Search NASA⌕ Search

SEARCH · Search NASA

Results for “sparse”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 1,351 records · Page 75

AWB-GCN: A Graph Convolutional Network Accelerator with Runtime Workload Rebalancing

The recent development of deep learning has been mostly focusing on Euclidean data, such as images, videos, audios, etc. However, most real-world information and relation are often expressed as graphs. To efficiently learn from graph data, graph convolutional networks (GCNs) emerge as a promising approach, showing advantages in several practical applications such as social network analysis, knowledge discovery, 3D modeling, motion capturing, etc. Real-world graphs are usually extremely large and imbalanced, posting significant performance demand and design challenges on the hardware dedicated for GCN inference. In this paper, we propose an architecture design called UW-GCN to accelerate graph convolutional network inference. To tackle the major performance bottleneck from workload imbalance, we propose dynamic neighborhood stealing and remote chunk shuffling techniques, relying on hardware flexibility to achieve hardware auto-tuning under negligible area or delay overhead. Specifically, UW-GCN is able to smartly profile the sparse graph pattern while continuously adjusting the workload distribution via routing reconfiguration among parallel processing elements (PEs). The ideal configuration is then reused in the remaining iterations. To the best of our knowledge, this is the first accelerator design particularly for GCN and the first work relying on hardware auto-tuning, which is normally based on software, to achieve near-optimal workload balance in processing sparse structures.

Geng, Tong↗

Automatic Code Generation for High-Performance Graph Algorithms

Graph problems are common across fields of scientific computing and social sciences. However, despite their importance, implementing graph algorithms effectively on modern computing systems is a challenging task that requires significant programming effort and generally results in customized implementations. Current computing and memory hierarchies are not architected for irregular computations resulting in challenges for graph algorithms to achieve high performance on those architectures. In this paper, we present GraphX, a novel compiler framework and DSL designed to simplify the development of efficient graph algorithms and achieve high performance on modern computing systems. GraphX consists of a DSL for efficient implementation of graph algorithms, various optimizations, such as support for sparse linear algebra and workspace transformations, optimized graph primitives, including semiring and masking, and a high-performance code generation engine. Using GraphX, users can implement graph algorithms using a semantically-rich language with graph-oriented operators. GraphX uses these semantics to automatically generate efficient code for target architectures, increasing performance and portability across architectures. The composable nature of GraphX makes it possible to extend the set of optimizations and architectures without modifying the source code. We demonstrate GraphX outperforms state-of-the-art graph libraries, such as LAGraph, up to $3.7 speedup in semiring operations, $2.19 speedup in an important sparse computational kernel, and $9.05 speedup in graph processing algorithms.

compiler, graph algorithms, semiring, masking, wor↗

Analyzing the Impacts of Increased Wind Power on Generation Revenue Sufficiency

The Revenue Sufficiency Guarantee (RSG), as part of make-whole (or uplift) payments in electricity markets, is designed to recover the generation resources' offer-based production costs that are not otherwise covered by their market revenues. Increased penetrations of wind power will bring significant impacts to the RSG payments in the markets. However, literature related to this topic is sparse. This paper first reviews the industrial practices of implementing RSG in major U.S. independent system operators (ISOs) and regional transmission operators (RTOs) and then develops a general RSG calculation method. Finally, an 18-bus test system is adopted to demonstrate the impacts of increased wind power on RSG payments.

17 WIND ENERGY↗

High performance radiation transport simulations: Preparing for TITAN

In this paper we describe the Denovo code system. Denovo solves the six-dimensional, steady-state, linear Boltzmann transport equation, of central importance to nuclear technology applications such as reactor core analysis (neutronics), radiation shielding, nuclear forensics and radiation detection. The code features multiple spatial differencing schemes, state-of-the-art linear solvers, the Koch-Baker-Alcouffe (KBA) parallel-wavefront sweep algorithm for inverting the transport operator, a new multilevel energy decomposition method scaling to hundreds of thousands of processing cores, and a modern, novel code architecture that supports straightforward integration of new features. In this paper we discuss the performance of Denovo on the 20+ petaflop ORNL GPU-based system, Titan. We describe algorithms and techniques used to exploit the capabilities of Titan's heterogeneous compute node architecture and the challenges of obtaining good parallel performance for this sparse hyperbolic PDE solver containing inherently sequential computations. Numerical results demonstrating Denovo performance on early Titan hardware are presented.

