Blacksmith Benchmark Snapshot

This page is generated from the Tool Benchmark workflow so published performance numbers can cite one reproducible runner, input corpus, and commit.

Latest Result

Measured: 2026-07-30T15:44:33.256Z Commit: 1511788d96ea (run) Runner: blacksmith-32vcpu-ubuntu-2404 (32 logical CPU, AMD EPYC, 32 vCPU / 128 GB RAM / 1.5 TB storage) Input: 15,000 generated SFC files (58.7 MB). Median of 5 measured run(s) after 1 warmup run(s). Versions: vize vize 0.303.0 · tsgo Version 7.0.0-dev.20260602.1 · vue-tsc Version 6.0.3 (typescript n/a) · vue 3.6.0-beta.10 · eslint v10.4.1 · prettier 3.8.3 · node v24.14.0 Binaries (sha256): Backend: native TypeScript engine ready at /home/runner/_work/vize/vize/node_modules/.bin/tsgo. Planted-diagnostic gating for the type-check rows lives in bench/check-gate.mjs (.github/workflows/check-bench.yml). Large SFC: 900 repeated template blocks (674.9 KB). Nuxt import set: 500 SFC files.

Large SFC type check — engine classes ranked separately

No cross-class ratio is published for Large SFC type check: the incumbent runs the JavaScript TypeScript compiler while Vize runs native tsgo, so a single number would credit TypeScript's Go rewrite to the Vue layer.

Type check — engine classes ranked separately

No cross-class ratio is published for Type check: the incumbent runs the JavaScript TypeScript compiler while Vize runs native tsgo, so a single number would credit TypeScript's Go rewrite to the Vue layer.

Fairness notes:

• All tools run on the same generated Vue SFC corpus from the same checkout and lockfile.
• The 15,000-SFC rows are the many-file workload; the large-SFC rows isolate one large component.
• Reported times are medians; measured runs alternate variant order after warmup runs.
• Destructive formatter runs receive a fresh copy of the same input before each invocation.
• SFC compile Vize max uses compileSfcBatchWithResults wall time so the primary number includes generated output crossing the JS/native boundary; the stats-only native timeMs is shown only in variant details. Explicit sequence variants run 1→max and max→1 in one Node process and measure the second call's full JavaScript wall time, including scoped pool creation.
• Vite build timings exclude fixture copy/setup; the Vize max lane sets precompileBatchSize to the benchmark file count so Blacksmith max runs one native precompile batch instead of the memory-safe default chunks.
• Nuxt SPA build timings exclude synthetic app generation and compare nuxt build with Nuxt's default compiler against the same app with @vizejs/nuxt installed. Vue SFC compilation is roughly 2% of that build (measured on the 500-file corpus in #3426), so this row cannot be moved by making the Vue compiler faster — it is dominated by Nitro, Rollup and Nuxt's own module graph work. bench/nuxt-bridge-transform.mjs isolates the part of this surface Vize does own, the Nuxt module's per-module bridge transform.
• Single-thread lanes are shown where useful, and the primary speedup compares the incumbent default/single-thread lane with Vize's max runner lane.
• Type-check rows span two TypeScript engines: vue-tsc runs the JavaScript compiler while Vize check runs native tsgo (Corsa). No cross-engine ratio is published for that surface — it is ranked within each engine class instead, so TypeScript's Go rewrite is never credited to the Vue layer; bench/check-gate.mjs publishes the same per-engine-class split with planted-diagnostic gating.

Commands:

gh workflow run tool-benchmark.yml --ref <branch> -f file_count=15000 -f check_file_count=500 -f vite_file_count=1000 -f nuxt_file_count=500 -f large_blocks=900 -f runs=5 -f warmups=1 -f commit_results=true
node bench/generate.mjs 15000
node bench/compare-tools.mjs --input bench/__in__ --vize-bin target/release/vize --runs 5 --warmups 1 --check-file-count 500 --vite-file-count 1000 --nuxt-file-count 500 --large-blocks 900 --runner-label "blacksmith-32vcpu-ubuntu-2404" --out tool-benchmark-summary.md --json tool-benchmark-results.json --doc performance-blacksmith.md