The CUDA-based micromagnetic simulator, rebuilt to run natively on Apple Silicon.
Upstream mumax³ is built around NVIDIA CUDA,
which is unavailable on current Macs. This fork replaces that GPU execution
layer with a native Metal implementation, so an Apple-silicon Mac can run
ordinary .mx3 simulations locally without an NVIDIA GPU, CUDA, a virtual
machine, or a remote Linux host.
The port changes the hardware backend, not the physical model. The .mx3
language, high-level Go solver, material terms, integration methods, and output
formats remain mumax³-compatible.
Important
The Metal backend supports Apple Silicon (M1 or newer) on macOS 14 or newer. Intel Macs are not supported. The original CUDA backend remains available for NVIDIA-equipped Linux and Windows systems.
On a fresh supported Mac, open Terminal and run:
/bin/bash -c "$(/usr/bin/curl -fsSL https://raw.githubusercontent.com/TaewoooPark/mumax3-for-mac/master/install-macos.sh)"The installer checks the machine and shell architecture, opens Apple's Command
Line Tools installer when necessary, installs native Homebrew and a compatible
Go toolchain if missing, clones and builds this Metal port, configures the
binary path, and finishes with mumax3 -test. It is safe to rerun after an
interruption because completed prerequisites are reused.
When installing from an existing checkout, run:
./install-macos.shOpen a new Terminal after installation so the updated path is loaded.
The default mode runs the simulation with the live Web UI:
mumax3 example.mx3The UI is served at http://127.0.0.1:35367, and simulation output is written
to example.out/. If a browser does not open automatically, open that address
manually.
For a headless benchmark, script, or batch job, disable the UI explicitly:
mumax3 -http="" example.mx3Both modes execute the same simulation and produce the same output. The full Xcode application and an offline Metal compiler are not required; the shader library can compile through the system Metal runtime.
| Upstream CUDA component | macOS implementation | Compatibility preserved |
|---|---|---|
| CUDA compute kernels | Metal compute shaders and an Objective-C++ runtime bridge | Existing mumax³ kernel contracts and FP32 model |
| cuFFT | MPSGraph FFT | Packed R2C/C2R layout used by the demagnetizing-field convolution |
cudaMalloc and CUDA streams |
Apple unified-memory buffers, command batching, and one ordered Metal queue | Buffer semantics and execution ordering |
| cuRAND thermal noise | Philox4x32-10 with Box–Muller normal generation | Seeded reproducibility on Metal and validated noise statistics |
| CUDA-only platform bindings | darwin/arm64 build tags and compatibility shims |
The .mx3 language and high-level Go API |
The implementation also uses X-contiguous 32-wide SIMD tiles suited to Apple GPUs and runtime shader compilation so the project works with Command Line Tools alone.
The physical equations and solvers were not reimplemented independently; the native backend executes the existing mumax³ model through Metal. Validation against unchanged upstream CUDA-era regression references produced:
| Validation measure | Result |
|---|---|
| Non-thermal upstream physics tests passed | 15/15 |
| Assertions within the original upstream tolerances | 103/103 (100.000%) |
| Mean average-magnetization vector agreement | 99.9923% |
| Minimum average-magnetization vector agreement | 99.9415% |
| Zero-tolerance assertions matched exactly | 19/19 |
| Official mumax³ example-page simulations completed | 15/15 |
The 100% result is tolerance conformance, not a claim of bit-for-bit identity. Parallel GPU reductions can differ in their last floating-point bits. Thermal simulations use Philox instead of cuRAND's XORWOW sequence, so the same seed is reproducible on Metal but does not generate the same sample-by-sample trajectory as CUDA; the statistical behavior is validated instead.
See the complete physics-validation method and data and the official examples, outputs, provenance, and GPL license.
For compact price-class context, the measured base M4 Metal result is compared with the RTX 4050 Laptop GPU in the official mumax³ 4-million-cell benchmark. Both appeared in consumer notebook families with overlapping launch-price ranges, although this is not an exact configuration- or price-normalized test.
| Comparable notebook CUDA reference | M4 Metal throughput ÷ CUDA throughput |
|---|---|
| RTX 4050 (mobile) | 37.5% |
The measured machine was a fanless, base 10-core-GPU M4 MacBook Air, so this is one conservative hardware datapoint rather than an Apple Silicon performance ceiling. Absolute results, raw runs, the full CUDA comparison, projections, limitations, and the measurement-submission protocol are kept in the benchmark documentation.
This project is derived from mumax³, whose design and verification are described in the original paper. NVIDIA CUDA users should follow the official mumax³ installation documentation. This fork is distributed under the GNU GPL v3 or later. The project notices preserve the upstream CUDA linking permission and identify the July 2026 macOS modifications. Licenses and attribution for Random123, Go, SVGo, and Freetype-Go are collected in the third-party notices.
I am Taewoo Park, and I am an undergraduate physics student at the Korea Advanced Institute of Science and Technology (KAIST). Since October 2025, I have conducted experimental spintronics research on magnetic domain wall motion and neuromorphic computing applications at the KAIST Ultrafast Spin Dynamics Laboratory (USDL), led by Professor Kab Jin Kim. From June 2023 through March 2024, I studied domain wall motion through theoretical modeling and micromagnetic simulation in the KAIST Quantum Spin Dynamics Laboratory under Professor Se Kwon Kim.