Baker, C.↗

JACC: Leveraging HPC Meta-Programming and Performance Portability with the Just-in-Time and LLVM-based Julia Language

We present JACC (Julia for Accelerators), the first high-level, and performance-portable model for the just-in-time and LLVM-based Julia language. JACC provides a unified and lightweight front end across different back ends available in Julia, enabling the same Julia code to run efficiently on many HPC CPU and GPU targets. We evaluated the performance of JACC for common HPC kernels as well as for the most computationally demanding kernels used in applications, HPCCG, a supercomputing benchmark test for sparse domains, and HARVEY, a blood flow simulator to assist in the diagnosis and treatment of patients suffering from vascular diseases. We carried out the performance analysis on the most advanced US DOE supercomputers: Aurora, Frontier, and Perlmutter. Overall, we show that JACC has a negligible overhead versus vendor-specific solutions, reporting GPU speedups with no extra cost to programmability.

Valero-Lara, Pedro↗

SGD-Net: Efficient Model-Based Deep Learning with Theoretical Guarantees

Deep unfolding networks have recently gained popularity for solving imaging inverse problems. However, the computational and memory complexity of data-consistency layers within traditional deep unfolding networks scales with the number of measurements, limiting their applicability to large-scale imaging inverse problems. We propose SGD-Net as a new methodology for improving the efficiency of deep unfolding through stochastic approximations of the data-consistency layers. Our theoretical analysis shows that SGD-Net can be trained to approximate batch deep unfolding networks to an arbitrary precision. Our simulations on intensity diffraction tomography and sparse-view computed tomography show that SGD-Net can match the performance of the traditional batch network at a fraction of training and testing complexity.Deep unfolding networks have recently gained popularity for solving imaging inverse problems. However, the computational and memory complexity of data-consistency layers within traditional deep unfolding networks scales with the number of measurements, limiting their applicability to large-scale imaging inverse problems. We propose SGD-Net as a new methodology for improving the efficiency of deep unfolding through stochastic approximations of the data-consistency layers. Our theoretical analysis shows that SGD-Net can be trained to approximate batch deep unfolding networks to an arbitrary precision. Our simulations on intensity diffraction tomography and sparse-view computed tomography show that SGD-Net can match the performance of the traditional batch network at a fraction of training and testing complexity.

97 MATHEMATICS AND COMPUTING↗

Geometric Mapping of Tasks to Processors on Parallel Computers with Mesh or Torus Networks

We present a new method for reducing parallel applications' communication time by mapping their MPI tasks to processors in a way that lowers the distance messages travel and the amount of congestion in the network. Assuming geometric proximity among the tasks is a good approximation of their communication interdependence, we use a geometric partitioning algorithm to order both the tasks and the processors, assigning task parts to the corresponding processor parts. In this way, interdependent tasks are assigned to “nearby” cores in the network. We also present a number of algorithmic optimizations that exploit specific features of the network or application to further improve the quality of the mapping. We specifically address the case of sparse node allocation, where the nodes assigned to a job are not necessarily located in a contiguous block nor within close proximity to each other in the network. Furthermore, our methods generalize to contiguous allocations as well, and results are shown for both contiguous and non-contiguous allocations. We show that, for the structured finite difference mini-application MiniGhost, our mapping methods reduced communication time up to 75% relative to MiniGhost's default mapping on 128K cores of a Cray XK7 with sparse allocation. For the atmospheric modeling code E3SM/HOMME, our methods reduced communication time up to 31% on 16K cores of an IBM BlueGene/Q with contiguous allocation.

97 MATHEMATICS AND COMPUTING↗

A Space-Filling Visualization Technique for Multivariate Small World Graphs

We introduce an information visualization technique, known as GreenCurve, for large sparse graphs that exhibit small world properties. Our fractal-based design approach uses spatial cues to approximate the node connections and thus eliminates the links between the nodes in the visualization. The paper describes a sophisticated algorithm to order the neighboring nodes of a large sparse graph by solving the Fiedler vector of its graph Laplacian, and then fold the graph nodes into a space-filling fractal curve based on the Fiedler vector. The result is a highly compact visualization that gives a succinct overview of the graph with guaranteed visibility of every graph node. We show in the paper that the GreenCurve technology is (1) theoretically sustainable by introducing an error estimation metric to measure the fidelity of the new graph representation, (2) empirically rigorous by conducting a usability study to investigate its strengths and weaknesses against the traditional graph layout, and (3) pragmatically feasible by applying it to analyze stressed conditions of the large scale electric power grid on the west coast.

