Computational Performance Research Articles

GPU-powered CFD-DEM framework for modelling large-scale gas–solid reacting flows (GPU- rCFD-DEM) and an industry application

The coupling of CFD (computational fluid dynamics) and DEM (discrete element method) is extensively used for simulating gas–solid reacting flows in various industrial processes, while its high computational cost limits its industry applications, especially large-scale systems with a large number of particles and complex geometries. This paper reports a robust highly efficient GPU (graphics processing unit) − powered CFD-DEM coupling approach that is, for the first time, capable of simulating large-scale gas–solid reacting flow systems with complex geometries (GPU- rCFD-DEM). The fluid flow calculations are performed using CPU parallelization, while the particle flow simulations leverage GPU parallelization, and the coupling calculations between CFD and DEM are fully implemented on GPU. The model includes advanced coupling strategies to enhance numerical stability, especially when handling complex geometries with unstructured CFD meshes. The developed model is validated through experimental measurements and its computational performance is evaluated by comparison with previous GPU-based simulations. It shows good agreement with the experiments and superior performance compared to the traditional coupling method. The model is then applied to simulate raceway dynamics and coke combustion in an industrial-scale blast furnace, showcasing its effectiveness in handling complex geometries and a huge number of particles in gas–solid reacting flows. This work provides an efficient and robust solution for numerically simulating industrial applications of gas–solid reacting flow systems.

Vertex Block Descent

We introduce vertex block descent, a block coordinate descent solution for the variational form of implicit Euler through vertex-level Gauss-Seidel iterations. It operates with local vertex position updates that achieve reductions in global variational energy with maximized parallelism. This forms a physics solver that can achieve numerical convergence with unconditional stability and exceptional computation performance. It can also fit in a given computation budget by simply limiting the iteration count while maintaining its stability and superior convergence rate. We present and evaluate our method in the context of elastic body dynamics, providing details of all essential components and showing that it outperforms alternative techniques. In addition, we discuss and show examples of how our method can be used for other simulation systems, including particle-based simulations and rigid bodies.

Robust Containment Queries over Collections of Rational Parametric Curves via Generalized Winding Numbers

Point containment queries for regions bound by watertight geometric surfaces, i.e., closed and without self-intersections, can be evaluated straightforwardly with a number of well-studied algorithms. When this assumption on domain geometry is not met, such methods are either unusable, or prone to misclassifications that can lead to cascading errors in downstream applications. More robust point classification schemes based on generalized winding numbers have been proposed, as they are indifferent to these imperfections. However, existing algorithms are limited to point clouds and collections of linear elements. We extend this methodology to encompass more general curved shapes with an algorithm that evaluates the winding number scalar field over unstructured collections of rational parametric curves. In particular, we evaluate the winding number for each curve independently, making the derived containment query robust to how the curves are arranged. We ensure geometric fidelity in our queries by treating each curve as equivalent to an adaptively constructed polyline that provably has the same generalized winding number at the point of interest. Our algorithm is numerically stable for points that are arbitrarily close to the model, and explicitly treats points that are coincident with curves. We demonstrate the improvements in computational performance granted by this method over conventional techniques as well as the robustness induced by its application.

GCMe: Efficient Implementation of the Gaussian Core Model with Smeared Electrostatic Interactions for Molecular Dynamics Simulations of Soft Matter Systems.

In recent years, molecular dynamics (MD) simulations have emerged as an essential tool for understanding the structure, dynamics, and phase behavior of charged soft matter systems. To explore phenomena across greater length and time scales in MD simulations, molecules are often coarse-grained for better computational performance. However, commonly used force fields represent particles as hard-core interaction centers with point charges, which often overemphasizes the packing effect and short-range electrostatics, especially in systems with bulky deformable organic molecules and systems with strong coarse-graining. This underscores the need for an efficient soft-core model to physically capture the effective interactions between coarse-grained particles. To this end, we implement a soft-core model uniting the Gaussian core model with smeared electrostatic interactions that is phenomenologically equivalent to recent theoretical models. We first parametrize it generically using water as the model solvent. Then, we benchmark its performance in the OpenMM toolkit for different boundary conditions to highlight a computational speedup of up to 34 × compared to commonly used force fields and existing implementations. Finally, we demonstrate its utility by investigating how boundary polarizability affects the adsorption behavior of a polyelectrolyte solution on perfectly conducting and nonmetal boundaries.

