Measured on Semurg r11.54 to r11.71, 29 September to 7 October 2026

Graph, key-value, object, document, search and knowledge graph: wins and losses in the same chart.

The same question, the same answer, the same machine. Every other engine in its fastest form. Nine more domains are added here as their rows pass the same checks.

Every benchmark

How to read the charts

Every chart compares Semurg with another engine on the same question.
Each bar is a ratio: how many times faster (red, to the right) or slower (black, to the left) Semurg is than the engine named on that row.
Beside each name are the two measured figures, Semurg's first.

Semurg fasterSemurg slowerThe centre line: the same speed. Hover a bar for the detail.

Inside your application (the left column): Semurg runs in your app, with no network hop, against the other engine's own embedded mode where it has one, otherwise its fastest connection, named on each bar.
As a server (the right column): Semurg over its fastest connection against the other engine's fastest connection.

Two ways to run Semurg, both first-class. Inside your application (the left column): Semurg runs in your application's own process, with no network hop between them. As a server (the right column): Semurg runs as its own service and your application reaches it over the network; several servers form a cluster. In both, a write is acknowledged only after the drive has completed it, so if your application or Semurg is stopped or killed, nothing it acknowledged is lost. On a single node, a sudden power cut or an operating-system crash can lose the most recent writes (the last few seconds); network durability starts at the second node.

Bigger than RAM: each domain will be measured again on a dataset at least twice the size of the machine's memory, where every engine has to read from disk; these runs are scheduled and their results appear here as they complete.

One Cherry Servers box: AMD EPYC 9354, 32 cores / 64 threads, 377 GB RAM, NVMe; CPU only, no GPU
Both engines in Docker on the same box
Each engine in its fastest query form, sized by its own tooling
Answers checked identical before anything is timed
The other engines' many-at-once clients run as separate processes; the count is in each bar's detail
Semurg r11.54 to r11.71, measured 29 September to 7 October 2026

Domains

Graph

pokec social graph, 1.6M people, 30.6M relationships. Every engine returns the same 176,181 people at 4 hops.

Some comparisons are still being measured; their results appear here as they complete.

In-memory dataset

Inside your application

Engines with no in-application mode are shown through their own fastest connection, which favours Semurg.

Inside your applicationOne query, 4 hops measured on r11.59, carriedmedian time per query

Inside your application64 queries at once, 4 hops measured on r11.59, carriedqueries per second

Inside your applicationSemurg vs Neo4j: one query, by depth measured on r11.59, carriedmedian time per query; Neo4j at its best measured figure

Inside your applicationSemurg vs Neo4j: 64 queries at once, by depth measured on r11.59, carriedqueries per second; Neo4j at its best measured figure

Inside your applicationWalk to 131,072 hops, total time measured on r11.54, carried17 steps, each checked against the reference answer

As a server

As a serverOne query, 4 hops measured on r11.59, carriedmedian time per query

Neo4j, ArcadeDB, SurrealDB, De Wit Graph Database A and De Wit Graph Database B: Results pending

Engines named 'De Wit … Database A' or 'B': some database vendors' licences forbid publishing benchmark results without their permission (a 'DeWitt clause', common in the industry since the 1980s). We run those engines exactly like every other engine, on the same machine and the same questions, and show their results unranked and without their names, to respect those terms.

As a server64 queries at once, 4 hops measured on r11.59, carriedqueries per second

Neo4j, ArcadeDB, SurrealDB, De Wit Graph Database A and De Wit Graph Database B: Results pending

As a serverSemurg vs Neo4j: one query, by depth

Results pending

As a serverSemurg vs Neo4j: 64 queries at once, by depth

Results pending

As a serverWalk to 131,072 hops, total time measured on r11.54, carried17 steps, each checked against the reference answer

