Carve / benchmarks

Recorded performance evidence

How fast does Carve render?

Measured engine snapshots on shared hardware. Compare direct conversion, full parsing of authored documents, and the stress corpus.

2026-10-21T03:24:46.274Z; serial processes; Node v22.22.2; PHP 8.5.11 (cli) (built: Sep 24 2026 13:49:29) (NTS); rustc 1.97.1 (8bab26f4f 2026-07-14).

Track A

Direct HTML conversion on core examples

Default public conversion APIs, without opt-in extensions. Peers use equivalent logical content in their native syntax. Features and output differ; these rows measure rendering cost.

These core examples use the direct HTML route in all three engines. The authored-document track below measures full parse (AST) on public documentation files. Both tracks use the default conversion APIs; document features decide the route.

LanguageCarveCarve MB/sFastest peerPeer MB/s
JavaScriptcarve-js8.06markdown-it4.27
PHPcarve-php17.62djot-php17.58
Rustcarve-rs110.77pulldown-cmark99.65

The archived Rust wall-time snapshot came from a loaded two-round run (one-minute host load 46.54). An eight-round recheck with the same binary found carve-rs faster in 6 of 8 rounds, at 0.716 versus 0.866 process CPU ms/op. The direct-HTML chart uses process CPU time from the verified 24-round rerun for these two Rust rows. First recheck.

The displayed 24-round paired run of that binary found carve-rs faster in 22 of 24 rounds. Pooled process CPU time was 0.539 ms/op for carve-rs and 0.605 ms/op for pulldown-cmark. This measures the two Rust libraries on the core fixture; the shared host remained busy. Rerun method · Raw rounds.

The PHP peer changed route on 2 October: fab953f6 rejected this fixture in its borrowed HTML renderer, while 9c52e611 accepted it. The benchmark updated its Djot pin that day. The old and new Carve rows use the borrowed route, so rankings across those snapshots include a peer route change.

The displayed PHP pair comes from a 32-round process-CPU rerun of merged carve-php 760df44b and the locked djot-php revision. Carve used less CPU in 15 of 32 rounds. Pooled medians were 3.391 and 3.405 ms/op. One-minute load at round starts stayed between 6.88 and 8.11. Method and load range · Raw rounds.

CSV with pipe tablesSnapshot JSON

With pipe tables · 18 points (2026-10-21 snapshot with verified pair reruns)

Median process CPU time; 24 paired Rust rounds, 32 paired PHP rounds, two recorded rounds for JavaScript and other peers

  • carve-rs (Rust, Carve)110.77 MB/s
  • pulldown-cmark (Rust, Markdown)99.65 MB/s
  • jotdown (Rust, Djot)24.28 MB/s
  • comrak (Rust, Markdown)19.08 MB/s
  • carve-php (PHP, Carve)17.62 MB/s
  • djot-php (PHP, Djot)17.58 MB/s
  • carve-js (JS, Carve)8.06 MB/s
  • markdown-it (JS, Markdown)4.27 MB/s
  • djot.js (JS, Djot)3.34 MB/s
  • league/commonmark-gfm (PHP, Markdown)0.99 MB/s

JavaScript without pipe tables · 14 points (2026-11-12, Carve main efef22a)

Median timing across all samples from both measurement rounds

  • carve-js (JS, Carve)16.88 MB/s
  • commonmark.js (JS, Markdown)9.22 MB/s
  • markdown-it (JS, Markdown)4.90 MB/s
  • djot.js (JS, Djot)3.58 MB/s
Core conversion throughput with and without pipe tables, in separate panels
Download SVG
Carve core throughput with pipe tables
Download SVG

JavaScript

EngineMB/s
carve-js8.06
djot.js3.34
markdown-it4.27

PHP

EngineMB/s
carve-php17.62
djot-php17.58
league/commonmark-gfm0.99

Rust

EngineMB/s
carve-rs110.77
jotdown24.28
comrak19.08
pulldown-cmark99.65

Historical release comparison and capability scoring · Current dev-main measurement report

JavaScript without pipe tables

Carve, Djot, markdown-it and commonmark.js on equivalent content. This 14-point workload excludes pipe tables and is separate from the comparison with pipe tables above. Final values use median timing across all samples from both measurement rounds. Carve JS uses merged main efef22a, with fast-path use verified. Peers use the released packages named in the report.

EngineMedian ms/opMB/s
carve-js2.809616.88
djot.js13.23803.58
markdown-it9.79764.90
commonmark.js5.20379.22
Final JavaScript throughput without pipe tables
Download SVG

Method, individual rounds and measurement provenance · Raw samples and controls

Paired full-feature comparison

Before and after commits render identical output within each engine. Both use the same CPU settings, with alternating run order. This workload has closed blocks and exercises formatting, quotations, lists, tables, raw HTML and footnotes. Changes describe the observed per-round ratios, not statistical significance. Positive means more elapsed time; negative means less. The range shows variation across rounds. The median paired ratio can differ from the ratio of the two median times.

