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

Bram Veenboer; John W. Romein; Leon Oostrum; Michael Brown; Pieter Kruizinga; Ronald Rook

REVIEW 3 major objections 4 minor 14 references

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

T0 review · 3 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read By rewriting beamforming as complex matrix multiplication, a single library runs it on GPU tensor cores at over 3 PetaOps/s in 1-bit mode and over 600 TOps/s in float16, with far lower energy use.

desk verdict A real, open-source tensor-core complex matmul library with impressive measured performance and two solid application demos, but the 1-bit complex-multiplication derivation in Eq. (5) is wrong as printed and needs a fix before the headline 3-PetaOps/s claim can be trusted. read the letter →

arxiv 2505.03269 v1 pith:D5FJYZWI submitted 2025-05-06 cs.DC astro-ph.IM

Leon Oostrum , Bram Veenboer , Ronald Rook , Michael Brown , Pieter Kruizinga , John W. Romein This is my paper · ORCID

classification cs.DCastro-ph.IM

keywords tensorcoresbeamformingGPUmatrix-matrixmultiplication1-bitarithmeticultrasoundimagingradioastronomyenergyefficiency

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

Beamforming combines sensor signals into directional beams, and for many problems it can be rewritten as a complex-valued matrix-matrix multiplication. This paper claims that by moving that multiplication onto GPU tensor cores, a single library can outrun traditional GPU-core beamformers by one to two orders of magnitude while using far less energy. The library supports 16-bit and 1-bit data, and the authors demonstrate it on medical ultrasound and radio-astronomical workloads, reporting more than 600 TOps/s in float16 mode on an AMD MI300X and more than 3 PetaOps/s in 1-bit mode on an NVIDIA A100.

What carries the argument

The key object is ccglib, a runtime-compiling GPU library that wraps tensor-core matrix multiplication behind a simple interface and hides GPU vendor differences. For float16, the core trick is the five-step decomposition of complex multiplication into four real matrix products. For 1-bit, the core trick is expressing a dot product of signed bits as $K - 2\,\mathrm{popc}(A \oplus B)$, with the XOR replaced by two ANDs on Hopper GPUs, and handling padding by cancelling the contribution of padded columns. The library auto-tunes block and warp sizes for each GPU and data size.

What would settle it

Implement a small complex matrix product in 1-bit mode with the published encoding (each component one sign bit), compute it by hand using Eqs. (4)–(6), and compare with a direct complex multiplication; any mismatch in sign or magnitude shows the printed formulas do not describe the algorithm.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the beamforming computation, normally a sum of complex products over antenna elements, can be implemented on tensor cores with real-valued matrix products. For 16-bit data, complex multiplication is decomposed into four real matrix multiplications with one fast in-register negation. For 1-bit data, each complex component is encoded as a single sign bit and the products are evaluated with bitwise XOR or AND operations followed by population counts, yielding a matrix product without a general-purpose multiply unit. The authors report that this approach reaches 50–85% of peak tensor-core throughput on large matrices and comes close to memory-bandwidth limits on small matrices.

Load-bearing premise

The load-bearing premise is that the printed 1-bit matrix-multiplication formulas match what ccglib actually executes; the paper's own text and its Figure 1 disagree about the meaning of the binary digits and about whether the real parts are added or subtracted, so this link is not established as published.

Editorial extensions

If this is right

If the claimed performance holds, real-time 3D computational ultrasound imaging becomes achievable with current GPUs, which was previously out of reach.
The LOFAR radio-astronomical beamformer can run 2–20 times faster and about 10 times more energy efficiently, enabling more simultaneous beams or reduced hardware.
Beamforming with 1-bit quantization yields usable images and signal products, suggesting that aggressive quantization is acceptable when many samples are accumulated.
The auto-tuning approach makes the same library code efficient across both NVIDIA and AMD GPUs, so applications do not need vendor-specific kernels.

