.011.008x nominal	3.21	4,012,231	4,265.0 KiB	4,064.9	0.42	2.011.007x nominal	2.02	4,665,621	4,252.6 KiB	5,266.6
W4-sat50	0.65	2.011.007x nominal	3.90	1,847,344	5,418.8 KiB	4,202.3	0.62	2.011.007x nominal	2.20	3,357,861	4,475.4 KiB	6,031.0
W4-sat70	0.59	2.011.008x nominal	3.91	2,033,879	5,071.9 KiB	4,166.3	0.57	2.011.007x nominal	2.27	3,676,178	4,410.4 KiB	5,936.2
W4-sat90	0.44	2.011.007x nominal	3.82	2,708,292	4,439.6 KiB	4,151.7	0.47	2.011.008x nominal	2.32	4,341,456	4,348.9 KiB	5,595.0
W5	0.28	2.011.008x nominal	3.34	4,003,662	4,316.6 KiB	4,012.4	0.39	2.011.007x nominal	2.00	4,631,461	4,215.1 KiB	5,060.4
W6-upload	0.86	2.011.007x nominal	2.06	-	-	-	1.74	2.011.008x nominal	1.44	-	-	-
W6	0.57	2.011.008x nominal	4.10	1,789,401	2,714.7 KiB	5,737.5	1.23	2.011.008x nominal	1.61	2,179,803	836.4 KiB	4,131.9
W6-ef32	0.67	2.011.007x nominal	2.70	613,693	839.4 KiB	1,872.6	1.36	2.011.007x nominal	1.32	871,491	299.1 KiB	1,603.1
W6-ef64	0.61	2.011.008x nominal	3.42	1,023,975	1,492.8 KiB	3,174.2	1.29	2.011.007x nominal	1.45	1,330,144	487.6 KiB	2,491.1
W6-ef128	0.57	2.011.007x nominal	4.11	1,788,262	2,716.0 KiB	5,737.5	1.22	2.011.008x nominal	1.62	2,183,417	833.1 KiB	4,119.7
W6-ef256	0.56	2.011.007x nominal	4.75	3,196,179	4,951.6 KiB	10,828.1	1.18	2.011.007x nominal	1.77	3,737,642	1,477.9 KiB	6,999.0
W6-ef512	0.56	2.011.008x nominal	5.34	5,791,619	9,039.9 KiB	21,074.3	1.15	2.011.009x nominal	1.92	6,616,785	2,703.4 KiB	12,148.0
W7-upload	1.33	2.011.007x nominal	6.13	-	-	-	2.11	2.011.006x nominal	4.58	-	-	-
W7	1.31	2.011.008x nominal	6.58	641,434	421.5 KiB	2,986.2	1.47	2.011.007x nominal	3.50	1,082,919	461.8 KiB	3,366.8
W7-2bit-upload	1.26	2.011.006x nominal	5.79	-	-	-	1.93	2.011.007x nominal	4.08	-	-	-
W7-2bit	1.23	2.011.007x nominal	5.99	726,849	435.7 KiB	3,124.1	1.39	2.011.007x nominal	3.32	1,192,431	474.9 KiB	3,416.8
W7-1p5bit-upload	1.29	2.011.006x nominal	5.99	-	-	-	2.00	2.011.007x nominal	4.27	-	-	-
W7-1p5bit	1.24	2.011.007x nominal	6.22	704,491	431.1 KiB	3,106.4	1.41	2.011.007x nominal	3.42	1,159,883	472.6 KiB	3,454.7
W8-upload	2.31	2.011.007x nominal	0.73	-	-	-	5.40	2.011.006x nominal	0.17	-	-	-
W8	4.33	2.011.006x nominal	0.52	2,341,563	445.2 KiB	3,731.2	3.24	2.011.006x nominal	0.49	3,124,364	514.6 KiB	4,069.6
W14-1bit-upload	0.51	2.011.007x nominal	1.87	-	-	-	2.53	2.011.007x nominal	2.08	-	-	-
W14-1bit	2.09	2.011.007x nominal	2.07	1,039,966	435.2 KiB	4,288.1	1.61	2.011.006x nominal	2.43	1,171,852	445.9 KiB	2,961.7
W14-1p5bit-upload	0.57	2.011.007x nominal	1.71	-	-	-	2.38	2.011.007x nominal	1.68	-	-	-
W14-1p5bit	2.27	2.011.007x nominal	1.58	1,263,192	444.2 KiB	4,499.8	1.49	2.011.006x nominal	2.04	1,514,098	506.8 KiB	3,115.5
W14-2bit-upload	0.52	2.011.007x nominal	1.87	-	-	-	2.90	2.011.007x nominal	1.09	-	-	-
W14-2bit	1.83	2.011.007x nominal	2.59	927,706	440.6 KiB	4,552.6	1.56	2.011.007x nominal	2.22	1,284,774	467.8 KiB	3,258.4
W14-4bit-upload	0.54	2.011.007x nominal	1.84	-	-	-	2.63	2.011.007x nominal	1.07	-	-	-
W14-4bit	1.55	2.011.009x nominal	2.44	1,128,835	492.6 KiB	5,011.2	1.39	2.011.007x nominal	2.16	1,501,938	491.6 KiB	3,675.9
W9	0.58	2.011.008x nominal	0.15	639,305,597	798,578.5 KiB	629,566.0	0.36	2.011.008x nominal	0.09	1,620,243,551	490,680.2 KiB	1,078,523.6
W10-ef32	0.35	2.011.008x nominal	2.32	1,375,044	1,410.3 KiB	1,541.6	0.63	2.011.008x nominal	1.94	1,543,102	1,535.4 KiB	2,355.9
W10-ef64	0.32	2.011.008x nominal	2.87	2,316,020	2,432.8 KiB	2,557.1	0.55	2.011.007x nominal	2.24	2,513,841	2,534.9 KiB	3,597.3
W10-ef128	0.30	2.011.007x nominal	3.22	3,953,554	4,204.2 KiB	4,489.3	0.51	2.011.007x nominal	2.32	4,167,431	4,284.8 KiB	5,705.7
W10-ef256	0.30	2.011.008x nominal	3.55	6,882,434	7,410.0 KiB	8,274.5	0.45	2.011.007x nominal	2.50	7,429,819	7,585.2 KiB	9,387.8
W10-ef512	0.30	2.011.009x nominal	3.89	12,106,011	13,186.5 KiB	15,818.5	0.42	2.011.008x nominal	2.79	13,402,278	13,528.1 KiB	15,758.4
W12-upload	0.52	2.011.009x nominal	2.43	-	-	-	0.76	2.011.007x nominal	1.95	-	-	-
W12-sel1	0.27	2.011.007x nominal	2.45	3,614,960	6,183.6 KiB	4,603.6	1.86	2.011.007x nominal	3.74	11,454,643	3,279.4 KiB	4,834.9
W12-sel10	0.85	2.011.008x nominal	3.62	5,599,476	8,230.0 KiB	10,628.8	1.59	2.011.007x nominal	2.76	19,156,807	8,694.8 KiB	12,206.2
W12-sel1-ef32	0.27	2.011.009x nominal	2.51	3,616,728	6,185.7 KiB	4,607.7	1.75	2.011.007x nominal	2.71	4,162,783	1,529.5 KiB	2,376.0
W12-sel1-ef64	0.27	2.011.009x nominal	2.45	3,617,223	6,177.0 KiB	4,603.5	1.82	2.011.008x nominal	3.21	6,853,731	2,237.2 KiB	3,366.3
W12-sel1-ef128	0.27	2.011.009x nominal	2.47	3,622,889	6,174.3 KiB	4,595.5	1.86	2.011.009x nominal	3.74	11,457,659	3,257.1 KiB	4,952.8
W12-sel1-ef256	0.27	2.011.009x nominal	2.46	3,615,054	6,176.1 KiB	4,602.5	1.90	2.021.010x nominal	4.16	18,707,877	4,492.1 KiB	6,906.8
W12-sel1-ef512	0.27	2.021.010x nominal	2.46	3,625,836	6,172.9 KiB	4,611.7	1.99	2.021.010x nominal	4.38	29,074,855	5,728.6 KiB	9,008.0
W12-sel10-ef32	0.94	2.011.008x nominal	2.72	1,921,812	2,753.6 KiB	3,373.2	1.57	2.011.007x nominal	2.14	7,385,543	3,454.1 KiB	4,876.3
W12-sel10-ef64	0.87	2.011.007x nominal	3.24	3,221,252	4,710.9 KiB	5,878.9	1.58	2.011.007x nominal	2.43	11,795,539	5,453.3 KiB	7,664.1
W12-sel10-ef128	0.84	2.011.007x nominal	3.63	5,641,464	8,251.3 KiB	10,630.8	1.58	2.011.007x nominal	2.76	19,228,198	8,669.6 KiB	12,227.3
W12-sel10-ef256	0.88	2.011.007x nominal	3.87	9,634,463	14,014.7 KiB	19,192.0	1.59	2.011.007x nominal	3.12	30,640,779	13,532.4 KiB	19,340.6
W12-sel10-ef512	0.25	2.011.009x nominal	0.75	32,006,997	38,879.8 KiB	38,130.1	1.61	2.011.006x nominal	3.49	47,360,202	20,298.6 KiB	29,442.9
W13	1.36	2.011.008x nominal	1.29	42,574	1.0 KiB	13.5	1.46	2.021.010x nominal	1.47	138,835	1.4 KiB	39.2
W11-steady	0.32	2.011.007x nominal	1.83	30,774,186	36,870.4 KiB	25,006.2	0.66	2.011.007x nominal	0.67	35,068,822	16,210.2 KiB	34,486.1
W11	0.29	2.011.007x nominal	1.10	149,880,481	177,081.7 KiB	95,234.7	0.51	2.011.007x nominal	0.49	244,593,270	118,339.6 KiB	239,539.1

Nothing here is scaled. Where perf had to multiplex the group, the values it prints are extrapolations from the fraction of the row each counter was on, and they are withheld rather than shown — the same rule the scheduler table applies to a thread that exited. An event this host does not implement is likewise blank, never zero.

