Apple Silicon Bench

A native macOS benchmark tool for Apple Silicon Macs (M-series and A-series)

A lightweight, native Swift benchmark tool designed specifically for Apple Silicon processors (M1, M2, M3 Family, M4 Family, M5 Family (like Pro, Max), A18 Pro). Supports all M-series and A-series chips with comprehensive CPU, GPU, memory, disk, and AI benchmarks.

Features

• CPU Single-Core & Multi-Core: Integer, floating-point, SIMD, cryptography, compression
• GPU Benchmark (Metal): Compute shaders, particle simulation, image processing
• Memory Benchmark: Bandwidth and latency measurements
• Disk Benchmark: Sequential and random I/O with cache bypass
• AI/ML Benchmark (NEW in v2.0): CoreML inference (CPU/GPU/Neural Engine), BNNS operations
• Thermal Monitoring: Real-time throttling detection
• HTML Reports: Beautiful interactive reports saved to Desktop
• Lightweight: ~2MB standalone binary, no dependencies

Quick Start

Download Binary

curl -LO https://github.com/carlosacchi/apple-silicon-bench/releases/latest/download/osx-bench-macos-arm64.tar.gz
tar -xzf osx-bench-macos-arm64.tar.gz
xattr -cr osx-bench && chmod +x osx-bench
./osx-bench run

Build from Source

git clone https://github.com/carlosacchi/apple-silicon-bench.git
cd apple-silicon-bench
swift build -c release
./.build/release/osx-bench run

Usage

# Full benchmark (recommended)
osx-bench run

# Quick mode (~3s per test, less accurate)
osx-bench run --quick

# Custom duration
osx-bench run --duration 30

# Stress test (60s per test)
osx-bench run --stress

# Selective benchmarks
osx-bench run --only cpu-single,gpu
osx-bench run --only memory,disk
osx-bench run --only ai

# AI benchmark options
osx-bench run --only ai --model-path /path/to/model.mlmodelc
osx-bench run --offline  # Skip AI if model not cached

# Advanced profiling (v2.1.0+)
osx-bench run --advanced  # Stride sweep, QD matrix, thread scaling

# Multiple passes with median (v2.2.0+)
osx-bench run --repeats 3  # Run 3 passes, report median scores

# System info
osx-bench info
osx-bench info --extended

# Export results
osx-bench run --export results.json

Scoring

Total Score

• Baseline: M1 base chip = 1000 points per category (calibrated from median of 5 full runs)
• Method: Geometric mean of ratios (commonly used in benchmark suites)
• Weights: CPU-Single 25%, CPU-Multi 25%, Memory 15%, Disk 15%, GPU 20%

Chip	Expected Score (Rule-of-Thumb)
M1	~1000
M2	~1100
M3	~1290
M4	~1600
M5	~2000
M5 Pro	~2650-2850
M4 Max	~2870
A18 Pro (Neo)	~1165

Note on Disk scores: SSD capacity directly affects disk benchmark results. Larger SSDs (512GB, 1TB+) have more NAND channels working in parallel, producing higher throughput than smaller SSDs (256GB). Disk scores are clamped (0.25x-4.0x vs baseline) to limit this effect, but comparisons across different SSD capacities should account for this hardware difference.

AI Score (Separate)

The AI/ML score is reported separately from the Total Score (similar to Geekbench AI):

• AI-CPU: CoreML inference with CPU-only compute
• AI-GPU: CoreML inference with GPU acceleration
• AI-Neural Engine: CoreML inference with Neural Engine (when available)
• AI-BNNS: Accelerate framework matrix operations

Note: Actual results may vary based on chassis, cooling, and configuration

For detailed methodology, see the Wiki.

Thermal States

• 🟢 Nominal: No throttling
• 🟡 Fair: Minor throttling possible
• 🟠 Serious: Significant throttling
• 🔴 Critical: Maximum throttling

Requirements

• macOS 13.0 (Ventura) or later
• Apple Silicon Mac

Supported Chips

M-series: M1, M1 Pro, M1 Max, M1 Ultra, M2, M2 Pro, M2 Max, M2 Ultra, M3, M3 Pro, M3 Max, M3 Ultra, M4, M4 Pro, M4 Max, M5, M5 Pro, M5 Max

A-series: A18 Pro (MacBook Neo)

Documentation

See the Wiki for:

• Scoring Methodology - How scores are calculated
• Benchmark Details - Technical details of each test
• Advanced Profiling - PassMark-inspired deep analysis
• FAQ - Common questions and troubleshooting
• Roadmap - Planned features

Why Apple Silicon Bench?

Feature	Apple Silicon Bench	Geekbench 6	Cinebench
Open Source	Yes	No	No
Offline	Yes	Account required	Yes
Transparent Scoring	Yes	Closed	Closed
Thermal Monitoring	Yes	No	No
Binary Size	~2MB	~200MB	~1GB
Price	Free	$15 (Pro)	Free

Contributing

Contributions welcome! See the Wiki for guidelines.

License

MIT License - see LICENSE file.

Author

Carlo Sacchi - @carlosacchi

---

Made with Swift for Apple Silicon