Reading between the lines

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

A hand-computed 2×2 complex 1-bit product would quickly settle whether the printed equations in Section III-D match the implemented code; the paper's text and its Figure 1 use opposite sign conventions for the binary digits.
Because the float16 kernel is memory-bound for small matrices, real-world pipelines with many small batches would probably benefit from fusing the packing and transpose steps with the tensor-core kernel, something the paper leaves for future work.
The same four-product complex-GEMM decomposition could equally accelerate other complex-valued tensor-core workloads, such as convolution or magnetic-resonance reconstruction, whenever inputs fit the 16-bit range.
If the 1-bit formulas are correct after a sign correction, the method could generalize to other clipped-signal processing, such as sign-based correlation or sparse-aperture imaging.

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, and a circularity audit.

Desk Editor's Note

read the letter

The paper is worth a serious look. It ships a genuine artifact, ccglib/TCBF, that maps complex-valued beamforming onto GPU tensor cores with 16-bit and 1-bit inputs, supports both NVIDIA and AMD, and is evaluated across seven GPUs with micro-benchmarks, roofline analysis, auto-tuning, and two real applications (LOFAR radio astronomy, 3D computational ultrasound). The measured numbers are plausible and internally consistent: the 16-bit mode reaches over 600 TOP/s on the MI300X, the ultrasound demo hits real-time frame rates, and the LOFAR comparison is fair because the reference is run without weight calculation. The library is open source, which is a real plus. This is the first time tensor-core beamforming has been applied to both domains, and the generic complex-matmuls layer is a useful contribution in its own right.

Now the soft spots. The biggest one is Section III-D. The 1-bit encoding is actually consistent throughout: bit 0 = −1, bit 1 = +1, and Eq. (4), K − 2 popc(A⊕B), correctly gives the dot product. But Eq. (5) for the real part of the complex product is wrong. Under the stated encoding, Re(A×B) = [K − 2 popc(Ar⊕Br)] − [K − 2 popc(Ai⊕Bi)], which equals 2(popc(Ai⊕Bi) − popc(Ar⊕Br)). The paper prints 2[K − (popc(Ar⊕Br) + popc(Ai⊕Bi))], which is the sum, i.e. Re(A×conj(B)), not Re(A×B). The prose directly above says the two matrix multiplications are subtracted, so the equation contradicts the derivation. If the code follows the printed equation, the 1-bit ultrasound images would be using the wrong complex multiplication; if it's a typo, then as published the 1-bit path cannot be checked without a unit test or worked example. Either way, the claim that 1-bit mode achieves 3 PetaOps/s and produces correct images is not currently verifiable from the paper.

Other issues are minor by comparison: no error bars on any benchmark, no quantitative accuracy comparison for the 1-bit outputs (just "usable image feedback"), and auto-tuning parameters were selected on the same matrix sizes used for the headline numbers, though the application demos do provide some independent validation. A commit hash or benchmark data bundle would improve reproducibility, but the library link is given.

Who is this for? Anyone working on GPU beamforming, tensor-core programming, or real-time computational ultrasound. It deserves peer review—not a desk reject—but the revision must fix Eq. (5), add a small worked example or unit test for 1-bit complex multiplication, and ideally add error bars and a basic accuracy check. Once that is done, I would be comfortable signing off on it.

Referee Report

3 major / 4 minor

Summary. The paper presents the Tensor-Core Beamformer (TCBF), implemented as the ccglib library, which performs complex-valued matrix-matrix multiplication on GPU tensor cores for beamforming workloads. The library supports float16 on NVIDIA and AMD GPUs and 1-bit integer on NVIDIA GPUs, with packing, transpose, and auto-tuned matrix-multiplication kernels. The authors benchmark the library on seven GPUs, report up to 603 TOPs/s in float16 on an AMD MI300X and 3080 TOPs/s in int1 on an NVIDIA A100, analyze performance with roofline models, and demonstrate applications in computational ultrasound imaging and LOFAR radio-astronomy beamforming. The central claim is that TCBF outperforms conventional GPU-core beamformers by one to two orders of magnitude in throughput and energy efficiency.