The two instruments disagree on these rows. strawmANN: perf and /proc disagree on minor faults by up to 14% on W12-upload; strawmANN: perf and /proc disagree on context switches by up to 7% on W11-steady; Qdrant: perf and /proc disagree on minor faults by up to 43% on W1, W2, W6-upload, W7-upload, W7-1p5bit-upload, W8-upload, W14-1bit-upload, W14-1p5bit-upload, W14-2bit-upload, W14-4bit-upload, W12-upload, W11-steady, W11; Qdrant: perf and /proc disagree on context switches by up to 70% on W8-upload, W14-1bit-upload, W14-1p5bit-upload, W14-2bit-upload, W14-4bit-upload. perf keeps an exiting thread's counts and the /proc sums do not, so a gap here is usually the same lost-thread effect the scheduler table reports as coverage.

Collection config what each engine says it built

Read back from the engine rather than taken from the request. --segments 1 sets Qdrant's default_segment_number, which its optimizer treats as a target, and the segment count is the largest confound in the ef comparison. strawmANN held 1 segment; Qdrant held 2 segments.

Where the engines disagree a cell reads strawmANN / Qdrant, and is marked.

collection	segments	populated segments	requested segments	points	indexed vectors	vector size	hnsw m	hnsw ef_construct	quantization	shards	vector residency
bench0	1 / 2	- / 1	- / 1	990,000	990,000	4	16	100	-	1	cached
bench1	1 / 5	- / 5	- / 1	990,000	0 / 59,500	1,536	16	100	-	1	cached
bench2	1 / 2	- / 1	- / 1	990,000	990,000	1,536	16	100	-	1	cached
bench6	1 / 2	- / 1	- / 1	990,000	990,000	1,536	16	100	scalar	1	cached
bench7	1 / 2	- / 1	- / 1	990,000	990,000	1,536	16	100	binary	1	cached
bench7b2	1 / 2	- / 1	- / 1	990,000	990,000	1,536	16	100	binary	1	cached
bench7b15	1 / 2	- / 1	- / 1	990,000	990,000	1,536	16	100	binary	1	cached
bench8	1 / 2	- / 1	- / 1	990,000	990,000	1,536	16	100	product	1	cached
bench14t1	1 / 2	- / 1	- / 1	990,000	990,000	1,536	16	100	turboquant	1	cached
bench14t15	1 / 2	- / 1	- / 1	990,000	990,000	1,536	16	100	turboquant	1	cached
bench14t2	1 / 2	- / 1	- / 1	990,000	990,000	1,536	16	100	turboquant	1	cached
bench14t4	1 / 2	- / 1	- / 1	990,000	990,000	1,536	16	100	turboquant	1	cached
bench12	1 / 2	- / 1	- / 1	200,000	200,000	1,536	16	100	-	1	cached

After the mutating rows bench2 held strawmANN 1,237,500 (+247,500), 1,075,740 of them indexed; Qdrant 1,237,500 (+247,500), 1,085,700 of them indexed. The table above is the state the throughput, latency and recall rows searched; this is what W11 left, and is that row's subject rather than theirs.

bench1 was read back straight after its row and dropped, with no settle between, so an engine still building it reports a snapshot taken mid-ingest. Its cells are not marked as a disagreement.

Per workload every metric of every row, engine beside engine

Metric down, engine across, so a comparison is two adjacent cells. Metrics a row did not measure are dropped rather than shown empty.

W0-upload transport floor: load

upload · collection=bench0 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=990,000

metric	strawmANN	Qdrant
wall clock	20.5 s	32.0 s	1.56x
of which upload	1.4 s	7.8 s	5.55x
of which index wait	17.1 s	22.1 s	1.29x
time to Green	17.1 s	22.1 s	1.29x
cpu	111.0 s	158.5 s	1.43x
cpu, % of wall	541%	495%
waiting for a core	43.2 ms	15,024.7 ms	347.40x
migrations	2	12,858	6,429.00x
faults min/maj	113,520 / 0	424,958 / 193,878
peak RSS	441.1 MiB	1.4 GiB	3.30x
disk written	60.5 MiB	383.8 MiB
disk read	4.0 KiB	114.2 MiB

W0 d=4 floor: graph traversal with the distance taken out 1.26x

closed-loop · collection=bench0 · client -p 1 -t 2 -c 1 (-t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	4,564	3,615	1.26x
client p50	215 µs	277 µs	1.29x
client p99	262 µs	348 µs	1.33x
client p99.9	307 µs	411 µs	1.34x
server p50	123 µs	174 µs	1.41x
server p99	135 µs	218 µs	1.62x
wall clock	11.0 s	13.9 s	1.26x
queries sent	50,000	50,000
cpu	7.7 s	14.2 s	1.83x
cpu, % of wall	70%	102%
waiting for a core	3.0 ms	3.7 ms	1.24x
migrations	0	197
faults min/maj	0 / 0	464 / 0
peak RSS	441.1 MiB	1.4 GiB	3.30x
disk written	0	0
disk read	0	0

1.26x is 1.74x less work per query x 0.67x cores busy during the row (0.67 against 1.01) x 1.00x clock x 1.09x counted on-CPU share: the middle term is occupancy, not search speed d=4 floor: this measures graph traversal and heaps with the distance taken out, not request plumbing (findings 28: 76% of the server's CPU was HNSW search, 24% kernel; the RPC path is a small remainder)

W1 ingest throughput

upload · collection=bench1 · client -p 8 -t 8 -c 1 (-c bfb default) · n=990,000

metric	strawmANN	Qdrant
wall clock	8.4 s	36.6 s	4.35x
of which upload	5.5 s	34.3 s	6.29x
cpu	11.5 s	260.1 s	22.52x
cpu, % of wall	137%	711%
waiting for a core	107.4 ms	108,197.3 ms	1,007.29x
migrations	0	31,954
faults min/maj	1,576,572 / 0	630,876 / 1,597,619
peak RSS	7.6 GiB	11.5 GiB	1.51x
disk written	5.7 GiB	15.7 GiB
disk read	180.0 KiB	4.1 MiB

W2 index build time

upload · collection=bench2 · client -p 8 -t 8 -c 1 (-c bfb default) · n=990,000

metric	strawmANN	Qdrant
wall clock	262.2 s	355.1 s	1.35x
of which upload	5.5 s	37.0 s	6.76x
of which index wait	253.9 s	315.8 s	1.24x
time to Green	253.9 s	315.8 s	1.24x
cpu	2,022.1 s	2,115.7 s	1.05x
cpu, % of wall	771%	596%
waiting for a core	1,260.2 ms	108,969.3 ms	86.47x
migrations	0	32,701
faults min/maj	1,640,235 / 0	1,414,315 / 1,626,490
peak RSS	7.9 GiB	22.6 GiB	2.87x
disk written	5.7 GiB	25.5 GiB
disk read	0	172.7 MiB

W3 search, fp32, single query 1.69x

ef=128 · closed-loop · collection=bench2 · client -p 1 -t 2 -c 1 (-t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	1,059	627	1.69x
recall@10	0.9667	0.9692	~equal
client p50	948 µs	1.59 ms	1.68x
client p99	1.38 ms	2.27 ms	1.65x
client p99.9	1.90 ms	2.51 ms	1.32x
server p50	846 µs	1.45 ms	1.71x
server p99	1.27 ms	2.11 ms	1.67x
wall clock	47.2 s	79.7 s	1.69x
queries sent	50,000	50,000
cpu	43.6 s	79.5 s	1.82x
cpu, % of wall	92%	100%
waiting for a core	4.9 ms	5.9 ms	1.21x
migrations	0	228
faults min/maj	0 / 0	1,385 / 0
peak RSS	7.9 GiB	22.6 GiB	2.87x
disk written	0	0
disk read	0	0

W4 search, saturating (closed loop) parity

ef=128 · closed-loop · collection=bench2 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	3,572	3,384	parity
recall@10	0.9667	0.9692	~equal
client p50	17.94 ms	18.77 ms	1.05x
client p99	20.22 ms	33.05 ms	1.63x
client p99.9	24.43 ms	44.66 ms	1.83x
server p50	17.83 ms	18.22 ms	1.02x
server p99	20.06 ms	27.99 ms	1.39x
wall clock	14.0 s	14.8 s	1.06x
queries sent	50,000	50,000
cpu	100.3 s	117.4 s	1.17x
cpu, % of wall	715%	792%
waiting for a core	7.0 ms	96,311.0 ms	13,694.37x
migrations	0	5,047
faults min/maj	119 / 0	30,332 / 0
peak RSS	7.9 GiB	22.6 GiB	2.87x
disk written	0	0
disk read	0	0

within the ±5.8% band this dataset's noise floor puts on W4: no measured difference, not a small one

