A Python implementation of k-Wave — an acoustics toolbox for time-domain simulation of acoustic wave fields. Includes a pure NumPy/CuPy solver (backend="python", supports any CUDA-capable GPU) and an interface to the pre-compiled k-Wave C++ binaries (backend="cpp") with NVIDIA GPU support for compute capability 7.5 (Turing) and newer — covers every consumer/datacenter GPU since 2018, including all shipping Blackwell variants (B200/GB200, B300/GB300, Jetson Thor, RTX 50xx, RTX PRO 6000 Blackwell, GB10/DGX Spark).
Increase the accessibility and reproducibility of k-Wave simulations for medical imaging, algorithmic prototyping, and testing.
A collection of examples covers common simulation scenarios. Run any example locally:
uv run examples/ivp_homogeneous_medium.pyNo GPU required — all examples run on CPU with NumPy.
Reference runtimes for the 3D scaling benchmark in benchmarks/benchmark.py. Values are total elapsed seconds for a single default run (3D initial-value problem, heterogeneous absorbing medium, 1000 timesteps, averaged over 3 repeats).
| Backend | OS | Accelerator | k-wave-python | Binary | 64³ | 128³ | 256³ | Hardware |
|---|---|---|---|---|---|---|---|---|
| python | Linux | CPU | 0.6.3rc1 | — | 127 | — | — | GCP n1-highmem-8, 8 vCPU |
| python | Linux | NVIDIA GPU | 0.6.3rc1 | — | 6 | 71 | 382 | NVIDIA T4, 15 GB (GCP n1-highmem-8) |
| python | macOS (Apple Silicon) | CPU | — | — | 81 | — | — | Apple M1, 8 GB |
| python | Windows | CPU | 0.6.3rc1 | — | 341 | — | — | GCP n1-highmem-8, 8 vCPU |
| python | Windows | NVIDIA GPU | 0.6.3rc1 | — | 6 | 44 | 751 | NVIDIA T4, 15 GB (GCP n1-highmem-8) |
| cpp | Linux | CPU (OMP) | 0.6.3rc1 | v1.4.2 | 6 | 46 | 669 | GCP n1-highmem-8, 8 vCPU |
| cpp | Linux | NVIDIA GPU (CUDA) | 0.6.3rc1 | v1.4.2 | 2 | 7 | 51 | NVIDIA T4, 15 GB (GCP n1-highmem-8) |
| cpp | macOS (Apple Silicon) | CPU (OMP) | 0.6.3rc1 | v1.4.2 | — | — | — | — |
| cpp | Windows | CPU (OMP) | 0.6.3rc1 | v1.4.2 | 18 | 157 | 1243 | GCP n1-highmem-8, 8 vCPU |
| cpp | Windows | NVIDIA GPU (CUDA) | 0.6.3rc1 | v1.4.2 | 2 | 8 | 57 | NVIDIA T4, 15 GB (GCP n1-highmem-8) |
Contributions welcome — open a PR filling a row with your k-wave-python version, and (for cpp backend) BINARY_VERSION. See benchmarks/README.md for the reproducer command.
Using uv (recommended):
uv add k-wave-pythonOr with pip:
pip install k-wave-pythonThe backend="cpp" binaries shipped in v0.6.3+ require compute capability 7.5 (Turing) or newer. CUDA Toolkit 13.0 removed offline-compilation support for older architectures, so the following hardware is not covered by the bundled binaries:
- Maxwell (GTX 9xx, Titan X Maxwell, Tesla M-series, Jetson Nano)
- Pascal (GTX 10xx, P100, P40, Titan X(p)/Xp, Jetson TX2)
- Volta (V100, Titan V, Quadro GV100, Jetson AGX Xavier)
Use backend="python" instead (NumPy/CuPy works on every CUDA-capable GPU), or build the C++ backend from source against CUDA Toolkit 12.x.
Development instructions can be found here.
- k-Wave: A MATLAB toolbox for the time-domain simulation of acoustic wave fields.
- j-wave: Differentiable acoustic simulations in JAX.
- ADSeismic.jl: a finite difference acoustic simulator with support for AD and JIT compilation in Julia.
- stride: a general optimisation framework for medical ultrasound tomography.
The documentation for k-wave-python can be found here.
@software{k-Wave-Python,
author = {Yagubbayli, Farid and Sinden, David and Simson, Walter},
license = {GPL-3.0},
title = {{k-Wave-Python}},
url = {https://github.com/waltsims/k-wave-python}
}e-mail walter.a.simson@gmail.com.