99 GENERAL AND MISCELLANEOUS↗

Soil and fertilizer amendments and edge effects on the floral succession of pulverized fuel ash

Plots of fresh pulverized fuel ash (PFA, an industrial waste) were inoculated with soils from existing PFA sites and fertilizers in a factorial design, then left unmanaged for 12 years during which time the floral development and soil chemistry were monitored annually. For the first 3 years, the site supported a sparse mix of chenopods (including the scarce Chenopodium glaucum) and halophytes. As salinity declined, ruderals, legumes, and grasses plus the fire-site moss Funaria hygrometrica colonized, followed by Festuca arundinacea grassland (NVC community MG12) and Hippophae rhamnoides scrub. Dactylorhiza incarnata (orchidacea) appeared after 7 years, but only in plots that had received soil from existing orchid colonies. Four years later, a larger second generation of Dactylorhiza appeared, but only in the central zone of the site where vegetation was thinnest. By year 12, the site was dominated by coarse grasses and scrub, with early successional species persisting only in the sparsely vegetated center, where nitrate levels were lowest. This edge effect is interpreted as centripetal encroachment, a process of potentially wider concern for the conservation of low-fertility habitat patches. Overall, seed bank inoculation seems to have introduced few but desirable species (D. incarnata, Pyrola rotundifolia, some halophytes, and annuals), whereas initial application of organic fertilizer had long-lasting ({ge} 10 years) effects on cover and soil composition.

01 COAL, LIGNITE, AND PEAT↗

Optimal test selection for prediction uncertainty reduction

Economic factors and experimental limitations often lead to sparse and/or imprecise data used for the calibration and validation of computational models. This paper addresses resource allocation for calibration and validation experiments, in order to maximize their effectiveness within given resource constraints. When observation data are used for model calibration, the quality of the inferred parameter descriptions is directly affected by the quality and quantity of the data. This paper characterizes parameter uncertainty within a probabilistic framework, which enables the uncertainty to be systematically reduced with additional data. The validation assessment is also uncertain in the presence of sparse and imprecise data; therefore, this paper proposes an approach for quantifying the resulting validation uncertainty. Since calibration and validation uncertainty affect the prediction of interest, the proposed framework explores the decision of cost versus importance of data in terms of the impact on the prediction uncertainty. Often, calibration and validation tests may be performed for different input scenarios, and this paper shows how the calibration and validation results from different conditions may be integrated into the prediction. Then, a constrained discrete optimization formulation that selects the number of tests of each type (calibration or validation at given input conditions) is proposed. Furthermore, the proposed test selection methodology is demonstrated on a microelectromechanical system (MEMS) example.

97 MATHEMATICS AND COMPUTING↗

Application of High Performance Computing for Studying Cyclic Variability in Dilute Internal Combustion Engines

Combustion instabilities in dilute internal combustion engines are manifest in cyclic variability (CV) in engine performance measures such as integrated heat release or shaft work. Understanding the factors leading to CV is important in model-based control, especially with high dilution where experimental studies have demonstrated that deterministic effects can become more prominent. Observation of enough consecutive engine cycles for significant statistical analysis is standard in experimental studies but is largely wanting in numerical simulations because of the computational time required to compute hundreds or thousands of consecutive cycles. We have proposed and begun implementation of an alternative approach to allow rapid simulation of long series of engine dynamics based on a low-dimensional mapping of ensembles of single-cycle simulations which map input parameters to output engine performance. This paper details the use Titan at the Oak Ridge Leadership Computing Facility to investigate CV in a gasoline direct-injected spark-ignited engine with a moderately high rate of dilution achieved through external exhaust gas recirculation. The CONVERGE™ CFD software was used to perform single-cycle simulations with imposed variations of operating parameters and boundary conditions selected according to a sparse grid sampling of the parameter space. Using an uncertainty quantification technique, the sampling scheme is chosen similar to a design of experiments grid but uses algorithms designed to minimize the number of samples required to achieve a desired degree of accuracy. The simulations map input parameters to output metrics of engine performance for a single cycle, and by mapping over a large parameter space, results can be interpolated from within that space. This interpolation scheme forms the basis for a low-dimensional ‘metamodel’ (or model of a model) which can be used to mimic the dynamical behavior of corresponding high-dimensional simulations. Simulations of high-EGR spark-ignition combustion cycles within a parametric sampling grid were performed and analyzed statistically, and sensitivities of the physical factors leading to high CV are presented. With these results, the prospect of producing low-dimensional metamodels to describe engine dynamics at any point in the parameter space will be discussed. Additionally, modifications to the methodology to account for nondeterministic effects in the numerical solution environment are proposed.