EngineSectionsMedian before msMedian after msMedian paired changePaired range
js12852.777956.3681+11.2%+3.0% to +15.7%
js512205.7019225.3598+7.6%+4.1% to +27.7%
php128119.2356129.5381+14.7%-0.8% to +24.6%
php512490.3246563.3013+11.7%+7.3% to +21.1%
rs1288.66488.0922-5.7%-49.0% to +2.3%
rs51241.437439.1201-6.5%-14.0% to +192.7%

Exact commits and method · Output controls, samples and provenance · Exact-corpus revision checks · Historical paired comparison · Historical Rust ranking check

Track B

Concatenated corpus stress test and extension tiers

This corpus joins independent conformance examples into one document. Open fences and containers can interact, so correctness fixes may change the output and nesting. It exercises the normal parser and public AST and is separate from the paired comparison above. This stress input is designed to expose parser depth and feature interactions; the authored-document track gives complete prose examples. Competitor parsers do not accept equivalent syntax, so this track compares Carve implementations and internal PHP tiers. These are shared-host wall-time snapshots; use paired runs to assess changes between revisions.

Check the output byte counts and hashes before treating cross-engine times as equal work.

Fixed committed corpus; input hashes are recorded in the full-run JSON.

The PHP tier table and chart use 15 alternating rounds in one PHP process at the current PHP commit. Tier 1 and Tier 2 render identical HTML on this input. Tier 3 changes output. The earlier separate-process measurements remain available.

Throughput of the three Carve engines for each corpus size
Download SVG

small (1.2 KB)

Enginems/opMB/srel
carve-js2.16690.5415.86x
carve-php2.96670.4021.72x
carve-rs0.13668.611.00x

medium (63.6 KB)

Enginems/opMB/srel
carve-js244.71020.253.90x
carve-php346.79120.185.52x
carve-rs62.79910.991.00x

large (508.6 KB)

Enginems/opMB/srel
carve-js1696.68310.293.53x
carve-php2622.74450.195.45x
carve-rs481.01561.031.00x

PHP authoritative extension tiers

ProfileRegistered extensionsms/opMB/scost vs Tier 1round range, ms
Tier 1 core/default06.537.19baseline3.77 to 7.53
Tier 2 stack87.226.50+11%4.33 to 10.17
Tier 3 stack208.555.49+31%5.29 to 13.92
PHP throughput with core, Tier 2, and Tier 3 extension profiles
Download SVG

Full corpus report · Worker results and provenance

Track C

Structured document conversion (AST)

Three public Carve documentation files were adapted to Carve once and saved as fixed inputs. All three pinned engines took the full parse route for every complete file. JavaScript and Rust produced identical HTML bytes; PHP added one final newline. This measures real prose, examples, links, and document structure through the default source-to-HTML API. The fixed documents show conversion cost for authored prose and structure; they do not measure how often users take this route.

DocumentSizeJavaScript CPU ms/opPHP CPU ms/opRust CPU ms/op
guide4.9 KiB3.014.010.191
migration21.4 KiB10.9716.831.41
contract96.2 KiB62.61108.9013.18
Median process CPU time on three adapted documentation files, with a logarithmic axis
Download SVG
Round median ranges
DocumentEngineRange, CPU ms/op
guideJavaScript1.94–3.11
guidePHP3.66–5.40
guideRust0.184–0.355
migrationJavaScript7.48–12.53
migrationPHP11.85–18.45
migrationRust1.18–1.76
contractJavaScript51.76–84.63
contractPHP90.99–142.51
contractRust12.71–18.98

Measured 2026-10-21. Lower process CPU milliseconds are faster. The chart has a logarithmic axis; the table gives exact medians. Elapsed wall samples remain in the raw data. 4 rotating rounds used 11 warmed samples each. Method and pinned commits · Raw samples · Output controls · Fixture provenance

Track D

Authored documents against peer libraries

Three public documentation files were converted to native Carve, Djot, and Markdown inputs. All three Carve engines took the AST route through their default APIs. All eleven libraries matched the same projected HTML. The djot-php code-span discrepancy in the earlier local run was fixed by PR #325. The chart shows median within-round speed relative to Carve for each runtime and document. Carve is 1×; higher means the peer used less process CPU time. Whiskers show the round range. JavaScript timing starts after 200 warmup conversions per worker.

