mach-beamform 0.1.1

Ultrafast GPU-accelerated beamforming kernel for ultrasound imaging

mach

An ultrafast CUDA-accelerated ultrasound beamformer for Python users. Developed at Forest Neurotech.

Benchmark: Beamforming PyMUST's rotating-disk Doppler dataset at 1.1 trillion points per second (6.5x the speed of sound).

Highlights

• ⚡ Ultra-fast beamforming: ~10x faster than prior state-of-the-art
• 🚀 GPU-accelerated: Leverages CUDA for maximum performance on NVIDIA GPUs
• 🎯 Optimized for research: Designed for functional ultrasound imaging (fUSI) and other ultrafast, high-channel-count, or volumetric-ensemble imaging
• 🐍 Python bindings: Zero-copy integration with CuPy, and JAX arrays via nanobind. NumPy support included.
• 🔬 Validated: Matches vbeam and PyMUST outputs

Installation

Install from PyPI (recommended):

pip install mach-beamform

Or: to include all optional dependencies, including to run the examples:

pip install mach-beamform[all]

Wheel prerequisites:

• Linux
• CUDA-enabled GPU with driver >= 12.3, compute-capability >= 7.5

Build from source

make compile

Build prerequisites:

• Linux
• make
• uv >= 0.9.7
• gcc >= 8
• nvcc >= 11.0

Examples

Try our examples:

• 📊 Plane Wave Imaging with PICMUS Dataset
• 🩸 Doppler Imaging

If you don't have a CUDA-enabled GPU, you can download the notebook from the docs and open in Google Colab (select a GPU instance).

Contributing

We welcome contributions! Please see CONTRIBUTING.md for guidelines.

Roadmap

Beta release (v0.Y.0)

• ✅ Single-wave transmissions (plane wave, focused, diverging)
• ✅ Linear interpolation beamforming
• ✅ Allow NumPy/CuPy/JAX/PyTorch inputs through Array API
• ✅ Comprehensive error handling
• ✅ PyPI packaging and distribution
• ✅ Interpolation options: nearest, linear, and quadratic

Numerically validated, but looking for feedback on API

• ✅ Coherent compounding

See the project page for our up-to-date roadmap. We welcome feature requests!

Acknowledgments

mach builds upon the excellent work of the ultrasound imaging community:

• vbeam - For educational examples and validation benchmarks
• PyMUST / PICMUS - For standardized evaluation datasets
• Community contributors - Gev and Qi for CUDA optimization guidance

This package was developed by the Forest Neurotech team, a Focused Research Organization supported by Convergent Research and generous philanthropic funders.

Citation

If you use mach in your research, you can cite:

@inproceedings{mach,
  title={{Mach: Beamforming one trillion points per second on a consumer GPU}},
  author={Guan, Charles and Rockhill, Alex and Pinton, Gianmarco},
  booktitle={Medical Imaging 2026: Ultrasonic Imaging and Tomography},
  year={2026},
  organization = {International Society for Optics and Photonics},
  publisher = {SPIE},
  URL={https://github.com/Forest-Neurotech/mach}
}