Finney, Charles E. A.↗

Remotely sensed physical property estimation from powder contaminated surfaces

It is the intent of this paper to demonstrate the veracity of a method to estimate the total solid contaminant mass present on a sparsely coated surface encompassing a pixel from longwave infrared spectra. Sparsely coated surfaces create complex radiometry due to the interactions of electromagnetic energy between intimate materials. Current algorithms can be used on intimate mixtures to identify the abundances of materials in a pixel, but they cannot provide additional property information. Radiative transfer models can create mixture signatures, but only with a set of well-characterized physical parameters that are typically not known or are difficult to retrieve. Finally, the approach described here creates a parameter inversion model from a radiative transfer model and uses empirically measured mixture data to retrieve physical characteristics of the contaminated surface and derive a total contaminant mass present.

97 MATHEMATICS AND COMPUTING↗

Evaluation of interpolation methods for surface-based motion compensated tomographic reconstruction for cardiac angiographic C-arm data

Purpose: For interventional cardiac procedures, anatomical and functional information about the cardiac chambers is of major interest. With the technology of angiographic C-arm systems it is possible to reconstruct intraprocedural three-dimensional (3D) images from 2D rotational angiographic projection data (C-arm CT). However, 3D reconstruction of a dynamic object is a fundamental problem in C-arm CT reconstruction. The 2D projections are acquired over a scan time of several seconds, thus the projection data show different states of the heart. A standard FDK reconstruction algorithm would use all acquired data for a filtered backprojection and result in a motion-blurred image. In this approach, a motion compensated reconstruction algorithm requiring knowledge of the 3D heart motion is used. The motion is estimated from a previously presented 3D dynamic surface model. This dynamic surface model results in a sparse motion vector field (MVF) defined at control points. In order to perform a motion compensated reconstruction, a dense motion vector field is required. The dense MVF is generated by interpolation of the sparse MVF. Therefore, the influence of different motion interpolation methods on the reconstructed image quality is evaluated. Methods: Four different interpolation methods, thin-plate splines (TPS), Shepard's method, a smoothed weighting function, and a simple averaging, were evaluated. The reconstruction quality was measured on phantom data, a porcine model as well as on in vivo clinical data sets. As a quality index, the 2D overlap of the forward projected motion compensated reconstructed ventricle and the segmented 2D ventricle blood pool was quantitatively measured with the Dice similarity coefficient and the mean deviation between extracted ventricle contours. For the phantom data set, the normalized root mean square error (nRMSE) and the universal quality index (UQI) were also evaluated in 3D image space. Results: The quantitative evaluation of all experiments showed that TPS interpolation provided the best results. The quantitative results in the phantom experiments showed comparable nRMSE of Almost-Equal-To 0.047 {+-} 0.004 for the TPS and Shepard's method. Only slightly inferior results for the smoothed weighting function and the linear approach were achieved. The UQI resulted in a value of Almost-Equal-To 99% for all four interpolation methods. On clinical human data sets, the best results were clearly obtained with the TPS interpolation. The mean contour deviation between the TPS reconstruction and the standard FDK reconstruction improved in the three human cases by 1.52, 1.34, and 1.55 mm. The Dice coefficient showed less sensitivity with respect to variations in the ventricle boundary. Conclusions: In this work, the influence of different motion interpolation methods on left ventricle motion compensated tomographic reconstructions was investigated. The best quantitative reconstruction results of a phantom, a porcine, and human clinical data sets were achieved with the TPS approach. In general, the framework of motion estimation using a surface model and motion interpolation to a dense MVF provides the ability for tomographic reconstruction using a motion compensation technique.

