Pith. sign in

REVIEW 1 major objections 15 references

ffdas: Volumetric ultrasound reconstruction at warp speed

T0 review · 1 major / 0 minor · reviewed 2026-06-27 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Three targeted GPU changes let delay-and-sum beamforming reach kilohertz rates on 128-cubed voxel grids with 1024-element arrays.

desk verdict The paper applies standard GPU memory and compute tweaks to delay-and-sum ultrasound reconstruction and releases the code, but the abstract supplies no performance numbers or quality metrics to support the kHz claims or the assertion that image quality is preserved. read the letter →

arxiv 2606.13259 v1 pith:OT2JEQT2 submitted 2026-06-11 physics.med-ph

Luuk Verhoef , Pieter Kruizinga This is my paper · ORCID
classification physics.med-ph
keywords volumetricultrasoundGPUbeamformingdelay-and-sumreal-timereconstructionmixedprecisiontensorcoresultrafastimaging
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper shows that standard delay-and-sum reconstruction on GPUs wastes capacity because memory reads are scattered and under-aligned with hardware transfer sizes. It introduces three fixes: matching access patterns to GPU granularity, storing data in mixed precision to cut traffic in half, and routing arithmetic through tensor cores by using spatial locality. These changes together produce frame rates in the kilohertz range for clinically sized volumes while the authors state that image quality remains comparable to unoptimized code. The result moves volumetric ultrasound from offline-only processing to real-time use in settings that need immediate feedback, such as intraoperative brain imaging.

What carries the argument

The three optimization strategies applied to delay-and-sum beamforming: memory-access alignment, mixed-precision storage, and tensor-core exploitation via spatial locality.

What would settle it

Side-by-side quantitative comparison of standard image-quality metrics (contrast, resolution, artifact levels) between the optimized and baseline reconstructions on identical raw data sets.

Watch

Extended reading notes

Core claim

Aligning memory access with GPU transfer granularity, halving memory traffic through mixed-precision storage, and exploiting spatial locality to utilize tensor core arithmetic together produce kilohertz frame rates for 128^3-voxel grids with 1024-element arrays, substantially outperforming existing implementations while maintaining image quality.

Load-bearing premise

The optimizations preserve diagnostic image quality without introducing artifacts that would require case-by-case validation.

Editorial extensions

If this is right

  • Real-time volumetric imaging becomes feasible at scales previously limited to offline analysis.
  • Applications requiring immediate feedback, such as intraoperative brain imaging and brain-computer interfaces, gain practical support.
  • The open-source ffdas library makes the optimized implementation available for direct use and further development.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same memory-alignment and tensor-core tactics could be tested on other beamforming algorithms that share similar access patterns.
  • If the mixed-precision step holds across varying array sizes, it may reduce hardware requirements for portable ultrasound systems.
  • Extending the approach to even larger grids or higher channel counts would test whether the performance gains scale linearly.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 0 minor

Summary. The manuscript describes three GPU optimizations for delay-and-sum volumetric ultrasound reconstruction: aligning memory accesses with GPU transfer granularity, halving traffic via mixed-precision storage, and exploiting spatial locality with tensor-core arithmetic. These are presented as enabling kilohertz frame rates for 128^3-voxel grids using 1024-element arrays while maintaining image quality, with the code released as the open-source ffdas library to support real-time applications such as intraoperative brain imaging.

Significance. If the reported performance gains hold and image quality is preserved, the work would enable real-time volumetric ultrasound at scales previously restricted to offline processing, with potential impact on time-critical medical applications. The open-source release is a clear strength that supports reproducibility.

major comments (1)
  1. [Abstract] Abstract: The claim that the optimizations 'maintain image quality' is load-bearing for the central assertion of practical utility, yet the manuscript provides no quantitative metrics (PSNR, SSIM, CNR), no comparison protocols against reference DAS, and no description of test datasets or error-propagation analysis under mixed precision. This leaves the quality-preservation step as an unverified assertion.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive feedback and positive evaluation of the work's potential impact. We address the single major comment below and will revise the manuscript accordingly to strengthen the evidence for image quality preservation.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The claim that the optimizations 'maintain image quality' is load-bearing for the central assertion of practical utility, yet the manuscript provides no quantitative metrics (PSNR, SSIM, CNR), no comparison protocols against reference DAS, and no description of test datasets or error-propagation analysis under mixed precision. This leaves the quality-preservation step as an unverified assertion.

