Benchmarks / turbopuffer workloads

turbopuffer's own benchmark, run on WarpSearch.

turbopuffer publishes latency for a fixed set of workloads and open-sources the harness that produces it. We pointed that same harness at WarpSearch's turbopuffer-compatible API, with the same data, query rate and duration: 10M Cohere Wikipedia embeddings, 1024 dimensions, cosine, top 10, 8 queries a second for 30 minutes on a warm cache.

2.7xfaster p99 on the 10M-vector workload, 10 ms vs 27 ms
1.6xfaster p50, 9 ms vs 14 ms
4.7xfaster p99 on VectorDBBench at recall 0.90
3xfaster keyword search at 1M passages

Vector, 10M, warm

Faster at every percentile

Shorter is better. All bars share one scale.

p501.6x faster
WarpSearch9 ms
turbopuffer14 ms
p901.7x faster
WarpSearch10 ms
turbopuffer17 ms
p992.7x faster
WarpSearch10 ms
turbopuffer27 ms

turbopuffer: homepage figures for this workload, client in the same region, as of 2026-11-04. WarpSearch: Serverless, server time reported by the service, 13,851 queries after the first 500, 0 errors, run 2026-11-14.

Scaling

Same harness, three sizes

The website configuration at 100K, 1M and 10M documents, at 8 queries a second with a warm cache. Median latency stays under 10 ms as the namespace grows 100x.

DocumentsQueriesp50p90p99Errors
100K x 10244,7995 ms6 ms6 ms0
1M x 102414,3995 ms6 ms6 ms0
10M x 102413,8519 ms10 ms10 ms0

Server time, WarpSearch Serverless, runs 2026-11-14. turbopuffer publishes its homepage figures for 10M only.

VectorDBBench

And at the same recall

turbopuffer also appears in VectorDBBench, which reports recall next to latency. On Cohere 1M x 768, WarpSearch returns slightly better neighbours in a fraction of the time.

Cohere 1M x 768, p99 at recall 0.904.7x faster
WarpSearch12 ms
turbopuffer56.7 ms

recall@10: WarpSearch 0.902, turbopuffer 0.899. See the vector search results.

turbopuffer: VectorDBBench result file, case Performance768D1M, serial p99, 2026-03. WarpSearch: Serverless, server p99, run 2026-11-13.

Full-text, warm

Keyword search, too

BM25 over MS MARCO passages, top 10, warm cache.

Keyword search, 1M passages, p503x faster
WarpSearch4.4 ms
turbopuffer13 ms
Keyword search, 1M passages, p992.4x faster
WarpSearch12 ms
turbopuffer29 ms

turbopuffer: /docs/performance, full-text on 1M MS MARCO documents, warm, as of 2026-11-04. WarpSearch: 1M MS MARCO passages, single client, warm.

Full-text, query by query

turbopuffer's own full-text queries

The 5M-article English Wikipedia corpus and the queries turbopuffer published for its FTS v2 engine, top 10. Median latency.

QueryWarpSearchturbopuffer FTS v2At p50
the incredibles1 ms3.3 ms3.3x faster
to be or not to be8 ms18.4 ms2.3x faster
lord of the rings2 ms4.2 ms2.1x faster
new york2 ms4.0 ms2x faster
united states constitution2 ms3.8 ms1.9x faster

turbopuffer: appendix of its blog post on FTS v2 (/blog/fts-v2-maxscore). WarpSearch: Serverless, through the turbopuffer-compatible API, server time, median of 150 runs per query, 2026-11-04.

Already on turbopuffer? Keep your SDK. Point its base URL at https://api.warpsearch.io/compat/turbopuffer and use a WarpSearch API key. The compatibility guide lists the supported operations, and the price comparison shows what the same workload costs.

Reproduce

turbopuffer's open-source harness (MIT), pointed at the compatible endpoint, using the website configuration unchanged. Our patch sends vectors as base64 and skips turbopuffer-specific index checks; it is published with the run.

# base URL: https://api.warpsearch.io/compat/turbopuffer
tpufbench run --warm-cache --namespace-setup-concurrency 8 \
  benchmarks/website/vector-10m-hot.toml

Run reports: 20261114T035946Z (10M), 20261114T025847Z (1M), 20261114T024015Z (100K).

Cookie settings

Optional tracking stays off until you choose it. Your choice applies to this browser for 180 days. Read the details.

Essential storage keeps you signed in, protects forms, and remembers this choice. It is always available.