Neo4j: Results pending

Bigger-than-RAM dataset

Results pending

The numbers behind the graph charts

Chart	Mode	Comparison	Semurg	Other engine	Result
One query, 4 hops	Inside your application	Semurg vs Neo4j	2.54 ms	52.1 ms	Semurg 20.5 times faster
One query, 4 hops	Inside your application	Semurg vs FalkorDB	2.54 ms	97.4 ms	Semurg 38.4 times faster
One query, 4 hops	Inside your application	Semurg vs ArangoDB	2.54 ms	99.3 ms	Semurg 39.1 times faster
One query, 4 hops	As a server	Semurg vs Neo4j		Results pending	Results pending
One query, 4 hops	As a server	Semurg vs FalkorDB	2.77 ms	97.4 ms	Semurg 35.2 times faster
One query, 4 hops	As a server	Semurg vs ArangoDB	2.77 ms	99.3 ms	Semurg 35.9 times faster
One query, 4 hops	As a server	Semurg vs ArcadeDB		Results pending	Results pending
One query, 4 hops	As a server	Semurg vs SurrealDB		Results pending	Results pending
One query, 4 hops	As a server	Semurg vs De Wit Graph Database A		Results pending	Results pending
One query, 4 hops	As a server	Semurg vs De Wit Graph Database B		Results pending	Results pending
64 queries at once, 4 hops	Inside your application	Semurg vs Neo4j	1,574 queries/s	183 queries/s	Semurg 8.61 times faster
64 queries at once, 4 hops	Inside your application	Semurg vs ArangoDB	1,574 queries/s	72.5 queries/s	Semurg 21.7 times faster
64 queries at once, 4 hops	Inside your application	Semurg vs FalkorDB	1,574 queries/s	44.6 queries/s	Semurg 35.3 times faster
64 queries at once, 4 hops	As a server	Semurg vs Neo4j		Results pending	Results pending
64 queries at once, 4 hops	As a server	Semurg vs ArangoDB	1,476 queries/s	72.5 queries/s	Semurg 20.4 times faster
64 queries at once, 4 hops	As a server	Semurg vs FalkorDB	1,476 queries/s	44.6 queries/s	Semurg 33.1 times faster
64 queries at once, 4 hops	As a server	Semurg vs ArcadeDB		Results pending	Results pending
64 queries at once, 4 hops	As a server	Semurg vs SurrealDB		Results pending	Results pending
64 queries at once, 4 hops	As a server	Semurg vs De Wit Graph Database A		Results pending	Results pending
64 queries at once, 4 hops	As a server	Semurg vs De Wit Graph Database B		Results pending	Results pending
Semurg vs Neo4j: one query, by depth	Inside your application	2 hops	0.42 ms	0.46 ms	Semurg 1.09 times faster
Semurg vs Neo4j: one query, by depth	Inside your application	4 hops	2.54 ms	52.1 ms	Semurg 20.5 times faster
Semurg vs Neo4j: one query, by depth	Inside your application	8 hops	50.6 ms	5,593 ms	Semurg 110 times faster
Semurg vs Neo4j: one query, by depth	Inside your application	16 hops	58.1 ms	5,632 ms	Semurg 96.9 times faster
Semurg vs Neo4j: one query, by depth	Inside your application	32 hops	59.2 ms	5,570 ms	Semurg 94 times faster
Semurg vs Neo4j: one query, by depth	Inside your application	64 hops	58.5 ms	5,600 ms	Semurg 95.8 times faster
Semurg vs Neo4j: one query, by depth	Inside your application	128 hops	57.7 ms	5,529 ms	Semurg 95.8 times faster
Semurg vs Neo4j: one query, by depth	Inside your application	256 hops	57.9 ms	5,454 ms	Semurg 94.2 times faster
Semurg vs Neo4j: one query, by depth	Inside your application	512 hops	56.9 ms	5,483 ms	Semurg 96.3 times faster
Semurg vs Neo4j: one query, by depth	Inside your application	1024 hops	51.7 ms	5,466 ms	Semurg 106 times faster
Semurg vs Neo4j: one query, by depth	As a server	2 hops		Results pending	Results pending
Semurg vs Neo4j: one query, by depth	As a server	4 hops		Results pending	Results pending
Semurg vs Neo4j: one query, by depth	As a server	8 hops		Results pending	Results pending