W4-sat50 search, fixed rate at 50% of saturation (open loop) offered

ef=128 · open-loop · offered=1,732/s (50% of saturation, pinned reference 3,463 qps) · collection=bench2 · client -p n/a under --rps -t 16 -c 2 · n=200,000

metric	strawmANN	Qdrant
queries/second	1,732	1,732	offered
recall@10	0.9667	0.9692	~equal
client p50	1.02 ms	1.74 ms	1.70x
client p99	1.62 ms	2.84 ms	1.76x
client p99.9	1.84 ms	3.71 ms	2.02x
server p50	920 µs	1.59 ms	1.73x
server p99	1.51 ms	2.63 ms	1.75x
wall clock	115.6 s	115.6 s	~equal
queries sent	200,000	200,000
cpu	190.5 s	351.1 s	1.84x
cpu, % of wall	165%	304%
waiting for a core	17.1 ms	11,591.5 ms	679.75x
migrations	0	469,270
faults min/maj	3 / 0	4,655 / 0
peak RSS	7.9 GiB	22.6 GiB	2.87x
disk written	0	0
disk read	0	0

W4-sat70 search, fixed rate at 70% of saturation (open loop) offered

ef=128 · open-loop · offered=2,424/s (70% of saturation, pinned reference 3,463 qps) · collection=bench2 · client -p n/a under --rps -t 16 -c 2 · n=200,000

metric	strawmANN	Qdrant
queries/second	2,424	2,424	offered
recall@10	0.9667	0.9692	~equal
client p50	1.11 ms	2.00 ms	1.81x
client p99	1.93 ms	3.79 ms	1.96x
client p99.9	2.28 ms	4.76 ms	2.09x
server p50	1.00 ms	1.82 ms	1.82x
server p99	1.82 ms	3.46 ms	1.90x
wall clock	82.6 s	82.6 s	~equal
queries sent	200,000	200,000
cpu	206.3 s	381.6 s	1.85x
cpu, % of wall	250%	462%
waiting for a core	21.2 ms	74,808.3 ms	3,531.15x
migrations	0	648,154
faults min/maj	12 / 0	3,791 / 0
peak RSS	7.9 GiB	22.6 GiB	2.87x
disk written	0	0
disk read	0	0

W4-sat90 search, fixed rate at 90% of saturation (open loop) offered

ef=128 · open-loop · offered=3,117/s (90% of saturation, pinned reference 3,463 qps) · collection=bench2 · client -p n/a under --rps -t 16 -c 2 · n=200,000

metric	strawmANN	Qdrant
queries/second	3,117	3,117	offered
recall@10	0.9667	0.9692	~equal
client p50	1.38 ms	3.13 ms	2.27x
client p99	2.76 ms	14.05 ms	5.10x
client p99.9	3.44 ms	23.35 ms	6.78x
server p50	1.27 ms	2.70 ms	2.13x
server p99	2.63 ms	10.43 ms	3.96x
wall clock	64.3 s	64.3 s	~equal
queries sent	200,000	200,000
cpu	270.9 s	444.9 s	1.64x
cpu, % of wall	421%	692%
waiting for a core	13.1 ms	217,282.7 ms	16,580.49x
migrations	0	286,321
faults min/maj	9 / 0	25,140 / 0
peak RSS	7.9 GiB	22.6 GiB	2.87x
disk written	0	0
disk read	0	0

W5 search batched (16 distinct dataset queries per request) parity

ef=128 · closed-loop · collection=bench2 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	3,542	3,441	parity
recall@10	0.9667	0.9692	~equal
client p50	288.12 ms	295.78 ms	1.03x
client p99	294.42 ms	339.82 ms	1.15x
client p99.9	303.98 ms	402.78 ms	1.33x
server p50	287.48 ms	294.25 ms	1.02x
server p99	293.67 ms	325.77 ms	1.11x
wall clock	14.1 s	14.5 s	1.03x
queries sent	50,000	50,000
cpu	99.7 s	115.8 s	1.16x
cpu, % of wall	706%	796%
waiting for a core	9.7 ms	86,320.1 ms	8,933.52x
migrations	0	1,552
faults min/maj	752 / 0	22,143 / 0
peak RSS	7.9 GiB	22.6 GiB	2.87x
disk written	0	0
disk read	0	0

per-batch latency (16 queries/request); queries: dataset, random-sample within the ±5.9% band this dataset's noise floor puts on W5: no measured difference, not a small one batched 16, so qps and rps differ by 16x; the ratio uses qps

W6-upload scalar quantization: load

upload · collection=bench6 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=990,000

metric	strawmANN	Qdrant
wall clock	155.1 s	241.5 s	1.56x
of which upload	5.8 s	50.8 s	8.73x
of which index wait	144.5 s	188.4 s	1.30x
time to Green	144.5 s	188.4 s	1.30x
cpu	1,106.8 s	1,048.0 s	0.95x
cpu, % of wall	714%	434%
waiting for a core	467.4 ms	65,433.9 ms	140.00x
migrations	0	30,253
faults min/maj	2,205,991 / 0	2,981,965 / 1,686,609
peak RSS	17.4 GiB	32.4 GiB	1.86x
disk written	5.7 GiB	34.4 GiB
disk read	0	225.8 MiB

W6 quantized: scalar 1.13x

ef=128 · closed-loop · collection=bench6 · client -p 64 -t 16 -c 2 · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	8,112	7,160	1.13x
recall@10	0.9666	0.9676	~equal
client p50	7.90 ms	8.80 ms	1.11x
client p99	8.72 ms	13.85 ms	1.59x
client p99.9	26.73 ms	24.76 ms	0.93x
server p50	7.79 ms	8.45 ms	1.09x
server p99	8.60 ms	12.95 ms	1.51x
wall clock	6.2 s	7.0 s	1.13x
queries sent	50,000	50,000
cpu	44.9 s	55.0 s	1.23x
cpu, % of wall	723%	784%
waiting for a core	6.8 ms	15,840.5 ms	2,344.63x
migrations	0	2,607
faults min/maj	18 / 0	2,779 / 0
peak RSS	17.4 GiB	32.4 GiB	1.86x
disk written	0	0
disk read	0	0

W6-ef32 SQ8 recall control, ef=32 (latency only) 1.45x

ef=32 · closed-loop · collection=bench6 · client -p 64 -t 16 -c 2 · n=50,000 · oversampling=3.2 · rescore=true

metric	strawmANN	Qdrant
queries/second	24,808	17,155	1.45x
recall@10	0.8916	0.8882	~equal
client p50	2.57 ms	3.54 ms	1.38x
client p99	2.87 ms	7.42 ms	2.59x
client p99.9	22.28 ms	24.74 ms	1.11x
server p50	2.46 ms	2.93 ms	1.19x
server p99	2.75 ms	5.90 ms	2.15x
wall clock	2.1 s	3.0 s	1.44x
queries sent	50,000	50,000
cpu	15.5 s	22.2 s	1.43x
cpu, % of wall	753%	750%
waiting for a core	1.6 ms	12,726.4 ms	7,863.76x
migrations	0	5,947
faults min/maj	6 / 0	438 / 0
peak RSS	17.4 GiB	32.4 GiB	1.86x
disk written	0	0
disk read	0	0

W6-ef64 SQ8 recall control, ef=64 (latency only) 1.25x

ef=64 · closed-loop · collection=bench6 · client -p 64 -t 16 -c 2 · n=50,000 · oversampling=6.4 · rescore=true

metric	strawmANN	Qdrant
queries/second	14,496	11,592	1.25x
recall@10	0.9395	0.9396	~equal
client p50	4.42 ms	5.32 ms	1.20x
client p99	4.89 ms	9.83 ms	2.01x
client p99.9	19.24 ms	23.49 ms	1.22x
server p50	4.32 ms	4.84 ms	1.12x
server p99	4.77 ms	8.40 ms	1.76x
wall clock	3.5 s	4.4 s	1.25x
queries sent	50,000	50,000
cpu	25.8 s	33.7 s	1.31x
cpu, % of wall	737%	772%
waiting for a core	2.0 ms	14,418.4 ms	7,077.00x
migrations	0	4,624
faults min/maj	0 / 0	205 / 0
peak RSS	17.4 GiB	32.4 GiB	1.86x
disk written	0	0
disk read	0	0

W6-ef128 SQ8 recall control, ef=128 (latency only) 1.13x

ef=128 · closed-loop · collection=bench6 · client -p 64 -t 16 -c 2 · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	8,120	7,209	1.13x
recall@10	0.9666	0.9676	~equal
client p50	7.90 ms	8.76 ms	1.11x
client p99	8.71 ms	13.74 ms	1.58x
client p99.9	24.79 ms	23.13 ms	0.93x
server p50	7.79 ms	8.37 ms	1.07x
server p99	8.60 ms	12.68 ms	1.48x
wall clock	6.2 s	7.0 s	1.13x
queries sent	50,000	50,000
cpu	44.9 s	55.1 s	1.23x
cpu, % of wall	725%	790%
waiting for a core	2.7 ms	15,464.4 ms	5,828.97x
migrations	0	2,507
faults min/maj	0 / 0	721 / 0
peak RSS	17.4 GiB	32.4 GiB	1.86x
disk written	0	0
disk read	0	0

W6-ef256 SQ8 recall control, ef=256 (latency only) 1.06x

ef=256 · closed-loop · collection=bench6 · client -p 64 -t 16 -c 2 · n=50,000 · oversampling=25.6 · rescore=true

metric	strawmANN	Qdrant
queries/second	4,485	4,251	1.06x
recall@10	0.9808	0.9818	~equal
client p50	14.29 ms	15.00 ms	1.05x
client p99	15.84 ms	17.98 ms	1.14x
client p99.9	23.82 ms	29.11 ms	1.22x
server p50	14.18 ms	14.63 ms	1.03x
server p99	15.69 ms	17.52 ms	1.12x
wall clock	11.2 s	11.8 s	1.06x
queries sent	50,000	50,000
cpu	80.0 s	93.9 s	1.17x
cpu, % of wall	715%	796%
waiting for a core	7.6 ms	18,336.9 ms	2,428.09x
migrations	0	482
faults min/maj	50 / 0	1,416 / 0
peak RSS	17.4 GiB	32.4 GiB	1.86x
disk written	0	0
disk read	0	0