Significance. If the published 1-bit derivation is corrected and validated, this is a practically significant systems contribution. The paper combines micro-benchmarks, roofline analysis, auto-tuning, and two real application integrations, and it is honest about the WMMA/WGMMA performance gap on Hopper. The cross-vendor evaluation on NVIDIA and AMD GPUs and the explicit reporting of energy efficiency are strengths. The measured performance is plausible against the tensor-core peak numbers in Table I, and the LOFAR comparison against a reference beamformer gives a concrete, non-self-referential speedup estimate. The main risk is the 1-bit complex-multiplication path, since the published equations are internally inconsistent and no correctness test is reported.

major comments (3)

[§III-D, Eq. (5)] The printed formula for the real part is not the difference of the two real-valued dot products that the surrounding text describes. With the encoding of Fig. 1 and Table II (bit 1 = +1, bit 0 = −1), the dot product in Eq. (4) is K − 2 popc(A⊕B). The real part of the complex product therefore requires (K − 2 popc(Ar⊕Br)) − (K − 2 popc(Ai⊕Bi)) = 2(popc(Ai⊕Bi) − popc(Ar⊕Br)). Equation (5) instead evaluates to 2(K − popc(Ar⊕Br) − popc(Ai⊕Bi)), which is the sum of the two real dot products, i.e. it equals Re(A×conj(B)) under the stated encoding. This contradicts the sentence immediately above the equation, which states that the two matrix-matrix multiplications are subtracted. Because Eq. (5) is the basis for the int1 results in Table III and for the ultrasound demonstration in Section V-A, the 1-bit path is currently not verifiable from the published derivation.
[§III-D, Eqs. (5)-(6)] The meaning of K in Eqs. (5)-(6) is ambiguous. In Section II, K is explicitly the number of receivers, i.e. the unpadded reduction dimension. With that definition, the raw tensor-core result for the imaginary part contains 2Kpad spurious +1 contributions, one for each of the two matrix products, so the printed imaginary-part formula is off by an extra Kpad inside the factor 2. If the authors instead intend K in Eqs. (5)-(6) to denote the padded dimension, this must be stated and the equations re-derived consistently from Eq. (4) and Table II. Please define K, Kpad, and the domain over which the popcounts are taken, and re-derive both equations.
[§V-A, Fig. 6] No quantitative correctness validation is provided for the 1-bit pipeline. The paper reports that the 1-bit beamformer produces usable image feedback, but it does not compare the beamformed volume against a float16 or float32 reference and no error metric is given. Given the inconsistency in Eq. (5), a validation experiment (for example, beamforming the same dataset in 1-bit and float32 and reporting a normalized error, or comparing against a known closed-form matrix product) is required to confirm that the 1-bit path computes the intended complex product.

minor comments (4)

[§III-D, paragraph below Table II] The sentence 'The final value of the dot product is then determined by subtracting the number of binary zeroes from the number of binary ones' describes the negative of Eq. (4); under the stated encoding the value is (number of zeroes) minus (number of ones).
[§III-D, Eq. (6)] The formula as typeset has popc(A∧B) twice; the text describes a second AND of the negated inputs, so the second term should be popc(¬A∧¬B) or the notation must be introduced explicitly.
[§IV-C, Table III] The paper should state whether the reported performance and energy numbers are single measurements, means, or medians, and ideally give run-to-run variance for the headline configurations, since the speedup and energy-efficiency claims rest on these measurements.
[§III, code availability] The GitLab URL is not versioned; a DOI or a release tag would improve reproducibility of the reported benchmark numbers.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the central claims are empirical measurements against hardware ceilings and reference kernels; the 1-bit equation inconsistency is a correctness/verifiability issue, not circularity.