60 APPLIED LIFE SCIENCES↗

Improving IMRT delivery efficiency with reweighted L1-minimization for inverse planning

Purpose: This study presents an improved technique to further simplify the fluence-map in intensity modulated radiation therapy (IMRT) inverse planning, thereby reducing plan complexity and improving delivery efficiency, while maintaining the plan quality.Methods: First-order total-variation (TV) minimization (min.) based on L1-norm has been proposed to reduce the complexity of fluence-map in IMRT by generating sparse fluence-map variations. However, with stronger dose sparing to the critical structures, the inevitable increase in the fluence-map complexity can lead to inefficient dose delivery. Theoretically, L0-min. is the ideal solution for the sparse signal recovery problem, yet practically intractable due to its nonconvexity of the objective function. As an alternative, the authors use the iteratively reweighted L1-min. technique to incorporate the benefits of the L0-norm into the tractability of L1-min. The weight multiplied to each element is inversely related to the magnitude of the corresponding element, which is iteratively updated by the reweighting process. The proposed penalizing process combined with TV min. further improves sparsity in the fluence-map variations, hence ultimately enhancing the delivery efficiency. To validate the proposed method, this work compares three treatment plans obtained from quadratic min. (generally used in clinic IMRT), conventional TV min., and our proposed reweighted TV min. techniques, implemented by a large-scale L1-solver (template for first-order conic solver), for five patient clinical data. Criteria such as conformation number (CN), modulation index (MI), and estimated treatment time are employed to assess the relationship between the plan quality and delivery efficiency.Results: The proposed method yields simpler fluence-maps than the quadratic and conventional TV based techniques. To attain a given CN and dose sparing to the critical organs for 5 clinical cases, the proposed method reduces the number of segments by 10-15 and 30-35, relative to TV min. and quadratic min. based plans, while MIs decreases by about 20%-30% and 40%-60% over the plans by two existing techniques, respectively. With such conditions, the total treatment time of the plans obtained from our proposed method can be reduced by 12-30 s and 30-80 s mainly due to greatly shorter multileaf collimator (MLC) traveling time in IMRT step-and-shoot delivery.Conclusions: The reweighted L1-minimization technique provides a promising solution to simplify the fluence-map variations in IMRT inverse planning. It improves the delivery efficiency by reducing the entire segments and treatment time, while maintaining the plan quality in terms of target conformity and critical structure sparing.

61 RADIATION PROTECTION AND DOSIMETRY↗

SU-E-I-72: First Experimental Study of On-Board CBCT Imaging Using 2.5MV Beam On a Radiotherapy Linac

Purpose: Varian TrueBeam version 2.0 comes with a new inline 2.5MV beam modality for image guided patient setup. In this work we develop an iterative volumetric image reconstruction technique specific to the beam and investigate the possibility of obtaining metal artifact free CBCT images using the new imaging modality. Methods: An iterative reconstruction algorithm with a sparse representation constraint based on dictionary learning is developed, in which both sparse projection and low dose rate (10 MU/min) are considered. Two CBCT experiments were conducted using the newly available 2.5MV beam on a Varian TrueBeam linac. First, a Rando anthropomorphic head phantom with and without a copper bar inserted in the center was scanned using both 2.5MV and kV (100kVp) beams. In a second experiment, an MRI phantom with many coils was scanned using 2.5MV, 6MV, and kV (100kVp) beams. Imaging dose and the resultant image quality is studied. Results: Qualitative assessment suggests that there were no visually detectable metal artifacts in MV CBCT images, compared with significant metal artifacts in kV CBCT images, especially in the MRI phantom. For a region near the metal object in the head phantom, the 2.5MV CBCT gave a more accurate quantification of the electron density compared with kV CBCT, with a ∼50% reduction in mean HU error. As expected, the contrast between bone and soft-tissue in 2.5MV CBCT decreased compared with kV CBCT. Conclusion: On-board CBCT imaging with the new 2.5MV beam can effectively reduce metal artifacts, although with a reduced softtissue contrast. Combination of kV and MV scanning may lead to metal artifact free CBCT images with uncompromised soft-tissue contrast.

60 APPLIED LIFE SCIENCES↗

WE-G-18A-04: 3D Dictionary Learning Based Statistical Iterative Reconstruction for Low-Dose Cone Beam CT Imaging