W6-ef512 SQ8 recall control, ef=512 (latency only) parity

ef=512 · closed-loop · collection=bench6 · client -p 64 -t 16 -c 2 · n=50,000 · oversampling=51.2 · rescore=true

metric	strawmANN	Qdrant
queries/second	2,457	2,411	parity
recall@10	0.9893	0.9898	~equal
client p50	26.07 ms	26.50 ms	~equal
client p99	28.96 ms	30.34 ms	1.05x
client p99.9	29.95 ms	41.10 ms	1.37x
server p50	25.96 ms	25.99 ms	~equal
server p99	28.81 ms	29.90 ms	1.04x
wall clock	20.4 s	20.8 s	~equal
queries sent	50,000	50,000
cpu	144.6 s	165.7 s	1.15x
cpu, % of wall	709%	797%
waiting for a core	8.1 ms	27,282.8 ms	3,386.46x
migrations	0	424
faults min/maj	5 / 0	2,482 / 0
peak RSS	17.4 GiB	32.4 GiB	1.86x
disk written	0	0
disk read	0	0

within the ±3.1% band this dataset's noise floor puts on W6-ef512: no measured difference, not a small one

W7-upload binary quantization: load

upload · collection=bench7 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=990,000

metric	strawmANN	Qdrant
wall clock	59.3 s	368.6 s	6.21x
of which upload	6.0 s	178.8 s	29.96x
of which index wait	50.1 s	171.5 s	3.42x
time to Green	50.1 s	171.5 s	3.42x
cpu	288.9 s	608.6 s	2.11x
cpu, % of wall	487%	165%
waiting for a core	148.7 ms	39,248.8 ms	263.93x
migrations	1	27,924	27,924.00x
faults min/maj	1,698,301 / 0	2,997,835 / 1,861,540
peak RSS	24.1 GiB	41.0 GiB	1.70x
disk written	5.7 GiB	37.2 GiB
disk read	12.0 KiB	317.9 MiB

W7 quantized: binary + oversampling 1.72x

ef=128 · closed-loop · collection=bench7 · client -p 64 -t 16 -c 2 · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	23,264	13,558	1.72x
recall@10	0.9418	0.9459	~equal
client p50	2.72 ms	4.16 ms	1.53x
client p99	2.94 ms	13.00 ms	4.42x
client p99.9	18.45 ms	20.16 ms	1.09x
server p50	2.62 ms	2.55 ms	0.97x
server p99	2.83 ms	9.87 ms	3.49x
wall clock	2.2 s	3.7 s	1.70x
queries sent	50,000	50,000
cpu	16.2 s	28.6 s	1.76x
cpu, % of wall	741%	767%
waiting for a core	1.5 ms	93,182.8 ms	61,915.16x
migrations	0	19,523
faults min/maj	45 / 0	4,456 / 0
peak RSS	24.1 GiB	41.0 GiB	1.70x
disk written	0	0
disk read	0	0

W7-2bit-upload binary quantization, 2 bits: load

upload · collection=bench7b2 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=990,000

metric	strawmANN	Qdrant
wall clock	77.0 s	175.4 s	2.28x
of which upload	5.9 s	39.4 s	6.71x
of which index wait	68.2 s	134.3 s	1.97x
time to Green	68.2 s	134.3 s	1.97x
cpu	312.3 s	636.3 s	2.04x
cpu, % of wall	406%	363%
waiting for a core	170.2 ms	44,457.0 ms	261.27x
migrations	3	29,749	9,916.33x
faults min/maj	1,756,298 / 11	2,162,663 / 1,824,782
peak RSS	36.2 GiB	43.6 GiB	1.20x
disk written	5.7 GiB	30.3 GiB
disk read	380.0 KiB	218.6 MiB

W7-2bit quantized: binary quantization, 2 bits + oversampling 1.66x

ef=128 · closed-loop · collection=bench7b2 · client -p 64 -t 16 -c 2 · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	20,459	12,314	1.66x
recall@10	0.9613	0.9646	~equal
client p50	3.11 ms	4.61 ms	1.49x
client p99	3.35 ms	13.92 ms	4.15x
client p99.9	14.11 ms	26.40 ms	1.87x
server p50	3.01 ms	3.05 ms	~equal
server p99	3.24 ms	10.37 ms	3.20x
wall clock	2.5 s	4.1 s	1.65x
queries sent	50,000	50,000
cpu	18.4 s	31.4 s	1.71x
cpu, % of wall	740%	766%
waiting for a core	2.3 ms	93,687.6 ms	40,046.37x
migrations	0	19,425
faults min/maj	0 / 0	3,632 / 0
peak RSS	36.2 GiB	43.6 GiB	1.20x
disk written	0	0
disk read	0	0

W7-1p5bit-upload binary quantization, 1.5 bits: load

upload · collection=bench7b15 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=990,000

metric	strawmANN	Qdrant
wall clock	75.0 s	255.5 s	3.41x
of which upload	5.9 s	88.9 s	15.02x
of which index wait	66.2 s	205.5 s	3.11x
time to Green	66.2 s	205.5 s	3.11x
cpu	298.1 s	627.1 s	2.10x
cpu, % of wall	397%	245%
waiting for a core	143.3 ms	46,578.9 ms	324.94x
migrations	6	27,665	4,610.83x
faults min/maj	1,736,859 / 2	2,437,715 / 1,816,141
peak RSS	36.2 GiB	43.6 GiB	1.20x
disk written	5.7 GiB	32.9 GiB
disk read	344.0 KiB	262.7 MiB

W7-1p5bit quantized: binary quantization, 1.5 bits + oversampling 1.66x

ef=128 · closed-loop · collection=bench7b15 · client -p 64 -t 16 -c 2 · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	21,077	12,685	1.66x
recall@10	0.9579	0.9599	~equal
client p50	3.00 ms	4.48 ms	1.49x
client p99	3.24 ms	13.51 ms	4.17x
client p99.9	25.79 ms	23.95 ms	0.93x
server p50	2.91 ms	2.91 ms	~equal
server p99	3.13 ms	10.08 ms	3.22x
wall clock	2.4 s	4.0 s	1.65x
queries sent	50,000	50,000
cpu	17.8 s	30.6 s	1.72x
cpu, % of wall	736%	768%
waiting for a core	2.0 ms	93,526.9 ms	46,718.83x
migrations	0	19,700
faults min/maj	0 / 0	2,705 / 0
peak RSS	36.2 GiB	43.6 GiB	1.20x
disk written	0	0
disk read	0	0

W8-upload PQ: load

upload · collection=bench8 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=990,000

metric	strawmANN	Qdrant
wall clock	386.1 s	917.1 s	2.38x
of which upload	6.1 s	28.4 s	4.68x
of which index wait	375.3 s	886.3 s	2.36x
time to Green	375.3 s	886.3 s	2.36x
cpu	2,968.7 s	5,981.8 s	2.01x
cpu, % of wall	769%	652%
waiting for a core	1,353.4 ms	27,983.8 ms	20.68x
migrations	0	100,504
faults min/maj	1,859,672 / 0	1,707,263 / 1,611,187
peak RSS	30.2 GiB	43.6 GiB	1.45x
disk written	5.7 GiB	25.3 GiB
disk read	92.0 KiB	175.9 MiB

W8 quantized: PQ 1.21x

ef=128 · closed-loop · collection=bench8 · client -p 64 -t 16 -c 2 · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	6,115	5,067	1.21x
recall@10	0.9650	0.9659	~equal
client p50	10.44 ms	12.55 ms	1.20x
client p99	11.27 ms	15.65 ms	1.39x
client p99.9	26.49 ms	24.60 ms	0.93x
server p50	10.34 ms	12.21 ms	1.18x
server p99	11.15 ms	15.00 ms	1.34x
wall clock	8.3 s	9.9 s	1.20x
queries sent	50,000	50,000
cpu	58.8 s	78.7 s	1.34x
cpu, % of wall	713%	794%
waiting for a core	3.4 ms	15,653.6 ms	4,626.48x
migrations	0	650
faults min/maj	0 / 0	3,577 / 0
peak RSS	30.2 GiB	43.6 GiB	1.45x
disk written	0	0
disk read	0	0

W14-1bit-upload TurboQuant, 1 bit: load

upload · collection=bench14t1 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=990,000

metric	strawmANN	Qdrant
wall clock	268.7 s	177.0 s	0.66x
of which upload	5.8 s	59.8 s	10.28x
of which index wait	259.9 s	115.3 s	0.44x
time to Green	259.9 s	115.3 s	0.44x
cpu	2,059.2 s	650.8 s	0.32x
cpu, % of wall	766%	368%
waiting for a core	1,109.5 ms	54,863.8 ms	49.45x
migrations	1	35,230	35,230.00x
faults min/maj	1,708,495 / 1	2,151,305 / 1,777,602
peak RSS	36.2 GiB	43.6 GiB	1.20x
disk written	5.7 GiB	30.2 GiB
disk read	40.0 KiB	233.0 MiB