full rationale

The paper's central claims are measured benchmark results, not derived predictions. The complex-matrix-multiply implementation is a straightforward five-step tensor-core scheme (Section III-B); the float16 performance is measured against cudapeak micro-benchmark ceilings and an independent roofline model (Section IV-B), and the application claims are compared against the existing LOFAR beamformer and Octave/OpenCL baselines. Auto-tuning (Section IV-A) selects kernel launch parameters on representative matrix sizes; Table III and Fig. 4 then report measured throughputs, including sizes different from the tuning size, so no fitted parameter is renamed as a prediction. The self-citation [4] is used only for an interleaved-storage idea in future work and is not load-bearing. The 1-bit path in Section III-D does contain a serious internal inconsistency: Eq. (5) prints the sum of two popcounts where the surrounding prose says the two matrix-matrix multiplications 'are subtracted from each other,' and Eq. (6) omits the negation of the second AND operand; if the printed formulas are implemented, the 1-bit complex product is wrong, and if they are misprints, the published path is not verifiable. That is a correctness/reproducibility problem, not a circularity, and therefore does not raise the circularity score.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper contributes a software library, not a new physical theory; the main assumptions are about GPU hardware behavior, the bit encoding for 1-bit arithmetic, and the power measurement methodology. No new particles, forces, or physical entities are introduced.

free parameters (1)

Kernel tuning parameters (tile sizes, warp counts, buffer count) = Varies by GPU; see Table III (e.g., MI300X float16: M per block 128, N per block 128, M per warp 64, N per warp 32…
The headline performance numbers are obtained after auto-tuning these parameters to specific benchmark matrix sizes; the reported peaks are tuned values, not typical-case performance across matrix shapes.

assumptions (3)

domain assumption GPU tensor cores execute the documented WMMA/rocWMMA matrix multiply-accumulate instructions, including the 1-bit fragment layout used via inline PTX.
The library's correctness and all measured throughputs depend on hardware semantics described in Section III; the paper attributes the GH200 shortfall to WMMA limitations.
ad hoc to paper The bit encoding for 1-bit values (binary 0/1 map to ±1) is applied consistently in the implementation.
Section III-D and Fig. 1 define the encoding, but the text conflicts on whether binary 0 means +1 or -1; the 1-bit derivation depends on this mapping.
domain assumption Power measurements obtained with PMT reflect the true average energy consumption of the GPU kernels.
Energy efficiency is computed as performance divided by average power during kernel execution, without calibration or error bounds (Section IV-A).

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Tensor-Core Beamformer: A High-Speed Signal-Processing Library for Multidisciplinary Use." pith.science (2026). https://pith.science/paper/D5FJYZWI

@misc{pith2026250503269,
  author       = {Pith},
  title        = {Pith review of: The Tensor-Core Beamformer: A High-Speed Signal-Processing Library for Multidisciplinary Use},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D5FJYZWI}},
  note         = {Machine review of arXiv:2505.03269}
}

read the original abstract

Beamforming is a well-known technique to combine signals from multiple sensors. It has a wide range of application domains. This paper introduces the Tensor-Core Beamformer: a generic, optimized beamformer library that harnesses the computational power of GPU tensor cores to accelerate beamforming computations. The library hides the complexity of tensor cores from the user, and supports 16-bit and 1-bit precision. An extensive performance evaluation on NVIDIA and AMD GPUs shows that the library outperforms traditional beamforming on regular GPU cores by a wide margin, at much higher energy efficiency. In the 16-bit mode, it achieves over 600 TeraOps/s on an AMD MI300X GPU, while approaching 1 TeraOp/J. In the 1-bit mode, it breaks the 3 PetaOps/s barrier and achieves over 10 TeraOps/J on an NVIDIA A100 GPU. The beamforming library can be easily integrated into existing pipelines. We demonstrate its use for medical ultrasound and radio-astronomical instruments.

Figures

Figures reproduced from arXiv: 2505.03269 by the authors.

**Figure 2.** Auto-tuning results of ccglib matrix-matrix multiplication kernel. The measured performance and energy efficiency of each combination of tuning parameters is shown. 2 3 2 6 2 9 2 12 2 15 2 18 10 0 10 1 10 2 10 3 TOPs/s small small small small int1 tensor float16 tensor float32 AD4000 float16 int1 2 3 2 6 2 9 2 12 2 15 2 18 10 1 10 2 10 3 small small big big int1 tensor float16 tensor float32 A100 float16 int1 2 3 2 … view at source ↗

