Benchmarks

---

Reading the Results

When comparing zenpix and Sharp, two axes matter: wall-clock and CPU user.

Metric	Meaning	zenpix position
wall-clock	Real elapsed time until completion	Sharp leads (faster for single requests)
CPU user	Total CPU time consumed across all cores	zenpix is significantly lower

Why CPU user matters: On a 2–4 core VPS handling concurrent requests, high CPU user time means cores compete for work, directly increasing latency for other requests. A faster wall-clock tool that burns more CPU will hurt overall throughput as concurrency grows.

---

Single-point Comparison (manual measurement)

Condition: 3840×2160 PNG → 1920×1080 AVIF (quality=60)
Environment: macOS aarch64 (Apple M4 Pro), illustration fixture (bench_chara_chika.png)
Measurement: /usr/bin/time, warm-up 3 / measure 7, median
Version: zenpix 0.8.0

Tool	wall-clock	CPU user	File size
Sharp quality=60 (libvips auto-thread)	0.422s	2.630s	63 KB
zenpix speed=10 (single-thread)	0.512s	0.530s	106 KB
zenpix speed=6 (single-thread)	0.992s	1.000s	73 KB
zenpix speed=6 (threads=14)	0.610s	1.060s	73 KB

Key takeaways:

• Sharp wins on wall-clock (0.422s)
• zenpix CPU user is ~40% of Sharp (1.060s vs 2.630s)
• threads=14 cuts speed=6 wall-clock from 0.992s → 0.610s (−38%) with nearly flat CPU user

---

Matrix Benchmark (bench/bench.ts)

Condition: PNG decode → Sharp resize to each resolution → AVIF encode (quality=60 / speed=6)
warm-up 2, measure 10; each cell is the median wall-clock (ms)
ratio = Sharp median ÷ zenpix median (> 1 means zenpix is faster)

VPS (Ubuntu · 2vCPU · 2GB)

3 runs; median of 3 values per cell. Measured 2026-05-04, zenpix 0.4.0.

Fixture	FHD (ratio)	WQHD (ratio)	4K (ratio)
bench_input (tile)	0.26	0.25	0.24
bench_chara_chika	1.35	1.26	1.21
bench_chara_kanata	1.36	1.27	1.21
bench_landscape_dark	1.13	1.20	0.97
bench_landscape_impasto	1.47	1.37	1.44
bench_landscape_light	1.03	0.93	0.79

Trend: Tile images (bench_input) favor Sharp. Character art and impasto landscapes favor zenpix.

Mac (M4 Pro · 14 cores)

3 runs; median of 3 values per cell. Measured 2026-04-15.

Fixture	FHD (ratio)	WQHD (ratio)	4K (ratio)
bench_input	0.26	0.18	0.15
bench_chara_chika	0.60	0.44	0.41
bench_chara_kanata	0.63	0.46	0.38
bench_landscape_dark	0.55	0.51	0.38
bench_landscape_impasto	0.57	0.44	0.37
bench_landscape_light	0.53	0.46	0.35

Trend: All cells < 1 on Mac (Sharp wins). The Mac table is used for regression detection; the VPS table is the externally relevant one.

Mac Multi-threaded (threads=14)

Same conditions, same fixtures, with threads=14 (os.cpus().length) passed to encodeAvif. Measured 2026-05-04.

Fixture	FHD (ratio)	WQHD (ratio)	4K (ratio)
bench_input	0.33	0.22	0.19
bench_chara_chika	1.01	0.77	0.69
bench_chara_kanata	1.00	0.76	0.70
bench_landscape_dark	0.69	0.76	0.54
bench_landscape_impasto	0.95	0.72	0.78
bench_landscape_light	0.74	0.64	0.48

Trend: All rows improve over single-thread. Character art FHD reaches parity with Sharp (1.00–1.01).

---

Summary

Use case	Recommendation
Single-request wall-clock priority	Sharp
VPS / low-core concurrent workloads	zenpix (CPU user ~40% of Sharp)
Per-call CPU budget control	zenpix (threads option)
Illustration / character AVIF on VPS	zenpix (ratio 1.2–1.5)

---

Visual Quality Comparison

Speed is one dimension — visual quality is another. At the same quality=60 setting, zenpix preserves more tonal nuance in illustration content. Sharp discards subtle gradients more aggressively to achieve a smaller file size.

Sharp (quality=60)	zenpix (quality=60)

Pastel beach illustration. Sharp produces a smaller file by discarding subtle color transitions; zenpix retains them at a slightly larger size.

---

Re-running Benchmarks

npm run build
npx tsx bench/bench.ts

# Filter fixtures
BENCH_FIXTURES=bench_input,bench_chara_chika npx tsx bench/bench.ts

# Multi-thread measurement
npm run bench:threads

# Quality-matched comparison
npm run bench:quality

Results are written to bench/results/benchmark.json and benchmark.md.