Semurg vs Neo4j: one query, by depth	As a server	16 hops		Results pending	Results pending
Semurg vs Neo4j: one query, by depth	As a server	32 hops		Results pending	Results pending
Semurg vs Neo4j: one query, by depth	As a server	64 hops		Results pending	Results pending
Semurg vs Neo4j: one query, by depth	As a server	128 hops		Results pending	Results pending
Semurg vs Neo4j: one query, by depth	As a server	256 hops		Results pending	Results pending
Semurg vs Neo4j: one query, by depth	As a server	512 hops		Results pending	Results pending
Semurg vs Neo4j: one query, by depth	As a server	1024 hops		Results pending	Results pending
Semurg vs Neo4j: 64 queries at once, by depth	Inside your application	2 hops	53,159 queries/s	166,536 queries/s	Semurg 3.13 times slower
Semurg vs Neo4j: 64 queries at once, by depth	Inside your application	4 hops	1,574 queries/s	183 queries/s	Semurg 8.61 times faster
Semurg vs Neo4j: 64 queries at once, by depth	Inside your application	8 hops	62.5 queries/s	7.15 queries/s	Semurg 8.74 times faster
Semurg vs Neo4j: 64 queries at once, by depth	Inside your application	16 hops	60.9 queries/s	6.82 queries/s	Semurg 8.93 times faster
Semurg vs Neo4j: 64 queries at once, by depth	Inside your application	32 hops	60.9 queries/s	7 queries/s	Semurg 8.7 times faster
Semurg vs Neo4j: 64 queries at once, by depth	Inside your application	64 hops	61 queries/s	6.94 queries/s	Semurg 8.79 times faster
Semurg vs Neo4j: 64 queries at once, by depth	Inside your application	128 hops	61.6 queries/s	7.59 queries/s	Semurg 8.12 times faster
Semurg vs Neo4j: 64 queries at once, by depth	Inside your application	256 hops	61.6 queries/s	7.17 queries/s	Semurg 8.59 times faster
Semurg vs Neo4j: 64 queries at once, by depth	Inside your application	512 hops	61.7 queries/s	6.72 queries/s	Semurg 9.18 times faster
Semurg vs Neo4j: 64 queries at once, by depth	Inside your application	1024 hops	61.6 queries/s	6.63 queries/s	Semurg 9.29 times faster
Semurg vs Neo4j: 64 queries at once, by depth	As a server	2 hops		Results pending	Results pending
Semurg vs Neo4j: 64 queries at once, by depth	As a server	4 hops		Results pending	Results pending
Semurg vs Neo4j: 64 queries at once, by depth	As a server	8 hops		Results pending	Results pending
Semurg vs Neo4j: 64 queries at once, by depth	As a server	16 hops		Results pending	Results pending
Semurg vs Neo4j: 64 queries at once, by depth	As a server	32 hops		Results pending	Results pending
Semurg vs Neo4j: 64 queries at once, by depth	As a server	64 hops		Results pending	Results pending
Semurg vs Neo4j: 64 queries at once, by depth	As a server	128 hops		Results pending	Results pending
Semurg vs Neo4j: 64 queries at once, by depth	As a server	256 hops		Results pending	Results pending
Semurg vs Neo4j: 64 queries at once, by depth	As a server	512 hops		Results pending	Results pending
Semurg vs Neo4j: 64 queries at once, by depth	As a server	1024 hops		Results pending	Results pending
Walk to 131,072 hops, total time	Inside your application	Semurg vs FalkorDB	1,052 ms	4,673 ms	Semurg 4.44 times faster
Walk to 131,072 hops, total time	Inside your application	Semurg vs Neo4j	1,052 ms	83,954 ms	Semurg 79.8 times faster
Walk to 131,072 hops, total time	Inside your application	Semurg vs ArangoDB	1,052 ms	278,200 ms	Semurg 264 times faster
Walk to 131,072 hops, total time	As a server	Semurg vs FalkorDB	988 ms	4,673 ms	Semurg 4.73 times faster
Walk to 131,072 hops, total time	As a server	Semurg vs Neo4j		Results pending	Results pending
Walk to 131,072 hops, total time	As a server	Semurg vs ArangoDB	988 ms	278,200 ms	Semurg 282 times faster