Paired process CPU speed relative to Carve across eleven libraries on three authored documents
Download SVG
Absolute CPU time and paired ratios
DocumentLibraryCPU ms/opSpeed vs Carve
termscarve-js1.1151.00×
termsdjot.js0.5432.07×
termsmarkdown-it0.5402.06×
termscommonmark.js0.2483.94×
termscarve-php2.3151.00×
termsdjot-php1.0862.00×
termsleague/commonmark-gfm2.1791.09×
termscarve-rs0.1501.00×
termsjotdown0.0582.60×
termscomrak0.0532.92×
termspulldown-cmark0.0227.07×
case-study-parsing-astcarve-js3.7541.00×
case-study-parsing-astdjot.js1.4622.68×
case-study-parsing-astmarkdown-it0.5966.63×
case-study-parsing-astcommonmark.js0.37610.95×
case-study-parsing-astcarve-php2.8271.00×
case-study-parsing-astdjot-php1.3171.95×
case-study-parsing-astleague/commonmark-gfm2.9460.93×
case-study-parsing-astcarve-rs0.2201.00×
case-study-parsing-astjotdown0.0902.53×
case-study-parsing-astcomrak0.0603.70×
case-study-parsing-astpulldown-cmark0.0259.13×
spec-historycarve-js2.4661.00×
spec-historydjot.js1.3471.87×
spec-historymarkdown-it1.5821.50×
spec-historycommonmark.js0.5494.57×
spec-historycarve-php6.4151.00×
spec-historydjot-php3.3541.88×
spec-historyleague/commonmark-gfm5.7961.11×
spec-historycarve-rs0.9061.00×
spec-historyjotdown0.2763.26×
spec-historycomrak0.1994.47×
spec-historypulldown-cmark0.08310.96×

Measured 2026-10-24 in 16 rotating rounds on a shared host. The libraries support different grammar and build different internal representations. Carve report · Raw samples and output controls · AST route checks · Fixture provenance.

Engine release history

Release-tag rows retain their 2026-10-02 sessions. Merged main points were refreshed 2026-10-20 at pinned commits. Compare elapsed time within each engine; graphs mark changed output. Separate sessions do not isolate code speedups.

The broad JavaScript rise against 0.1.9 did not repeat across ordinary cases in a fresh paired run. PHP's quoted-fence rise did repeat; merged PR #3595 and PR #3598 reduced it. The pinned Rust recheck was faster than 0.1.6 on quoted fences and mixed documents; the latest 0.1.10 tag is close. The paired history check shows the sources and samples; the dated structured-document run is measured separately.

Explore the dated version history · History CSV · Samples and source commits

js elapsed time across release tags and pinned merged main
Download js history SVG
php elapsed time across release tags and pinned merged main
Download php history SVG
rs elapsed time across release tags and pinned merged main
Download rs history SVG

Read the measurements

Method and source commits

Higher MB/s is better. The direct HTML chart uses median process CPU time: 24 paired rounds for the two leading Rust libraries and two recorded rounds for the other engines. The separate JavaScript chart without pipe tables uses wall time. Full-corpus rows average many in-process iterations. Each track has its own input mix and route; compare figures within that workload.

These are machine-specific snapshots. Shared host activity affects timings; controlled paired runs are needed to establish improvements or regressions.

Core comparison host

2026-10-21T02:51:04.375Z; AMD Ryzen 9 PRO 7940HS w/ Radeon 780M Graphics; Node v22.22.2; PHP PHP 8.5.11 (cli) (built: Sep 24 2026 13:49:29) (NTS); rustc 1.97.1 (8bab26f4f 2026-07-14); Allowed CPUs: 12. Child processes inherit this affinity, including Node compiler and GC threads.; PHP pair completed 2026-10-21T21:28:55Z

Core peer versions

djot.js 0.3.2; markdown-it 15.0.0; Djot PHP 77e5b6c83978b28b72f314fe52c0a200856ad3f9; league/commonmark 2.10.0. Carve source commits: js 2ce388c2897f2a93e13de8832b3d1c3a5788735c; php 760df44b19013a6b1ae07f7a2cb71da08da84d0c; rs 47fdd7e487fc793133cf6af8396f653975454dd9. Rust peer versions are recorded in the downloadable Cargo lock.

Full-corpus engines measured

carve-js @markup-carve/carve 0.1.12 (local checkout carve-js-main @ 2ce388c28), carve-php markup-carve/carve-php (local checkout carve-php-main @ 2a7a38145), carve-rs carve-lang 0.1.10 (local checkout carve-rs-main @ 47fdd7e48)

Site source: ff8815e3d44eb8ed3596de01cfe33dca39dc67c3.

Reproduce these runs · Feature scoring · HTML import comparison

Download reports

Current merged core peers · Main versus retained tags · Released core comparison · Full corpus · Reproduction guide · Feature scoring · Historical findings · All site data