A Fast Bundler
We ran the rolldown/benchmarks suite (10,000 React JSX components) against 7 bundlers. howth came out on top at 275ms — ahead of bun, esbuild, rolldown, vite, rspack, and rsbuild.
Results
GCP c3-highcpu-8 — Linux x64
Intel Xeon Platinum 8481C @ 2.70GHz, 4 cores / 8 threads, 16GB RAM. Warm cache, 10 runs, measured with hyperfine.
| Tool | Version | Time | vs howth |
|---|---|---|---|
| howth | 0.4.0 | 275ms | 1.0x |
| Bun | 1.3.9 | 307ms | 1.1x |
| esbuild | 0.27.3 | 589ms | 2.1x |
| Rolldown | 1.0.0-rc.3 | 735ms | 2.7x |
| Vite | 7.3.1 | 922ms | 3.3x |
| rspack | 1.7.5 | 2,067ms | 7.5x |
| Rsbuild | 1.7.3 | 2,176ms | 7.9x |
macOS — Apple M3 Pro
| Tool | Version | Time | vs howth |
|---|---|---|---|
| howth | 0.4.0 | 276ms | 1.0x |
| Bun | 1.3.9 | 350ms | 1.3x |
| esbuild | 0.27.3 | 724ms | 2.6x |
| Rolldown | 1.0.0-rc.3 | 765ms | 2.8x |
| Vite | 7.3.1 | 1,203ms | 4.4x |
| Rsbuild | 1.7.3 | 1,587ms | 5.8x |
| rspack | 1.7.5 | 1,648ms | 6.0x |
All tools configured identically: production mode, minification, sourcemaps, no gzip.
How we got here
A day earlier, howth was at 634ms on this benchmark — slower than bun. Here's what changed.
The bottleneck: SWC
Profiling showed 90% of time was spent in build_graph_parallel, and the single biggest cost was SWC transpilation — 48% of all worker thread samples. Every .jsx file was being parsed three times:
- By howth-parser for import extraction
- By SWC for JSX →
_jsx()transformation - By howth-parser again to re-extract imports from transpiled output
The apps/10000 benchmark is 100% .jsx files with no TypeScript. SWC was entirely unnecessary.
Optimization 1: JSX fast path — 634ms → 520ms
howth-parser already had complete JSX codegen — it just wasn't wired into the bundler. We added a transform_jsx() function that does parse + codegen + import extraction in a single pass, and routed .jsx files through it instead of SWC.
For this benchmark, SWC is completely eliminated from the hot path. TypeScript files still use SWC for type stripping.
Optimization 2: Parallel resolver — 520ms → 384ms
The module resolver was sequential. We made the Resolver struct thread-safe with RwLock<HashMap> caches and moved resolution into rayon parallel closures. Imports now resolve across all CPU cores simultaneously.
Optimization 3: Directory listing cache — 384ms → 246ms
The resolver was doing ~12 stat() syscalls per import, trying extensions like .js, .jsx, .ts, .tsx, /index.js, etc. For 10,000 modules, that's over 120,000 stat() calls.
The fix: read each directory once and cache file names as HashSets. One readdir replaces all per-extension probing. We also replaced canonicalize() (which calls realpath()) with in-memory path normalization that resolves . and .. without any syscalls.
Optimization 4: Drop SWC minifier — cleaner pipeline
Replaced SWC's minifier with howth-parser's CodegenOptions { minify: true }. Performance-neutral, but removes the last SWC dependency from the JS/JSX bundling path.
Why howth is fast
- Native binary — no Node.js startup cost
- Single-pass JSX — howth-parser does parse + import extraction + codegen in one pass
- Rayon parallelism — file I/O and transpilation across all cores
- Arena allocation — bump allocator for AST nodes, better cache locality
- Directory listing cache — one readdir per directory, HashSet lookups
- Dense module graph —
Vec<Module>indexed byusize, no pointer chasing - In-memory path normalization — zero realpath() syscalls
Methodology
All benchmarks use the rolldown/benchmarks suite (forked with howth added). The GCP benchmark ran on a dedicated c3-highcpu-8 spot instance with no other workloads. All tools configured identically. Measured with hyperfine (10 runs).
Update: v0.5.0 — Variable Name Mangling (February 10, 2026)
v0.5.0 adds variable name mangling to the minifier. --minify now shortens local variable names by default (myVariable → a), matching the behavior of esbuild and bun.