Vector

SIFT1M, 1M vectors of 128 dimensions. The question: the exact 10 nearest vectors to a query. Every engine returns the same answer as the exact oracle.

Results pending

Relational

TPC-C at 150 warehouses: nine tables, 74,959,250 rows. The question: read 256 rows of the item table by primary key in one request.

Results pending

Columnar

The dataset and the question are published with its first gated run.

Results pending

Key-value

9,000,000 key-value pairs; every engine stores each key with its whole value. The question: read one key's value. Every engine returns the same value.

Some comparisons are still being measured; their results appear here as they complete.

In-memory dataset

Inside your application

Engines with no in-application mode are shown through their own fastest connection, which favours Semurg.

Inside your applicationOne read (point get) measured on r11.71median time per read

LMDB: Results pending

Inside your application64 reads at once measured on r11.71reads per second; every other engine on all 64 threads

Redis and Valkey: Results pending

As a server

Engines with no server mode are shown in-process, which favours them.

As a serverOne read (point get) measured on r11.71median time per read

RocksDB and LMDB: Results pending

As a server64 reads at once measured on r11.71reads per second; every other engine on all 64 threads

Redis and Valkey: Results pending

Bigger-than-RAM dataset

Results pending

The numbers behind the key-value charts

Chart	Mode	Comparison	Semurg	Other engine	Result
One read (point get)	Inside your application	Semurg vs Redis	0.039 ms	0.063 ms	Semurg 1.62 times faster
One read (point get)	Inside your application	Semurg vs Valkey	0.039 ms	0.056 ms	Semurg 1.44 times faster
One read (point get)	Inside your application	Semurg vs RocksDB	0.039 ms	0.002 ms	Semurg 19.5 times slower
One read (point get)	Inside your application	Semurg vs LMDB		Results pending	Results pending
One read (point get)	As a server	Semurg vs Redis	0.16 ms	0.063 ms	Semurg 2.54 times slower
One read (point get)	As a server	Semurg vs Valkey	0.16 ms	0.056 ms	Semurg 2.86 times slower
One read (point get)	As a server	Semurg vs RocksDB		Results pending	Results pending
One read (point get)	As a server	Semurg vs LMDB		Results pending	Results pending
64 reads at once	Inside your application	Semurg vs Redis		Results pending	Results pending
64 reads at once	Inside your application	Semurg vs Valkey		Results pending	Results pending
64 reads at once	Inside your application	Semurg vs RocksDB	64,908 reads/s	93,553 reads/s	Semurg 1.44 times slower
64 reads at once	Inside your application	Semurg vs LMDB	64,908 reads/s	548,843 reads/s	Semurg 8.46 times slower
64 reads at once	As a server	Semurg vs Redis		Results pending	Results pending
64 reads at once	As a server	Semurg vs Valkey		Results pending	Results pending
64 reads at once	As a server	Semurg vs RocksDB	6,556 reads/s	93,553 reads/s	Semurg 14.3 times slower
64 reads at once	As a server	Semurg vs LMDB	6,556 reads/s	548,843 reads/s	Semurg 83.7 times slower

Time-series

The dataset and the question are published with its first gated run.

Results pending

Object

1,000 images from Wikimedia Commons (a subset listed by the 2026-09-01 Commons image dump), 140,275,854 bytes in total, 467 B to 262,100 B each, one object per file, keyed by its path. The question: get one object by its key; every engine returns the same bytes.

Some comparisons are still being measured; their results appear here as they complete. Ceph RGW: not measured.

In-memory dataset

Inside your application

Engines with no in-application mode are shown through their own fastest connection, which favours Semurg.

Inside your applicationOne read by key measured on r11.71median time per read

Inside your application64 reads at once measured on r11.71reads per second; every other engine on all 64 threads

Inside your applicationLoad 1,000 objects measured on r11.71time to store the whole dataset, median of 3 loads

Inside your applicationList objects by prefix

Results pending

Inside your applicationRead one object as byte ranges

Results pending

As a server

As a serverOne read by key measured on r11.71median time per read

As a server64 reads at once measured on r11.71reads per second; every other engine on all 64 threads

As a serverLoad 1,000 objects measured on r11.71time to store the whole dataset, median of 3 loads