**Figure 3.** Roofline analysis of the ccglib matrix-matrix multiplication kernel. For each GPU, we show the roofline ceiling of the float16 and int1 (NVIDIA only) tensor cores, as well as the normal float32 cores for comparison [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Complex matrix-matrix multiplication benchmark results for (a) 16-bit data and (b) 1-bit data. The left panels show [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: Performance of beamforming for ultrasound. The [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 6.** Figure 6: Three orthogonal (sagittal, coronal and axial) maximum [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 7.** Figure 7: Performance (left) and energy efficiency (right) of the LOFAR TCBF. The reference lines for A100 and GH200 [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 13 canonical work pages

[1]

Beamforming: a versatile approach to spatial filtering,

B. V . Veen and K. Buckley, “Beamforming: a versatile approach to spatial filtering,” IEEE ASSP Magazine, vol. 5, pp. 4–24, 4 1988. [Online]. Available: http://ieeexplore.ieee.org/document/665/

work page 1988
[2]

The Apertif Radio Transient System (ARTS): Design, commissioning, data release, and detection of the first five fast radio bursts,

J. van Leeuwen et al. , “The Apertif Radio Transient System (ARTS): Design, commissioning, data release, and detection of the first five fast radio bursts,” A&A, vol. 672, p. A117, Apr. 2023

work page 2023
[3]

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

V . Perrot, M. Polichetti, F. Varray, and D. Garcia, “So you think you can DAS? A viewpoint on delay-and- sum beamforming,” Ultrasonics, vol. 111, p. 106309,

work page
[4]

The Tensor-Core Correlator,

J. W. Romein, “The Tensor-Core Correlator,” A&A, vol. 656, p. A52, Dec. 2021

work page 2021
[5]

Benchmarking and Dissecting the Nvidia Hopper GPU Architecture,

W. Luo, R. Fan, Z. Li, D. Du, Q. Wang, and X. Chu, “Benchmarking and Dissecting the Nvidia Hopper GPU Architecture,” in 2024 IEEE International Parallel and Distributed Processing Symposium (IPDPS) . Los Alamitos, CA, USA: IEEE Computer Society, may 2024, pp. 656–667. [Online]. Available: https://doi.ieeecomputersociety.org/ 10.1109/IPDPS57955.2024.00064

arXiv 2024
[6]

Kernel Tuner: A search-optimizing GPU code auto-tuner,

B. van Werkhoven, “Kernel Tuner: A search-optimizing GPU code auto-tuner,” Future Generation Computer Systems, vol. 90, pp. 347–358, 2019. [Online]. Available: https://www.sciencedirect.com/science/article/ pii/S0167739X18313359

work page 2019
[7]

Bringing Auto-Tuning to HIP: Analysis of Tuning Impact and Difficulty on AMD and Nvidia GPUs,

M. Lurati, S. Heldens, A. Sclocco, and B. van Werkhoven, “Bringing Auto-Tuning to HIP: Analysis of Tuning Impact and Difficulty on AMD and Nvidia GPUs,” in Euro-Par 2024: Parallel Processing , J. Car- retero, S. Shende, J. Garcia-Blas, I. Brandic, K. Olcoz, and M. Schreiber, Eds. Cham: Springer Nature Switzer- land, 2024, pp. 91–106

work page 2024
[8]

PMT: Power Measurement Toolkit,

S. Corda, B. Veenboer, and E. Tolley, “PMT: Power Measurement Toolkit,” in 2022 IEEE/ACM International Workshop on HPC User Support Tools (HUST) , 2022, pp. 44–47

work page 2022

Show all 14 references

[9]

Compressive 3D ultrasound imaging using a single sensor,

P. Kruizinga et al. , “Compressive 3D ultrasound imaging using a single sensor,” Science Advances , vol. 3, no. 12, p. e1701423, 2017. [Online]. Available: https://www.science.org/doi/abs/10.1126/sciadv.1701423

2017 doi
[10]

Four-dimensional computational ultrasound imaging of brain hemodynamics,

M. D. Brown et al. , “Four-dimensional computational ultrasound imaging of brain hemodynamics,” Science Advances, vol. 10, no. 3, p. eadk7957, 2024. [Online]. Available: https://www.science.org/doi/abs/10. 1126/sciadv.adk7957