Purpose: To develop a 3D dictionary learning based statistical reconstruction algorithm on graphic processing units (GPU), to improve the quality of low-dose cone beam CT (CBCT) imaging with high efficiency. Methods: A 3D dictionary containing 256 small volumes (atoms) of 3x3x3 voxels was trained from a high quality volume image. During reconstruction, we utilized a Cholesky decomposition based orthogonal matching pursuit algorithm to find a sparse representation on this dictionary basis of each patch in the reconstructed image, in order to regularize the image quality. To accelerate the time-consuming sparse coding in the 3D case, we implemented our algorithm in a parallel fashion by taking advantage of the tremendous computational power of GPU. Evaluations are performed based on a head-neck patient case. FDK reconstruction with full dataset of 364 projections is used as the reference. We compared the proposed 3D dictionary learning based method with a tight frame (TF) based one using a subset data of 121 projections. The image qualities under different resolutions in z-direction, with or without statistical weighting are also studied. Results: Compared to the TF-based CBCT reconstruction, our experiments indicated that 3D dictionary learning based CBCT reconstruction is able to recover finer structures, to remove more streaking artifacts, and is less susceptible to blocky artifacts. It is also observed that statistical reconstruction approach is sensitive to inconsistency between the forward and backward projection operations in parallel computing. Using high a spatial resolution along z direction helps improving the algorithm robustness. Conclusion: 3D dictionary learning based CBCT reconstruction algorithm is able to sense the structural information while suppressing noise, and hence to achieve high quality reconstruction. The GPU realization of the whole algorithm offers a significant efficiency enhancement, making this algorithm more feasible for potential clinical application. A high zresolution is preferred to stabilize statistical iterative reconstruction. This work was supported in part by NIH(1R01CA154747-01), NSFC((No. 61172163), Research Fund for the Doctoral Program of Higher Education of China (No. 20110201110011), China Scholarship Council.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗

Multi-GPU implementation of a VMAT treatment plan optimization algorithm

Purpose: Volumetric modulated arc therapy (VMAT) optimization is a computationally challenging problem due to its large data size, high degrees of freedom, and many hardware constraints. High-performance graphics processing units (GPUs) have been used to speed up the computations. However, GPU’s relatively small memory size cannot handle cases with a large dose-deposition coefficient (DDC) matrix in cases of, e.g., those with a large target size, multiple targets, multiple arcs, and/or small beamlet size. The main purpose of this paper is to report an implementation of a column-generation-based VMAT algorithm, previously developed in the authors’ group, on a multi-GPU platform to solve the memory limitation problem. While the column-generation-based VMAT algorithm has been previously developed, the GPU implementation details have not been reported. Hence, another purpose is to present detailed techniques employed for GPU implementation. The authors also would like to utilize this particular problem as an example problem to study the feasibility of using a multi-GPU platform to solve large-scale problems in medical physics. Methods: The column-generation approach generates VMAT apertures sequentially by solving a pricing problem (PP) and a master problem (MP) iteratively. In the authors’ method, the sparse DDC matrix is first stored on a CPU in coordinate list format (COO). On the GPU side, this matrix is split into four submatrices according to beam angles, which are stored on four GPUs in compressed sparse row format. Computation of beamlet price, the first step in PP, is accomplished using multi-GPUs. A fast inter-GPU data transfer scheme is accomplished using peer-to-peer access. The remaining steps of PP and MP problems are implemented on CPU or a single GPU due to their modest problem scale and computational loads. Barzilai and Borwein algorithm with a subspace step scheme is adopted here to solve the MP problem. A head and neck (H and N) cancer case is then used to validate the authors’ method. The authors also compare their multi-GPU implementation with three different single GPU implementation strategies, i.e., truncating DDC matrix (S1), repeatedly transferring DDC matrix between CPU and GPU (S2), and porting computations involving DDC matrix to CPU (S3), in terms of both plan quality and computational efficiency. Two more H and N patient cases and three prostate cases are used to demonstrate the advantages of the authors’ method. Results: The authors’ multi-GPU implementation can finish the optimization process within ∼1 min for the H and N patient case. S1 leads to an inferior plan quality although its total time was 10 s shorter than the multi-GPU implementation due to the reduced matrix size. S2 and S3 yield the same plan quality as the multi-GPU implementation but take ∼4 and ∼6 min, respectively. High computational efficiency was consistently achieved for the other five patient cases tested, with VMAT plans of clinically acceptable quality obtained within 23–46 s. Conversely, to obtain clinically comparable or acceptable plans for all six of these VMAT cases that the authors have tested in this paper, the optimization time needed in a commercial TPS system on CPU was found to be in an order of several minutes. Conclusions: The results demonstrate that the multi-GPU implementation of the authors’ column-generation-based VMAT optimization can handle the large-scale VMAT optimization problem efficiently without sacrificing plan quality. The authors’ study may serve as an example to shed some light on other large-scale medical physics problems that require multi-GPU techniques.

60 APPLIED LIFE SCIENCES↗