W14-1bit quantized: TurboQuant, 1 bit + oversampling 1.12x

ef=128 · closed-loop · collection=bench14t1 · client -p 64 -t 16 -c 2 · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	14,063	12,605	1.12x
recall@10	0.9642	0.9557	~equal
client p50	4.53 ms	4.48 ms	~equal
client p99	4.92 ms	14.03 ms	2.85x
client p99.9	15.31 ms	25.03 ms	1.63x
server p50	4.44 ms	3.00 ms	0.68x
server p99	4.81 ms	10.63 ms	2.21x
wall clock	3.6 s	4.0 s	1.11x
queries sent	50,000	50,000
cpu	26.2 s	30.8 s	1.17x
cpu, % of wall	730%	769%
waiting for a core	2.0 ms	94,570.4 ms	46,769.80x
migrations	0	18,635
faults min/maj	0 / 0	5,125 / 0
peak RSS	36.2 GiB	43.6 GiB	1.20x
disk written	0	0
disk read	0	0

W14-1p5bit-upload TurboQuant, 1.5 bits: load

upload · collection=bench14t15 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=990,000

metric	strawmANN	Qdrant
wall clock	271.8 s	239.6 s	0.88x
of which upload	5.8 s	61.9 s	10.59x
of which index wait	262.9 s	159.4 s	0.61x
time to Green	262.9 s	159.4 s	0.61x
cpu	2,082.3 s	849.4 s	0.41x
cpu, % of wall	766%	354%
waiting for a core	1,126.5 ms	53,287.1 ms	47.31x
migrations	1	29,537	29,537.00x
faults min/maj	1,739,075 / 0	2,294,444 / 1,803,171
peak RSS	36.2 GiB	43.6 GiB	1.20x
disk written	5.7 GiB	30.9 GiB
disk read	20.0 KiB	235.1 MiB

W14-1p5bit quantized: TurboQuant, 1.5 bits + oversampling 1.16x

ef=128 · closed-loop · collection=bench14t15 · client -p 64 -t 16 -c 2 · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	11,493	9,944	1.16x
recall@10	0.9648	0.9577	~equal
client p50	5.55 ms	5.88 ms	1.06x
client p99	6.03 ms	16.09 ms	2.67x
client p99.9	19.45 ms	24.04 ms	1.24x
server p50	5.45 ms	4.68 ms	0.86x
server p99	5.92 ms	12.53 ms	2.12x
wall clock	4.4 s	5.1 s	1.15x
queries sent	50,000	50,000
cpu	31.9 s	39.1 s	1.23x
cpu, % of wall	726%	771%
waiting for a core	2.9 ms	97,130.4 ms	33,761.35x
migrations	0	17,066
faults min/maj	0 / 0	6,949 / 0
peak RSS	36.2 GiB	43.6 GiB	1.20x
disk written	0	0
disk read	0	0

W14-2bit-upload TurboQuant, 2 bits: load

upload · collection=bench14t2 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=990,000

metric	strawmANN	Qdrant
wall clock	269.6 s	253.9 s	0.94x
of which upload	5.9 s	77.2 s	13.14x
of which index wait	260.9 s	177.4 s	0.68x
time to Green	260.9 s	177.4 s	0.68x
cpu	2,059.1 s	989.6 s	0.48x
cpu, % of wall	764%	390%
waiting for a core	1,112.1 ms	72,137.4 ms	64.87x
migrations	0	33,208
faults min/maj	1,772,660 / 0	2,223,495 / 1,827,480
peak RSS	36.2 GiB	43.6 GiB	1.20x
disk written	5.7 GiB	30.7 GiB
disk read	332.0 KiB	213.0 MiB

W14-2bit quantized: TurboQuant, 2 bits + oversampling 1.37x

ef=128 · closed-loop · collection=bench14t2 · client -p 64 -t 16 -c 2 · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	15,832	11,533	1.37x
recall@10	0.9663	0.9624	~equal
client p50	4.02 ms	4.93 ms	1.23x
client p99	4.34 ms	14.78 ms	3.41x
client p99.9	19.09 ms	29.95 ms	1.57x
server p50	3.92 ms	3.55 ms	0.90x
server p99	4.23 ms	11.15 ms	2.64x
wall clock	3.2 s	4.4 s	1.37x
queries sent	50,000	50,000
cpu	23.5 s	33.6 s	1.43x
cpu, % of wall	734%	767%
waiting for a core	2.5 ms	94,664.6 ms	37,682.99x
migrations	0	19,444
faults min/maj	0 / 0	4,576 / 0
peak RSS	36.2 GiB	43.6 GiB	1.20x
disk written	0	0
disk read	0	0

W14-4bit-upload TurboQuant, 4 bits: load

upload · collection=bench14t4 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=990,000

metric	strawmANN	Qdrant
wall clock	271.7 s	321.0 s	1.18x
of which upload	5.9 s	114.2 s	19.47x
of which index wait	262.9 s	204.5 s	0.78x
time to Green	262.9 s	204.5 s	0.78x
cpu	2,071.0 s	1,050.4 s	0.51x
cpu, % of wall	762%	327%
waiting for a core	1,135.9 ms	62,528.4 ms	55.05x
migrations	1	33,566	33,566.00x
faults min/maj	1,900,954 / 2	2,695,357 / 1,919,687
peak RSS	36.2 GiB	43.6 GiB	1.20x
disk written	5.7 GiB	32.1 GiB
disk read	24.0 KiB	216.3 MiB

W14-4bit quantized: TurboQuant, 4 bits + oversampling 1.29x

ef=128 · closed-loop · collection=bench14t4 · client -p 64 -t 16 -c 2 · n=50,000 · oversampling=12.8 · rescore=true

metric	strawmANN	Qdrant
queries/second	12,937	10,006	1.29x
recall@10	0.9672	0.9680	~equal
client p50	4.92 ms	5.79 ms	1.18x
client p99	5.37 ms	16.51 ms	3.07x
client p99.9	27.84 ms	26.39 ms	0.95x
server p50	4.82 ms	4.56 ms	0.95x
server p99	5.27 ms	12.56 ms	2.38x
wall clock	3.9 s	5.0 s	1.29x
queries sent	50,000	50,000
cpu	28.4 s	38.9 s	1.37x
cpu, % of wall	728%	772%
waiting for a core	2.2 ms	97,214.2 ms	44,450.37x
migrations	0	16,353
faults min/maj	0 / 0	4,632 / 0
peak RSS	36.2 GiB	43.6 GiB	1.20x
disk written	0	0
disk read	0	0

W9 exact / brute force 1.27x

exact · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=2,000

metric	strawmANN	Qdrant
queries/second	13	10	1.27x
client p50	620.80 ms	810.76 ms	1.31x
client p99	825.19 ms	848.33 ms	1.03x
client p99.9	836.42 ms	878.97 ms	1.05x
server p50	620.05 ms	809.32 ms	1.31x
server p99	824.50 ms	846.86 ms	1.03x
wall clock	159.0 s	202.1 s	1.27x
queries sent	2,000	2,000
cpu	636.2 s	1,612.5 s	2.53x
cpu, % of wall	400%	798%
waiting for a core	71.2 ms	14,216.7 ms	199.53x
migrations	0	6,641
faults min/maj	0 / 0	1,513 / 0
peak RSS	30.2 GiB	43.6 GiB	1.45x
disk written	0	0
disk read	0	0

1.27x is 2.53x less work per query x 0.50x cores busy during the row (4.00 against 7.98) x 1.00x clock: the middle term is occupancy, not search speed brute force over the whole collection, no index involved

W10-ef32 recall control, ef=32 (latency only) 1.37x

ef=32 · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	10,076	7,343	1.37x
recall@10	0.8928	0.8936	~equal
client p50	783 µs	1.01 ms	1.29x
client p99	1.26 ms	2.00 ms	1.58x
client p99.9	1.57 ms	2.53 ms	1.61x
server p50	682 µs	818 µs	1.20x
server p99	1.16 ms	1.73 ms	1.49x
wall clock	5.0 s	6.8 s	1.37x
queries sent	50,000	50,000
cpu	34.7 s	40.3 s	1.16x
cpu, % of wall	694%	589%
waiting for a core	4.0 ms	14,951.6 ms	3,779.75x
migrations	0	84,216
faults min/maj	0 / 0	1,035 / 0
peak RSS	30.2 GiB	43.6 GiB	1.45x
disk written	0	0
disk read	0	0

W10-ef64 recall control, ef=64 (latency only) 1.27x

ef=64 · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	6,082	4,790	1.27x
recall@10	0.9400	0.9434	~equal
client p50	1.31 ms	1.56 ms	1.19x
client p99	2.03 ms	3.09 ms	1.52x
client p99.9	2.48 ms	3.86 ms	1.56x
server p50	1.21 ms	1.33 ms	1.10x
server p99	1.92 ms	2.73 ms	1.42x
wall clock	8.3 s	10.5 s	1.27x
queries sent	50,000	50,000
cpu	58.1 s	66.3 s	1.14x
cpu, % of wall	704%	633%
waiting for a core	5.9 ms	41,212.9 ms	7,024.29x
migrations	0	157,577
faults min/maj	0 / 0	535 / 0
peak RSS	30.2 GiB	43.6 GiB	1.45x
disk written	0	0
disk read	0	0

W10-ef128 recall control, ef=128 (latency only) 1.17x

ef=128 · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	3,581	3,062	1.17x
recall@10	0.9667	0.9692	~equal
client p50	2.23 ms	2.54 ms	1.14x
client p99	3.47 ms	4.62 ms	1.33x
client p99.9	3.89 ms	5.73 ms	1.47x
server p50	2.12 ms	2.28 ms	1.08x
server p99	3.36 ms	4.25 ms	1.26x
wall clock	14.0 s	16.4 s	1.17x
queries sent	50,000	50,000
cpu	98.9 s	108.4 s	1.10x
cpu, % of wall	706%	662%
waiting for a core	9.6 ms	57,485.6 ms	6,004.56x
migrations	0	177,203
faults min/maj	0 / 0	472 / 0
peak RSS	30.2 GiB	43.6 GiB	1.45x
disk written	0	0
disk read	0	0