2024
[11]

LOFAR: The LOw-Frequency ARray,

M. P. van Haarlem et al., “LOFAR: The LOw-Frequency ARray,” A&A, vol. 556, p. A2, Aug. 2013

2013
[12]

Cobalt: A GPU- based correlator and beamformer for LOFAR,

P. C. Broekema et al. , “Cobalt: A GPU- based correlator and beamformer for LOFAR,” Astronomy and Computing , vol. 23, pp. 180–192, 4 2018. [Online]. Available: https://linkinghub.elsevier. com/retrieve/pii/S2213133717301439

2018
[13]

A Medium-Scale Distributed System for Computer Science Research: Infrastructure for the Long Term,

H. Bal et al. , “A Medium-Scale Distributed System for Computer Science Research: Infrastructure for the Long Term,” Computer, vol. 49, no. 5, pp. 54–63, 2016

2016
[2021]

Available: https://www.sciencedirect

[Online]. Available: https://www.sciencedirect. com/science/article/pii/S0041624X20302444

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.

[1] [1]

Beamforming: a versatile approach to spatial filtering,

B. V . Veen and K. Buckley, “Beamforming: a versatile approach to spatial filtering,” IEEE ASSP Magazine, vol. 5, pp. 4–24, 4 1988. [Online]. Available: http://ieeexplore.ieee.org/document/665/

work page 1988

[2] [2]

The Apertif Radio Transient System (ARTS): Design, commissioning, data release, and detection of the first five fast radio bursts,

J. van Leeuwen et al. , “The Apertif Radio Transient System (ARTS): Design, commissioning, data release, and detection of the first five fast radio bursts,” A&A, vol. 672, p. A117, Apr. 2023

work page 2023

[3] [3]

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

V . Perrot, M. Polichetti, F. Varray, and D. Garcia, “So you think you can DAS? A viewpoint on delay-and- sum beamforming,” Ultrasonics, vol. 111, p. 106309,

work page

[4] [4]

The Tensor-Core Correlator,

J. W. Romein, “The Tensor-Core Correlator,” A&A, vol. 656, p. A52, Dec. 2021

work page 2021

[5] [5]

Benchmarking and Dissecting the Nvidia Hopper GPU Architecture,

W. Luo, R. Fan, Z. Li, D. Du, Q. Wang, and X. Chu, “Benchmarking and Dissecting the Nvidia Hopper GPU Architecture,” in 2024 IEEE International Parallel and Distributed Processing Symposium (IPDPS) . Los Alamitos, CA, USA: IEEE Computer Society, may 2024, pp. 656–667. [Online]. Available: https://doi.ieeecomputersociety.org/ 10.1109/IPDPS57955.2024.00064

arXiv 2024

[6] [6]

Kernel Tuner: A search-optimizing GPU code auto-tuner,

B. van Werkhoven, “Kernel Tuner: A search-optimizing GPU code auto-tuner,” Future Generation Computer Systems, vol. 90, pp. 347–358, 2019. [Online]. Available: https://www.sciencedirect.com/science/article/ pii/S0167739X18313359

work page 2019

[7] [7]

Bringing Auto-Tuning to HIP: Analysis of Tuning Impact and Difficulty on AMD and Nvidia GPUs,

M. Lurati, S. Heldens, A. Sclocco, and B. van Werkhoven, “Bringing Auto-Tuning to HIP: Analysis of Tuning Impact and Difficulty on AMD and Nvidia GPUs,” in Euro-Par 2024: Parallel Processing , J. Car- retero, S. Shende, J. Garcia-Blas, I. Brandic, K. Olcoz, and M. Schreiber, Eds. Cham: Springer Nature Switzer- land, 2024, pp. 91–106

work page 2024

[8] [8]

PMT: Power Measurement Toolkit,

S. Corda, B. Veenboer, and E. Tolley, “PMT: Power Measurement Toolkit,” in 2022 IEEE/ACM International Workshop on HPC User Support Tools (HUST) , 2022, pp. 44–47

work page 2022