As a serverList objects by prefix

Results pending

As a serverRead one object as byte ranges

Results pending

Bigger-than-RAM dataset

Results pending

The numbers behind the object charts

Chart	Mode	Comparison	Semurg	Other engine	Result
One read by key	Inside your application	Semurg vs SeaweedFS	0.1 ms	0.52 ms	Semurg 5.2 times faster
One read by key	As a server	Semurg vs SeaweedFS	0.69 ms	0.52 ms	a tie
64 reads at once	Inside your application	Semurg vs SeaweedFS	75,347 reads/s	54,099 reads/s	Semurg 1.39 times faster
64 reads at once	As a server	Semurg vs SeaweedFS	29,165 reads/s	54,099 reads/s	Semurg 1.85 times slower
Load 1,000 objects	Inside your application	Semurg vs SeaweedFS	934 ms	5,439 ms	Semurg 5.82 times faster
Load 1,000 objects	As a server	Semurg vs SeaweedFS	929 ms	5,439 ms	Semurg 5.85 times faster

Document

GH Archive events: 2,318,659 JSON documents (8.6 GB). The question: read one document by its id. Every engine returns the same document.

Some comparisons are still being measured; their results appear here as they complete.

In-memory dataset

Inside your application

Engines with no in-application mode are shown through their own fastest connection, which favours Semurg.

Inside your applicationOne read by id measured on r11.69median time per read

Inside your application64 reads at once measured on r11.69reads per second; every other engine on all 64 threads

Inside your applicationLoad 2,318,659 documents

Results pending

As a server

As a serverOne read by id measured on r11.69median time per read

As a server64 reads at once measured on r11.69reads per second; every other engine on all 64 threads

As a serverLoad 2,318,659 documents

Results pending

Bigger-than-RAM dataset

Results pending

The numbers behind the document charts

Chart	Mode	Comparison	Semurg	Other engine	Result
One read by id	Inside your application	Semurg vs MongoDB	0.34 ms	0.47 ms	Semurg 1.4 times faster
One read by id	As a server	Semurg vs MongoDB	3.91 ms	0.47 ms	Semurg 8.33 times slower
64 reads at once	Inside your application	Semurg vs MongoDB	28,423 reads/s	79,506 reads/s	Semurg 2.8 times slower
64 reads at once	As a server	Semurg vs MongoDB	26,889 reads/s	79,506 reads/s	Semurg 2.96 times slower
Load 2,318,659 documents	Inside your application	Semurg vs MongoDB		Results pending	Results pending
Load 2,318,659 documents	As a server	Semurg vs MongoDB		Results pending	Results pending

Search

English Wikipedia full text (3.9 GB). The question: every article whose body contains the word "submarine". Every engine returns the same 7,195 articles.

Some comparisons are still being measured; their results appear here as they complete.

In-memory dataset

Inside your application

Engines with no in-application mode are shown through their own fastest connection, which favours Semurg.

Inside your applicationOne search measured on r11.68median time per search

De Wit Search Database A: not published (licence)

Inside your application64 searches at once

Results pending

As a server

As a serverOne search

Results pending

As a server64 searches at once

Results pending

Bigger-than-RAM dataset

Results pending

The numbers behind the search charts

Chart	Mode	Comparison	Semurg	Other engine	Result
One search	Inside your application	Semurg vs OpenSearch	0.18 ms	0.64 ms	Semurg 3.48 times faster
One search	Inside your application	Semurg vs Typesense	0.18 ms	2.04 ms	Semurg 11.2 times faster
One search	Inside your application	Semurg vs De Wit Search Database A	0.18 ms	not published (licence)	pending
One search	As a server	Semurg vs OpenSearch		Results pending	Results pending
One search	As a server	Semurg vs Typesense		Results pending	Results pending
One search	As a server	Semurg vs De Wit Search Database A		Results pending	Results pending
64 searches at once	Inside your application	Semurg vs OpenSearch		Results pending	Results pending
64 searches at once	Inside your application	Semurg vs Typesense		Results pending	Results pending
64 searches at once	Inside your application	Semurg vs De Wit Search Database A		not published (licence)	pending
64 searches at once	As a server	Semurg vs OpenSearch		Results pending	Results pending
64 searches at once	As a server	Semurg vs Typesense		Results pending	Results pending
64 searches at once	As a server	Semurg vs De Wit Search Database A		Results pending	Results pending

Streaming

The dataset and the question are published with its first gated run.

Results pending

Universal

Universal (multi-model)

The dataset and the question are published with its first gated run.

Results pending

Converged

The dataset and the question are published with its first gated run.