W10-ef256 recall control, ef=256 (latency only) 1.14x

ef=256 · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	2,058	1,811	1.14x
recall@10	0.9811	0.9822	~equal
client p50	3.88 ms	4.32 ms	1.11x
client p99	6.13 ms	7.57 ms	1.24x
client p99.9	6.98 ms	9.01 ms	1.29x
server p50	3.76 ms	4.00 ms	1.06x
server p99	6.01 ms	7.11 ms	1.18x
wall clock	24.3 s	27.6 s	1.14x
queries sent	50,000	50,000
cpu	171.9 s	190.4 s	1.11x
cpu, % of wall	706%	689%
waiting for a core	17.3 ms	83,088.3 ms	4,804.97x
migrations	0	196,685
faults min/maj	0 / 0	904 / 0
peak RSS	30.2 GiB	43.6 GiB	1.45x
disk written	0	0
disk read	0	0

W10-ef512 recall control, ef=512 (latency only) parity

ef=512 · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=50,000

metric	strawmANN	Qdrant
queries/second	1,170	1,051	parity
recall@10	0.9894	0.9900	~equal
client p50	6.79 ms	7.53 ms	1.11x
client p99	10.95 ms	12.35 ms	1.13x
client p99.9	12.34 ms	14.29 ms	1.16x
server p50	6.67 ms	7.09 ms	1.06x
server p99	10.82 ms	11.75 ms	1.09x
wall clock	42.8 s	47.6 s	1.11x
queries sent	50,000	50,000
cpu	301.4 s	338.6 s	1.12x
cpu, % of wall	704%	711%
waiting for a core	24.7 ms	92,337.9 ms	3,739.12x
migrations	0	146,149
faults min/maj	0 / 0	2,196 / 0
peak RSS	30.2 GiB	43.6 GiB	1.45x
disk written	0	0
disk read	0	0

within the ±11.6% band this dataset's noise floor puts on W10-ef512: no measured difference, not a small one

W12-upload filtered search: load 200,000 with payloads

upload · collection=bench12 · client -p 2 -t 2 -c 1 (-p, -t, -c bfb default) · n=200,000

metric	strawmANN	Qdrant
wall clock	48.8 s	83.0 s	1.70x
of which upload	1.6 s	6.7 s	4.20x
of which index wait	39.1 s	74.2 s	1.90x
time to Green	39.1 s	74.2 s	1.90x
cpu	299.7 s	473.9 s	1.58x
cpu, % of wall	614%	571%
waiting for a core	181.1 ms	17,205.5 ms	95.00x
migrations	0	6,853
faults min/maj	453,100 / 41	346,436 / 446,599
peak RSS	34.6 GiB	43.6 GiB	1.26x
disk written	1.1 GiB	5.3 GiB
disk read	768.0 KiB	53.5 MiB

W12-sel1 filtered search, one keyword (~1% of bench12) 2.84x

ef=128 · closed-loop · collection=bench12 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	3,958	1,395	2.84x
recall@10	1.0000	0.9995	~equal
client p50	16.08 ms	45.57 ms	2.83x
client p99	17.90 ms	50.14 ms	2.80x
client p99.9	26.35 ms	58.15 ms	2.21x
server p50	15.94 ms	44.95 ms	2.82x
server p99	17.69 ms	49.62 ms	2.81x
wall clock	12.7 s	35.9 s	2.83x
queries sent	50,000	50,000
cpu	90.0 s	286.5 s	3.18x
cpu, % of wall	710%	799%
waiting for a core	6.6 ms	39,910.4 ms	6,010.56x
migrations	0	471
faults min/maj	48 / 0	4,266 / 0
peak RSS	34.6 GiB	43.6 GiB	1.26x
disk written	0	0
disk read	0	0

filtered search over a keyword payload index both engines were asked to build; whether each did is read back from the engine and refused beside the row when it did not. Recall is measured under the filter, against ground truth restricted to the matching set

W12-sel10 filtered search, any of 10 keywords (~10% of bench12) 3.04x

ef=128 · closed-loop · collection=bench12 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	2,536	835	3.04x
recall@10	0.9893	0.9972	~equal
client p50	25.30 ms	76.84 ms	3.04x
client p99	27.51 ms	84.25 ms	3.06x
client p99.9	32.76 ms	98.42 ms	3.00x
server p50	25.19 ms	75.96 ms	3.02x
server p99	27.26 ms	83.84 ms	3.08x
wall clock	19.8 s	59.9 s	3.03x
queries sent	50,000	50,000
cpu	139.8 s	478.6 s	3.42x
cpu, % of wall	708%	798%
waiting for a core	8.3 ms	57,403.3 ms	6,914.33x
migrations	0	1,105
faults min/maj	6 / 2	6,255 / 0
peak RSS	34.6 GiB	43.6 GiB	1.26x
disk written	0	0
disk read	12.0 KiB	0

W12-sel1-ef32 filtered recall control, one keyword, ef=32 (latency only)

ef=32 · closed-loop · collection=bench12 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	3,959	3,802	-
recall@10	1.0000	0.9829
client p50	16.09 ms	16.52 ms
client p99	17.88 ms	30.08 ms
client p99.9	30.97 ms	44.02 ms
server p50	15.94 ms	16.12 ms
server p99	17.68 ms	25.77 ms
wall clock	12.7 s	13.2 s
queries sent	50,000	50,000
cpu	89.8 s	104.7 s
cpu, % of wall	709%	793%
waiting for a core	7.2 ms	68,961.4 ms
migrations	0	2,672
faults min/maj	0 / 0	19,528 / 0
peak RSS	34.6 GiB	43.6 GiB
disk written	0	0
disk read	0	0

recall unequal: 1.0000 vs 0.9829; §7.4 compares at equal recall

W12-sel1-ef64 filtered recall control, one keyword, ef=64 (latency only) 1.70x

ef=64 · closed-loop · collection=bench12 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	3,958	2,329	1.70x
recall@10	1.0000	0.9968	~equal
client p50	16.10 ms	27.30 ms	1.70x
client p99	17.85 ms	30.88 ms	1.73x
client p99.9	23.13 ms	38.66 ms	1.67x
server p50	15.98 ms	26.92 ms	1.68x
server p99	17.66 ms	30.28 ms	1.71x
wall clock	12.7 s	21.5 s	1.70x
queries sent	50,000	50,000
cpu	90.1 s	171.6 s	1.90x
cpu, % of wall	710%	798%
waiting for a core	6.1 ms	27,271.3 ms	4,442.49x
migrations	0	303
faults min/maj	0 / 0	2,525 / 0
peak RSS	34.6 GiB	43.6 GiB	1.26x
disk written	0	0
disk read	0	0

W12-sel1-ef128 filtered recall control, one keyword, ef=128 (latency only) 2.82x

ef=128 · closed-loop · collection=bench12 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	3,941	1,398	2.82x
recall@10	1.0000	0.9995	~equal
client p50	16.10 ms	45.80 ms	2.84x
client p99	17.86 ms	49.57 ms	2.78x
client p99.9	52.45 ms	56.59 ms	1.08x
server p50	15.96 ms	45.02 ms	2.82x
server p99	17.67 ms	48.95 ms	2.77x
wall clock	12.7 s	35.8 s	2.81x
queries sent	50,000	50,000
cpu	90.0 s	285.9 s	3.18x
cpu, % of wall	707%	798%
waiting for a core	6.1 ms	39,808.4 ms	6,481.60x
migrations	0	550
faults min/maj	13 / 10	3,145 / 0
peak RSS	34.6 GiB	43.6 GiB	1.26x
disk written	0	0
disk read	88.0 KiB	0

W12-sel1-ef256 filtered recall control, one keyword, ef=256 (latency only) 4.63x

ef=256 · closed-loop · collection=bench12 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	3,966	857	4.63x
recall@10	1.0000	0.9999	~equal
client p50	16.08 ms	74.30 ms	4.62x
client p99	17.93 ms	80.77 ms	4.50x
client p99.9	37.02 ms	88.55 ms	2.39x
server p50	15.95 ms	73.85 ms	4.63x
server p99	17.72 ms	79.83 ms	4.51x
wall clock	12.7 s	58.4 s	4.61x
queries sent	50,000	50,000
cpu	89.9 s	466.5 s	5.19x
cpu, % of wall	709%	799%
waiting for a core	6.2 ms	54,864.6 ms	8,851.00x
migrations	0	1,177
faults min/maj	54 / 0	6,221 / 0
peak RSS	34.6 GiB	43.6 GiB	1.26x
disk written	0	0
disk read	0	0

W12-sel1-ef512 filtered recall control, one keyword, ef=512 (latency only) 7.16x

ef=512 · closed-loop · collection=bench12 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	3,956	552	7.16x
recall@10	1.0000	1.0000	~equal
client p50	16.13 ms	115.60 ms	7.17x
client p99	17.84 ms	124.34 ms	6.97x
client p99.9	30.91 ms	134.45 ms	4.35x
server p50	15.98 ms	114.90 ms	7.19x
server p99	17.66 ms	123.81 ms	7.01x
wall clock	12.7 s	90.6 s	7.14x
queries sent	50,000	50,000
cpu	90.2 s	723.8 s	8.03x
cpu, % of wall	711%	799%
waiting for a core	5.4 ms	68,963.1 ms	12,822.75x
migrations	0	2,374
faults min/maj	0 / 0	8,405 / 0
peak RSS	34.6 GiB	43.6 GiB	1.26x
disk written	0	0
disk read	0	0