    Authors: We agree that the claim requires quantitative substantiation beyond the visual comparisons currently presented in the results. The manuscript does not include the requested metrics, protocols, dataset descriptions, or mixed-precision error analysis. In the revised version we will add a dedicated image quality evaluation section that reports PSNR, SSIM, and CNR on both synthetic point-target phantoms and in-vivo brain datasets, using a double-precision reference DAS implementation as baseline. The section will also describe the comparison protocol and include a brief error-propagation study for the mixed-precision storage scheme. These additions will be supported by new tables and figures. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity in optimization claims or derivation chain

full rationale

The manuscript describes three independent engineering optimizations (memory alignment to GPU granularity, mixed-precision storage, and tensor-core exploitation of spatial locality) applied to a standard delay-and-sum baseline. These steps are presented as direct modifications to memory access patterns and arithmetic, with resulting kHz frame rates claimed as empirical outcomes rather than any first-principles derivation or prediction. No equations, fitted parameters, or self-citations are shown that reduce the central performance or quality claims to their own inputs by construction. The paper is self-contained as an applied GPU engineering contribution with an open-source release, and the absence of quantitative image-quality metrics is a validation gap rather than circular reasoning.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

The work rests on standard assumptions about GPU memory hierarchy and ultrasound physics; no free parameters, new axioms, or invented entities are introduced in the abstract.

assumptions (1)
  • domain assumption Standard delay-and-sum is the appropriate baseline reconstruction method whose memory-access pattern limits GPU utilization.
    The optimizations are defined relative to this baseline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ffdas: Volumetric ultrasound reconstruction at warp speed." pith.science (2026). https://pith.science/paper/OT2JEQT2

@misc{pith2026260613259,
  author       = {Pith},
  title        = {Pith review of: ffdas: Volumetric ultrasound reconstruction at warp speed},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OT2JEQT2}},
  note         = {Machine review of arXiv:2606.13259}
}
abstract

Volumetric ultrafast ultrasound imaging demands reconstruction of images with millions of voxels thousands of times per second, creating computational challenges that limit both real-time feedback and easy offline analysis. Graphics processing units (GPUs) are well suited to this workload, yet we show that standard delay-and-sum implementations underutilize GPU resources through fragmented memory access patterns, even when sufficient computational capacity is available. Three optimization strategies address this: aligning memory access with GPU transfer granularity, halving memory traffic through mixed-precision storage, and exploiting spatial locality to utilize tensor core arithmetic. Together, these achieve kilohertz frame rates for $128^3$-voxel grids with 1024-element arrays, substantially outperforming existing implementations while maintaining image quality. This enables real-time volumetric imaging at scales previously restricted to offline processing, supporting applications such as intraoperative brain imaging and brain-computer interfaces where immediate feedback is essential. We release our implementation as part of the open-source \texttt{ffdas} library.

Figures

Figures reproduced from arXiv: 2606.13259 by the authors.

Figure 1
Figure 1. Delay-and-sum geometry and data layout. (a) Imaging geometry for voxel pn, showing transmit delay τ tx n and receive delays τ rx mn to receivers rm. Required samples (dark) are highlighted in the channel data. Neighboring voxel pn+1 has similar path lengths, causing nearby sample indices. (b) Baseline memory layout with samples k contiguous within each receiver m. Samples required for pn fall at irregular positions … view at source ↗
Figure 2
Figure 2. Performance scaling and roofline analysis. (A-C) Runtime scaling across grid size, batch size, and array size for baseline (BL), batch-inner (BI), and tensor core (TC) implementations, with FP16 and FP32 inputs. (D-E) DRAM and L2 roofline analysis showing operational intensity vs throughput. put storage halves bandwidth requirements for the op￾timized implementations while providing minimal ben￾efit to the BL implem… view at source ↗
Figure 3
Figure 3. Memory access efficiency. (A) Payload efficiency increases from 40% to 100% with batch-inner layout. (B) Request consolidation reduces memory fragmentation, with wider transactions improving bandwidth utilization. (C) Cache hit rates across L1 and L2 levels, showing trade-offs between hit rate and request granularity. B B  [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Warp stall breakdown. Distribution of warp stall reasons across implementations, showing the shift from memory-bound (Long Scoreboard) to more balanced execution with increased arithmetic pipeline utilization (Math Throttle) in optimized variants. 4.5 External library …
Figure 5
Figure 5. Figure 5: External library performance compar￾ison. Runtime comparison with established delay￾and-sum implementations (mach, vbeam) showing sub￾stantial performance improvements while maintaining visual quality under matched reconstruction settings. Each image is normalized indi…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 8 canonical work pages

  1. [1]

    Ultrafast imaging in biomedical ultrasound

    Mickael Tanter and Mathias Fink. “Ultrafast imaging in biomedical ultrasound” . In: IEEE transactions on ultrasonics, ferroelectrics, and 10 frequency control 61.1 (2014), pp. 102–119. doi: 10.1109/TUFFC.2014.6689779

    work page doi:10.1109/tuffc.2014.6689779 2014

  2. [2]