Results pending

Functionality

Vector + graph RAG

pokec social graph (30.6M relationships) with SIFT1M's 1M vectors attached to its first 1M people: a pairing of two public datasets. The question: the 10 nearest vectors to a query (exact), then everyone within 2 hops of those 10 people. Every engine returns the same people.

Results pending

Graph RAG

The dataset and the question are published with its first gated run.

Results pending

Knowledge graph

HotpotQA's knowledge graph: 292,851 passage-to-passage links (3.4 MB). The question: one HotpotQA question's seed passages, then every distinct passage within 2 hops (10 questions; question 1 reaches 92 passages). Every engine returns the same passages.

Some comparisons are still being measured; their results appear here as they complete.

In-memory dataset

Inside your application

Engines with no in-application mode are shown through their own fastest connection, which favours Semurg.

Inside your applicationOne question, 2 hops measured on r11.71median time per question

Neo4j: not measured

Inside your application64 questions at once, 2 hops measured on r11.71questions per second; every other engine on all 64 threads

Neo4j: not measured

Inside your applicationLoad the knowledge graph

Results pending

As a server

As a serverOne question, 2 hops measured on r11.71median time per question

Neo4j: Results pending

As a server64 questions at once, 2 hops measured on r11.71questions per second; every other engine on all 64 threads

Neo4j: Results pending

As a serverLoad the knowledge graph

Results pending

Bigger-than-RAM dataset

Results pending

The numbers behind the knowledge graph charts

Chart	Mode	Comparison	Semurg	Other engine	Result
One question, 2 hops	Inside your application	Semurg vs Neo4j	0.13 ms	not measured	not measured
One question, 2 hops	Inside your application	Semurg vs Postgres	0.13 ms	0.36 ms	Semurg 2.71 times faster
One question, 2 hops	As a server	Semurg vs Neo4j		Results pending	Results pending
One question, 2 hops	As a server	Semurg vs Postgres	0.19 ms	0.36 ms	Semurg 1.9 times faster
64 questions at once, 2 hops	Inside your application	Semurg vs Neo4j	335,638 questions/s	not measured	not measured
64 questions at once, 2 hops	Inside your application	Semurg vs Postgres	335,638 questions/s	95,130 questions/s	Semurg 3.53 times faster
64 questions at once, 2 hops	As a server	Semurg vs Neo4j		Results pending	Results pending
64 questions at once, 2 hops	As a server	Semurg vs Postgres	119,840 questions/s	95,130 questions/s	Semurg 1.26 times faster
Load the knowledge graph	Inside your application	Semurg vs Neo4j		not measured	not measured
Load the knowledge graph	Inside your application	Semurg vs Postgres		Results pending	Results pending
Load the knowledge graph	As a server	Semurg vs Neo4j		Results pending	Results pending
Load the knowledge graph	As a server	Semurg vs Postgres		Results pending	Results pending

How a lane is run

The answer check comes before the stopwatch. A faster wrong answer is not a result.

Run it yourself

curl -fsSL https://one.semurg.io/bench | sudo bash

The box: Ubuntu 24.04 on x86-64, root, and a data disk separate from the OS disk, mounted, with about 1 TB free. Dedicate it to this.
What the command does: downloads the benchmark runner from one.semurg.io, checks it against the checksum published with it, finds your data disks, installs Docker and Elixir if they are missing, runs graph and vector against the same engines as the charts above, and prints the table.
If it refuses: every refusal says what to do next and leaves the box unchanged.
Time: hours, mostly downloading and loading the datasets, which come from their original public sources and are checksummed when they arrive.
Your release: each chart above names the Semurg release it was measured on; your run measures the release you install, on your hardware, as its own series.
One domain only: curl -fsSL https://one.semurg.io/bench | sudo bash -s -- --graph
Read the script first: curl -fsSL https://one.semurg.io/bench | less

The four states a row can have

Measured

Both figures, the release, both sides' utilisation, and the command that reproduces it.

Results pending

The lane runs; its figures are not yet measured on current hardware.

Not published

That engine's licence forbids publishing benchmark results, so we publish none. No name, no figure.

Not run

No independent reference answer exists for this test yet, so a number could not be checked.

A loss is drawn at the same size and weight as a win.