W12-sel10-ef32 filtered recall control, any of 10 keywords, ef=32 (latency only)

ef=32 · closed-loop · collection=bench12 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	7,481	2,160	-
recall@10	0.9378	0.9699
client p50	8.51 ms	29.69 ms
client p99	9.45 ms	32.99 ms
client p99.9	20.99 ms	43.04 ms
server p50	8.41 ms	29.05 ms
server p99	9.29 ms	32.60 ms
wall clock	6.7 s	23.2 s
queries sent	50,000	50,000
cpu	48.2 s	185.2 s
cpu, % of wall	717%	799%
waiting for a core	2.3 ms	29,128.7 ms
migrations	0	303
faults min/maj	0 / 0	3,330 / 0
peak RSS	34.6 GiB	43.6 GiB
disk written	0	0
disk read	0	0

recall unequal: 0.9378 vs 0.9699; §7.4 compares at equal recall

W12-sel10-ef64 filtered recall control, any of 10 keywords, ef=64 (latency only)

ef=64 · closed-loop · collection=bench12 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	4,433	1,354	-
recall@10	0.9728	0.9901
client p50	14.46 ms	47.13 ms
client p99	15.77 ms	52.55 ms
client p99.9	29.24 ms	64.03 ms
server p50	14.35 ms	46.77 ms
server p99	15.62 ms	51.90 ms
wall clock	11.3 s	37.0 s
queries sent	50,000	50,000
cpu	80.7 s	294.3 s
cpu, % of wall	713%	796%
waiting for a core	3.7 ms	41,081.1 ms
migrations	0	545
faults min/maj	0 / 0	4,735 / 0
peak RSS	34.6 GiB	43.6 GiB
disk written	0	0
disk read	0	0

recall unequal: 0.9728 vs 0.9901; §7.4 compares at equal recall

W12-sel10-ef128 filtered recall control, any of 10 keywords, ef=128 (latency only) 3.03x

ef=128 · closed-loop · collection=bench12 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	2,517	832	3.03x
recall@10	0.9893	0.9972	~equal
client p50	25.52 ms	76.99 ms	3.02x
client p99	27.65 ms	84.22 ms	3.05x
client p99.9	28.54 ms	92.99 ms	3.26x
server p50	25.41 ms	76.18 ms	3.00x
server p99	27.52 ms	83.78 ms	3.04x
wall clock	19.9 s	60.2 s	3.02x
queries sent	50,000	50,000
cpu	141.0 s	479.6 s	3.40x
cpu, % of wall	708%	797%
waiting for a core	5.0 ms	57,549.0 ms	11,424.40x
migrations	0	1,042
faults min/maj	0 / 0	5,123 / 0
peak RSS	34.6 GiB	43.6 GiB	1.26x
disk written	0	0
disk read	0	0

W12-sel10-ef256 filtered recall control, any of 10 keywords, ef=256 (latency only) 2.81x

ef=256 · closed-loop · collection=bench12 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	1,468	522	2.81x
recall@10	0.9959	0.9993	~equal
client p50	43.72 ms	122.27 ms	2.80x
client p99	47.08 ms	133.88 ms	2.84x
client p99.9	48.66 ms	143.82 ms	2.96x
server p50	43.60 ms	121.76 ms	2.79x
server p99	46.93 ms	132.97 ms	2.83x
wall clock	34.1 s	95.7 s	2.81x
queries sent	50,000	50,000
cpu	240.2 s	765.9 s	3.19x
cpu, % of wall	704%	800%
waiting for a core	11.3 ms	74,054.5 ms	6,565.53x
migrations	0	2,862
faults min/maj	3 / 0	10,499 / 0
peak RSS	34.6 GiB	43.6 GiB	1.26x
disk written	0	0
disk read	0	0

W12-sel10-ef512 filtered recall control, any of 10 keywords, ef=512 (latency only) 1.31x

ef=512 · closed-loop · collection=bench12 · client -p 64 -t 16 -c 2 · n=50,000

metric	strawmANN	Qdrant
queries/second	443	338	1.31x
recall@10	1.0000	0.9998	~equal
client p50	147.64 ms	189.02 ms	1.28x
client p99	160.17 ms	205.24 ms	1.28x
client p99.9	163.18 ms	225.01 ms	1.38x
server p50	147.26 ms	188.12 ms	1.28x
server p99	159.79 ms	204.73 ms	1.28x
wall clock	112.8 s	148.0 s	1.31x
queries sent	50,000	50,000
cpu	796.9 s	1,183.0 s	1.48x
cpu, % of wall	706%	799%
waiting for a core	70.1 ms	87,860.9 ms	1,253.53x
migrations	0	5,854
faults min/maj	0 / 0	15,545 / 0
peak RSS	34.6 GiB	43.6 GiB	1.26x
disk written	0	0
disk read	0	0

W13 scroll / pagination 1.09x

closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=200,000

metric	strawmANN	Qdrant
queries/second	50,411	46,314	1.09x
client p50	139 µs	159 µs	1.15x
client p99	233 µs	292 µs	1.25x
client p99.9	271 µs	446 µs	1.64x
server p50	11 µs	28 µs	2.57x
server p99	16 µs	89 µs	5.58x
wall clock	4.1 s	2.2 s	0.54x
queries sent	200,000	100,000
cpu	5.6 s	9.2 s	1.63x
cpu, % of wall	137%	411%
waiting for a core	4.6 ms	812.8 ms	176.61x
migrations	0	113,205
faults min/maj	0 / 0	677 / 0
peak RSS	34.6 GiB	43.6 GiB	1.26x
disk written	0	0
disk read	0	0

strawmANN: -n 200000 (4x the table's 50000) Qdrant: -n 100000 (2x the table's 50000) strawmANN: 0.13 ms of the client's 0.14 ms p50 is not server time (13x), so 92% of this row is the load generator and the socket Qdrant: 0.13 ms of the client's 0.16 ms p50 is not server time (6x), so 82% of this row is the load generator and the socket 1.09x is 3.26x less work per query x 0.33x cores busy during the row (1.38 against 4.21) x 1.00x clock x 1.02x counted on-CPU share: the middle term is occupancy, not search speed

W11-steady mixed read/write below the rebuild threshold: search bench2 while 49,500 synthetic points append

ef=128 · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=100,000,000

metric	strawmANN	Qdrant
queries/second	541	494	-
client p50	13.96 ms	12.14 ms
client p99	40.01 ms	68.87 ms
client p99.9	153.28 ms	87.99 ms
server p50	12.87 ms	11.29 ms
server p99	39.02 ms	67.79 ms
wall clock	32.7 s	32.7 s
queries sent	16,480	14,203
cpu	256.2 s	252.7 s
cpu, % of wall	784%	773%
waiting for a core	195,850.3 ms	212,991.1 ms
migrations	779	44,937
faults min/maj	485,661 / 15,720	652,509 / 127,904
peak RSS	36.2 GiB	43.6 GiB
disk written	290.1 MiB	12.2 GiB
disk read	64.4 MiB	73.6 MiB

strawmANN: qps over the 26.1 s of the append the search saw (505 over the whole search); append 1,900 points/s Qdrant: qps over the 26.1 s of the append the search saw (427 over the whole search); append 1,900 points/s strawmANN pass 1 +15% against two passes that agree: one pass set the spread on this row, so a band built from it is not a noise band search-during-write; no recall join

W11 mixed read/write: search bench2 while 198,000 synthetic points append (runs last)

ef=128 · closed-loop · collection=bench2 · client -p 8 -t 2 -c 1 (-t, -c bfb default) · n=100,000,000

metric	strawmANN	Qdrant
queries/second	121	72	-
client p50	58.76 ms	115.38 ms
client p99	196.61 ms	230.21 ms
client p99.9	225.00 ms	371.97 ms
server p50	57.49 ms	113.96 ms
server p99	194.82 ms	228.68 ms
wall clock	75.2 s	187.5 s
queries sent	8,006	12,040
cpu	603.3 s	1,488.1 s
cpu, % of wall	802%	794%
waiting for a core	506,643.3 ms	1,211,680.4 ms
migrations	1,011	37,046
faults min/maj	769,326 / 1,133	1,053,039 / 352,436
peak RSS	36.2 GiB	43.6 GiB
disk written	1.1 GiB	9.8 GiB
disk read	7.5 MiB	167.0 MiB

strawmANN: qps over the 60.0 s of the append the search saw (106 over the whole search); append 3,300 points/s Qdrant: qps over the 186.7 s of the append the search saw (68 over the whole search); append 1,322 points/s strawmANN -11% across its passes, monotonically: the spread on this row is drift rather than noise, so the median is a trend's midpoint and not a repeatable measurement search-during-write; no recall join

Host discipline the §7.1 gate, per check, and ambient load per row

Ambient load per row

The §7.1 gate checks the machine once, at the start. Colour is the engine, as everywhere else; a hatched bar with a red edge is a row that had another process on the box while it ran, which the gate cannot see.

