Magnum.np.distributed brings 7x faster micromagnetic simulations with 8 GPUs

Micromagnetic simulations are critical for nanomagnetism and spintronics research, but existing GPU-accelerated solvers like Mumax3 and Magnum.np are limited to single-device computation. A new framework, Magnum.np.distributed, breaks that barrier by extending the Python-native Magnum.np with PyTorch Distributed. This enables high-speed communication and computation across multiple GPUs while maintaining ease of installation, platform-agnostic design, and Python compatibility. For the most computationally intensive demagnetisation effective-field calculations, the system achieves a 7.0x speedup across 8 GPUs connected via NVLink.

Beyond GPU scaling, the framework also demonstrates a 6.8x speedup in demagnetisation field computation on CPU using NUMA pinning through PyTorch Distributed's MPI backend. However, the Halo exchange required for Heisenberg exchange shows limited scaling due to kernel dispatch latency. By enabling faster turnaround times, Magnum.np.distributed allows researchers to tackle larger and more complex magnetic systems, accelerating the design cycle for novel spintronic devices. The paper is available on arXiv (2606.01114).