GCP c3-highcpu-8 — Linux x64 (updated)
Intel Xeon Platinum 8481C @ 2.70GHz, 4 cores / 8 threads, 16GB RAM. 10 runs, hyperfine.
With mangling (default --minify behavior):
| Tool | Version | Time | JS Size | vs fastest |
|---|---|---|---|---|
| Bun | 1.3.9 | 528ms | 5.34 MB | 1.0x |
| howth | 0.5.0 | 670ms | 4.13 MB | 1.3x |
| Rolldown | 1.0.0-rc.3 | 1,144ms | 5.22 MB | 2.2x |
| esbuild | 0.27.3 | 1,248ms | 5.90 MB | 2.4x |
| Vite | 7.3.1 | 1,498ms | 5.28 MB | 2.8x |
| Rsbuild | 1.7.3 | 2,550ms | 5.70 MB | 4.8x |
| rspack | 1.7.5 | 2,676ms | 5.18 MB | 5.1x |
Without mangling (--minify --no-mangle):
| Tool | Version | Time | JS Size | vs fastest |
|---|---|---|---|---|
| Bun | 1.3.9 | 528ms | 5.34 MB | 1.0x |
| howth | 0.5.0 | 549ms | 5.26 MB | 1.04x |
Without mangling, howth and bun are neck and neck. Mangling adds a re-parse and rename pass (~120ms overhead) but drops the JS output by 25% (5.26 MB → 4.13 MB) — the smallest in the benchmark.
macOS — Apple M3 Pro (updated)
| Tool | Version | Time | JS Size | vs fastest |
|---|---|---|---|---|
| Bun | 1.3.9 | 321ms | 5.34 MB | 1.0x |
| howth | 0.5.0 | 460ms | 4.02 MB | 1.4x |
| esbuild | 0.27.3 | 796ms | 5.91 MB | 2.5x |
| Rolldown | 1.0.0-rc.3 | 813ms | 5.22 MB | 2.5x |
| Vite | 7.3.1 | 1,274ms | 5.28 MB | 4.0x |
| Rsbuild | 1.7.3 | 1,607ms | 5.70 MB | 5.0x |
| rspack | 1.7.5 | 1,696ms | 5.18 MB | 5.3x |
See Removing SWC: Building a Custom TypeScript Parser and Minifier for details on how the mangler works.
Update: Per-Module Minification (February 10, 2026)
v0.5.0's minifier re-parsed the entire concatenated bundle (~5 MB) as a single pass. This added ~170ms on M3 Pro. The fix: minify and mangle each module inside the existing par_iter() loop instead of re-parsing the full bundle. Each module wrapper is only ~500 bytes, so 10,000 parallel parses are near-instant.
GCP c3-highcpu-8 — Linux x64 (updated)
Intel Xeon Platinum 8481C @ 2.70GHz, 4 cores / 8 threads, 16GB RAM. 10 runs, hyperfine.
| Tool | Version | Time | JS Size | vs howth |
|---|---|---|---|---|
| howth | 0.5.0 | 290ms | 4.12 MB | 1.0x |
| Bun | 1.3.9 | 541ms | 5.34 MB | 1.9x |
| esbuild | 0.27.3 | 1,090ms | 5.90 MB | 3.8x |
| Rolldown | 1.0.0-rc.3 | 1,179ms | 5.22 MB | 4.1x |
| Vite | 7.3.1 | 1,530ms | 5.28 MB | 5.3x |
| Rsbuild | 1.7.3 | 2,775ms | 5.70 MB | 9.6x |
| rspack | 1.7.5 | 2,930ms | 5.18 MB | 10.1x |
macOS — Apple M3 Pro (updated)
| Tool | Version | Time | JS Size | vs fastest |
|---|---|---|---|---|
| Bun | 1.3.9 | 315ms | 5.34 MB | 1.0x |
| howth | 0.5.0 | 317ms | 4.01 MB | 1.0x |
| esbuild | 0.27.3 | 736ms | 5.91 MB | 2.3x |
| Rolldown | 1.0.0-rc.3 | 799ms | 5.22 MB | 2.5x |
| Vite | 7.3.1 | 1,229ms | 5.28 MB | 3.9x |
| Rsbuild | 1.7.3 | 1,569ms | 5.70 MB | 5.0x |
| rspack | 1.7.5 | 1,646ms | 5.18 MB | 5.2x |
1.86x faster than Bun on C3, tied on M3 Pro. howth produces the smallest output in both benchmarks (4 MB vs 5.3 MB, 23% smaller).