strawmANN gate pass

environment hash 373e8531fd3d1bf6

okAMD Ryzen AI 9 HX PRO 370 w/ Radeon 890M (24 logical)

okgovernor=performance

okboost disabled

oksmt=on

declared rather than required (§7.1 asks for SMT on or off, not

for off), and hashed, so these rows never share a chart with

SMT-off ones

cpu N and cpu N+12 share a core (12 pairs);

a --server-cpus or --client-cpus set holding both cpus of a

pair measures the engine on half the cores it names

okprofile=as-deployed (isolcpus=none nohz_full=none)

the scheduler is running as it ships, so these numbers describe a

deployment rather than the engine in isolation; they may not be

compared against an `isolated` run, and the environment hash

enforces that

okthp=madvise numa_nodes=1 numa_balancing=0

okperf_event_paranoid=-1

oktask_delayacct=1, so time blocked on a device is measurable

okperf at /usr/bin/perf, so `workloads.py run --perf` can attach

okcgroup io controller reaches the engines' own scope, so block-layer read/write operations are measurable

okzig=0.16.0

busy: Xorg(7%)

okmachine is quiescent: 0.11 cores of foreign load, under the 1.00 budget

measurement profile: as-deployed

Qdrant gate pass

environment hash 373e8531fd3d1bf6

okAMD Ryzen AI 9 HX PRO 370 w/ Radeon 890M (24 logical)

okgovernor=performance

okboost disabled

oksmt=on

declared rather than required (§7.1 asks for SMT on or off, not

for off), and hashed, so these rows never share a chart with

SMT-off ones

cpu N and cpu N+12 share a core (12 pairs);

a --server-cpus or --client-cpus set holding both cpus of a

pair measures the engine on half the cores it names

okprofile=as-deployed (isolcpus=none nohz_full=none)

the scheduler is running as it ships, so these numbers describe a

deployment rather than the engine in isolation; they may not be

compared against an `isolated` run, and the environment hash

enforces that

okthp=madvise numa_nodes=1 numa_balancing=0

okperf_event_paranoid=-1

oktask_delayacct=1, so time blocked on a device is measurable

okperf at /usr/bin/perf, so `workloads.py run --perf` can attach

okcgroup io controller reaches the engines' own scope, so block-layer read/write operations are measurable

okzig=0.16.0

busy: Xorg(6%)

okmachine is quiescent: 0.10 cores of foreign load, under the 1.00 budget

measurement profile: as-deployed

What this does not establish

No Qdrant developer has used this tool. Nothing here has been run, checked against something already known, or disagreed with by anyone outside the project: every number has one author and one reviewer, and they are the same person. A result you can refute is more useful to us than one you accept.

It has never been pointed at a real Qdrant regression. The harness detects a known ISA slowdown and correctly reports no difference on a control row. That is internal consistency, not evidence it would flag a regression in your tree or stay quiet through a refactor. bench/harness/qdrant_ab.py exists to run two of your commits through it blind, and that experiment has not been done.

Part of the measured throughput gap is kernel width, not architecture. Read from a dev checkout: Qdrant's distance path tops out at four 256-bit accumulators and has no AVX-512, while these kernels use 512-bit ones on a machine that has them. That is a real difference and it is not the same claim as "the design is faster".

The engines are not equivalent, by construction. §1 removes sharding, replication, consensus, snapshots, sparse vectors, multivectors and disk-resident operation — permanently. Payload storage and filtering were in that list until 2026-09-03 and are now built (M7), so W12 measures a filtered search with a keyword index on both engines. Anything still unbuilt answers UNIMPLEMENTED naming the construct and reports n/a, never a silent degradation. A strawman that was not faster would mean it was badly built; the question is by how much, and where the model was wrong.

Reproducing this

Everything below runs from a clone. The engine has no dependencies; the analysis path is a uv project pinned by bench/uv.lock.

scripts/doctor.py                    # what can this machine measure?
bench/setup.py check                 # §7.1 host gate: governor, boost, isolation, idle
conformance/datasets/datasets.py fetch

# one engine at a time, §7.1
zig build -Doptimize=ReleaseFast
# --connections matters: W4 opens ~32 sockets (-t 16 -c 2) and a server with
# fewer closes the excess before the HTTP/2 preface, which the client reports
# only as "transport error".
./zig-out/bin/strawmann --port 6334 --connections 64
bench/harness/workloads.py run http://localhost:6334 strawmann --sink --report

# the other half of every number: recall, and the licence to compare at all
cd conformance
cargo run --release -- relevance --engine http://localhost:6334 --label strawmann \
  --base $DATA/dbpedia-openai-1m/base.fbin --queries $DATA/dbpedia-openai-1m/queries.fbin \
  --ground-truth $DATA/dbpedia-openai-1m/gt/dbpedia-openai-1m.cosine.k100.gt.json --metric cosine \
  --ef 32 64 128 256 512 --json ../bench/results/strawmann/recall.json
cargo run --release -- differ --strawmann http://localhost:6334 --qdrant http://localhost:6434 \
  --base $DATA/dbpedia-openai-1m/base.fbin --queries $DATA/dbpedia-openai-1m/queries.fbin \
  --ground-truth $DATA/dbpedia-openai-1m/gt/dbpedia-openai-1m.cosine.k100.gt.json \
  --json ../bench/results/strawmann/conformance.json

uv run --project bench bench/harness/report.py strawmann qdrant

Disagreements are the point. docs/spec.md is what all of this is measured against, docs/bugs.md lists the measurement bugs found so far (each produced a plausible wrong number rather than a failure), and docs/validation.md records what has and has not been checked.

Generated by bench/harness/report.py from bench/results/. Self-contained: charts are embedded, nothing is fetched at open time. Regenerate with uv run --project bench bench/harness/report.py sm-dbp1m-perf-1008 qd-dbp1m-perf-1008.

Summary

Throughput against recall

Throughput

Throughput by workload

Recall

At matched recall

At matched recall, SQ8

At matched recall, filtered to 10%

Recall by ef

What each quantization costs in recall

Latency

Ingest and index build

Where the load time goes

Memory and disk

Appendix

Dataset

Host

W10: throughput against ef

Per workload

Time spent waiting for a core

Peak memory per row

Time stalled on I/O

Time stalled on memory

What one query cost, in cycles

Demand loads from DRAM per query

TLB walks per query

Instructions per cycle

Branch mispredictions per 1k instructions

W0-upload transport floor: load

W0 d=4 floor: graph traversal with the distance taken out 1.26x

W1 ingest throughput

W2 index build time

W3 search, fp32, single query 1.69x

W4 search, saturating (closed loop) parity

W4-sat50 search, fixed rate at 50% of saturation (open loop) offered

W4-sat70 search, fixed rate at 70% of saturation (open loop) offered

W4-sat90 search, fixed rate at 90% of saturation (open loop) offered

W5 search batched (16 distinct dataset queries per request) parity

W6-upload scalar quantization: load

W6 quantized: scalar 1.13x

W6-ef32 SQ8 recall control, ef=32 (latency only) 1.45x

W6-ef64 SQ8 recall control, ef=64 (latency only) 1.25x

W6-ef128 SQ8 recall control, ef=128 (latency only) 1.13x

W6-ef256 SQ8 recall control, ef=256 (latency only) 1.06x

W6-ef512 SQ8 recall control, ef=512 (latency only) parity

W7-upload binary quantization: load

W7 quantized: binary + oversampling 1.72x

W7-2bit-upload binary quantization, 2 bits: load

W7-2bit quantized: binary quantization, 2 bits + oversampling 1.66x

W7-1p5bit-upload binary quantization, 1.5 bits: load

W7-1p5bit quantized: binary quantization, 1.5 bits + oversampling 1.66x

W8-upload PQ: load

W8 quantized: PQ 1.21x

W14-1bit-upload TurboQuant, 1 bit: load

W14-1bit quantized: TurboQuant, 1 bit + oversampling 1.12x

W14-1p5bit-upload TurboQuant, 1.5 bits: load

W14-1p5bit quantized: TurboQuant, 1.5 bits + oversampling 1.16x

W14-2bit-upload TurboQuant, 2 bits: load

W14-2bit quantized: TurboQuant, 2 bits + oversampling 1.37x

W14-4bit-upload TurboQuant, 4 bits: load

W14-4bit quantized: TurboQuant, 4 bits + oversampling 1.29x

W9 exact / brute force 1.27x

W10-ef32 recall control, ef=32 (latency only) 1.37x

W10-ef64 recall control, ef=64 (latency only) 1.27x

W10-ef128 recall control, ef=128 (latency only) 1.17x

W10-ef256 recall control, ef=256 (latency only) 1.14x

W10-ef512 recall control, ef=512 (latency only) parity

W12-upload filtered search: load 200,000 with payloads

W12-sel1 filtered search, one keyword (~1% of bench12) 2.84x

W12-sel10 filtered search, any of 10 keywords (~10% of bench12) 3.04x

W12-sel1-ef32 filtered recall control, one keyword, ef=32 (latency only)

W12-sel1-ef64 filtered recall control, one keyword, ef=64 (latency only) 1.70x

W12-sel1-ef128 filtered recall control, one keyword, ef=128 (latency only) 2.82x

W12-sel1-ef256 filtered recall control, one keyword, ef=256 (latency only) 4.63x

W12-sel1-ef512 filtered recall control, one keyword, ef=512 (latency only) 7.16x

W12-sel10-ef32 filtered recall control, any of 10 keywords, ef=32 (latency only)

W12-sel10-ef64 filtered recall control, any of 10 keywords, ef=64 (latency only)

W12-sel10-ef128 filtered recall control, any of 10 keywords, ef=128 (latency only) 3.03x

W12-sel10-ef256 filtered recall control, any of 10 keywords, ef=256 (latency only) 2.81x

W12-sel10-ef512 filtered recall control, any of 10 keywords, ef=512 (latency only) 1.31x

Recall by `ef`