Ultrasound Engineering
Faster Ultrasound Beamforming Without a Giant Delay Table
A short engineering look at mach's hybrid delay computation: precompute transmit arrival times, reuse receive delays and evaluate GPU memory alongside reconstruction speed.
By ALEVA | California, USA
For an ultrasound OEM team, faster reconstruction does not always require a larger GPU. A useful question is whether the beamformer spends too much memory storing calculations that could be organized differently.
A 2026 study introduces mach, an open-source CUDA delay-and-sum beamformer with Python bindings. Its practical contribution is a hybrid strategy for handling propagation delays, rather than a new transmit sequence or an AI image-enhancement model. [1]
What changes?
Delay-and-sum reconstruction aligns echoes using their travel times. Precomputing every transmit/receive delay can demand substantial memory; calculating everything again for every frame also costs work.
mach precomputes transmit-wavefront arrival times, calculates receive delays inside GPU thread blocks, and reuses receive-delay calculations across frames through shared memory. It also arranges channel data so neighboring GPU threads can access neighboring memory locations efficiently. [1]
Why this matters for OEM development
For repeated acquisitions with unchanged geometry and reconstruction assumptions, this suggests a way to evaluate memory efficiency before selecting a higher-capacity processor. Reuse must be reconsidered when probe geometry, the reconstruction grid or propagation assumptions change.
The authors' core-kernel benchmark excludes CPU-to-GPU transfers and preprocessing such as IQ demodulation or file conversion. Its timing is therefore not complete scanner latency. [1]
The trade-off to test
Efficient reuse depends on the workload. Temporal batching, GPU memory layout and the target hardware can affect the result; a benchmark configuration should not become a universal speed claim. The official project provides source code and examples for evaluation. [2]
Our proposed OEM evaluation is straightforward: replay the same channel data through a trusted reference and the candidate implementation. Compare numerical reconstruction differences, peak GPU memory, complete processing time and sustained operation. Include transfers and preprocessing in the system-level measurement, and test configuration changes explicitly.
Bottom line: hybrid delay computation is a practical software technique to investigate when ultrasound beamforming is constrained by memory and repeated calculations. It is not proof that every system needs less hardware, and computational acceleration alone does not establish better diagnostic performance.
This article reports no ALEVA benchmark results and makes no claim of commercial implementation.
Sources
Guan C., Rockhill A.P., Sode M., Pinton G. mach: ultrafast ultrasound beamforming. Journal of Medical Imaging 13(6), 062203 (2026). DOI 10.1117/1.JMI.13.6.062203. Accessible author version, April 6, 2026, particularly Sections 4.1, 4.3–4.4. The author record identifies journal publication; the directly resolved publisher page was unavailable during this check.
Forest Neurotech — official mach repository. Rolling project documentation checked October 14, 2026; not an independent benchmark.