    Shear modulus imag- ing with 2-D transient elastography

    Laurent Sandrin et al. “Shear modulus imag- ing with 2-D transient elastography” . In: IEEE transactions on ultrasonics, ferroelectrics, and frequency control 49.4 (2002), pp. 426–435. doi: 10.1109/58.996560

    work page doi:10.1109/58.996560 2002

  3. [3]

    Ultrafast compound Doppler imaging: Providing full blood flow char- acterization

    Jeremy Bercoff et al. “Ultrafast compound Doppler imaging: Providing full blood flow char- acterization” . In: IEEE transactions on ultra- sonics, ferroelectrics, and frequency control 58.1 (2011), pp. 134–147. doi: 10.1109/TUFFC.2011. 1780

    work page doi:10.1109/tuffc.2011 2011

  4. [4]

    High-definition imaging of carotid artery wall dynamics

    Pieter Kruizinga et al. “High-definition imaging of carotid artery wall dynamics” . In: Ultrasound in Medicine & Biology 40.10 (2014), pp. 2392–

    2014

  5. [5]

    doi: 10.1016/j.ultrasmedbio.2014.03. 009

    work page doi:10.1016/j.ultrasmedbio.2014.03 2014

  6. [6]

    Electromechanical wave imaging for arrhythmias

    Jean Provost et al. “Electromechanical wave imaging for arrhythmias” . In: Physics in Medicine & Biology 56.22 (2011), pp. L1–L11. doi: 10.1088/0031-9155/56/22/F01

    work page doi:10.1088/0031-9155/56/22/f01 2011

  7. [7]

    Functional ultrasound imag- ing of the brain

    Emilie Macé et al. “Functional ultrasound imag- ing of the brain” . In: Nature methods 8.8 (2011), pp. 662–664. doi: https://doi.org/10.1038/ nmeth.1641

    2011

  8. [8]

    Ultrafast ultrasound lo- calization microscopy for deep super-resolution vascular imaging

    Claudia Errico et al. “Ultrafast ultrasound lo- calization microscopy for deep super-resolution vascular imaging” . In: Nature 527.7579 (2015), pp. 499–502. doi: https://doi.org/10.1038/ nature16066

    2015

Show all 15 references
  1. [9]

    So you think you can DAS? A viewpoint on delay-and-sum beamforming

    Vincent Perrot et al. “So you think you can DAS? A viewpoint on delay-and-sum beamforming” . In: Ultrasonics 111 (2021), p. 106309. issn: 0041- 624X. doi: https : / / doi . org / 10 . 1016 / j . ultras.2020.106309

    2021

  2. [10]

    Miniaturized Four- Dimensional Functional Ultrasound for Mapping Human Brain Activity

    Luuk Verhoef et al. “Miniaturized Four- Dimensional Functional Ultrasound for Mapping Human Brain Activity” . In:medRxiv (2025). doi: 10.1101/2025.08.19.25332261

    2025 doi

  3. [11]

    The Tensor-Core Beam- former: A High-Speed Signal-Processing Library for Multidisciplinary Use

    Leon Oostrum et al. The Tensor-Core Beam- former: A High-Speed Signal-Processing Library for Multidisciplinary Use . 2025. arXiv: 2505 . 03269 [cs.DC]

    2025

  4. [12]

    NVIDIA Nsight Compute

    NVIDIA Corporation. NVIDIA Nsight Compute . https : / / developer . nvidia . com / nsight - compute. Version 2024.x. 2024

    2024

  5. [13]

    mach: ultrafast ultrasound beamforming

    Charles Guan et al. “mach: ultrafast ultrasound beamforming” . In: Journal of Medical Imaging 13.6 (2026), pp. 062203–062203. doi: 10.1117/ 1.JMI.13.6.062203 . url: https://github. com/Forest-Neurotech/mach

    2026

  6. [14]

    vbeam: a Fast and Differentiable Beamformer for Optimizing Ultra- sound Imaging

    Magnus Dalen Kvalevåg et al. “vbeam: a Fast and Differentiable Beamformer for Optimizing Ultra- sound Imaging” . In:2023 IEEE International Ul- trasonics Symposium (IUS) . 2023, pp. 1–4. doi: 10.1109/IUS51837.2023.10307255

    2023 doi

  7. [15]

    JAX: composable trans- formations of Python+NumPy programs

    James Bradbury et al. JAX: composable trans- formations of Python+NumPy programs . Ver- sion 0.3.13. 2018. url: http : / / github . com / jax-ml/jax. 11

    2018

Pith tools

Reviewed June 27, 2026 · model on record in the stance chip above.