A Model-Free Approach for Solving Choice-Based Competitive Facility Location Problems Using Simulation and Submodularity

This paper considers facility location problems in which a firm entering a market seeks to open facilities on a subset of candidate locations so as to maximize its expected market share, assuming that customers choose the available alternative that maximizes a random utility function. We introduce a deterministic equivalent reformulation of this stochastic problem as a maximum covering location problem with an exponential number of demand points, each of which is covered by a different set of candidate locations. Estimating the prevalence of these preference profiles through simulation generalizes a sample average approximation method from the literature and results in a maximum covering location problem of manageable size. To solve it, we develop a partial Benders reformulation in which the contribution to the objective of the least influential preference profiles is aggregated and bounded by submodular cuts. This set of profiles is selected by a knee detection method that seeks to identify the best tradeoff between the fraction of the demand that is retained in the master problem and the size of the model. We develop a theoretical analysis of our approach and show that the solution quality it provides for the original stochastic problem, its computational performance, and the automatic profile-retention strategy it exploits are directly connected to the entropy of the preference profiles in the population. Computational experiments on existing and new benchmark sets indicate that our approach dominates the classical sample average approximation method on large instances of the competitive facility location problem, can outperform the best heuristic method from the literature under the multinomial logit model, and achieves state-of-the-art results under the mixed multinomial logit model. We characterize a broader class of problems, which includes assortment optimization, to which the solving methodology and the analyses developed in this paper can be extended. History: Accepted by Andrea Lodi, Area Editor for Design & Analysis of Algorithms—Discrete. Funding: This research was supported by Fonds de Recherche du Québec-Nature et Technologies and Institut de Valorisation des Données through scholarships to R. Legault. E. Frejinger was partially supported by the Canada Research Chair program [Grant 950-232244]. Supplemental Material: The online appendices are available at https://doi.org/10.1287/ijoc.2023.0280 .

GeCKO: user-friendly workflows for genotyping complex genomes using target enrichment capture. A use case on the large tetraploid durum wheat genome

BackgroundGenotyping of individuals plays a pivotal role in various biological analyses, with technology choice influenced by multiple factors including genomic constraints, number of targeted loci and individuals, cost considerations, and the ease of sample preparation and data processing. Target enrichment capture of specific polymorphic regions has emerged as a flexible and cost-effective genomic reduction method for genotyping, especially adapted to the case of very large genomes. However, this approach necessitates complex bioinformatics treatment to extract genotyping data from raw reads. Existing workflows predominantly cater to phylogenetic inference, leaving a gap in user-friendly tools for genotyping analysis based on capture methods. In response to these challenges, we have developed GeCKO (Genotyping Complexity Knocked-Out). To assess the effectiveness of combining target enrichment capture with GeCKO, we conducted a case study on durum wheat domestication history, involving sequencing, processing, and analyzing variants in four relevant durum wheat groups.ResultsGeCKO encompasses four distinct workflows, each designed for specific steps of genomic data processing: (i) read demultiplexing and trimming for data cleaning, (ii) read mapping to align sequences to a reference genome, (iii) variant calling to identify genetic variants, and (iv) variant filtering. Each workflow in GeCKO can be easily configured and is executable across diverse computational environments. The workflows generate comprehensive HTML reports including key summary statistics and illustrative graphs, ensuring traceable, reproducible results and facilitating straightforward quality assessment. A specific innovation within GeCKO is its 'targeted remapping' feature, specifically designed for efficient treatment of targeted enrichment capture data. This process consists of extracting reads mapped to the targeted regions, constructing a smaller sub-reference genome, and remapping the reads to this sub-reference, thereby enhancing the efficiency of subsequent steps.ConclusionsThe case study results showed the expected intra-group diversity and inter-group differentiation levels, confirming the method's effectiveness for genotyping and analyzing genetic diversity in species with complex genomes. GeCKO streamlined the data processing, significantly improving computational performance and efficiency. The targeted remapping enabled straightforward SNP calling in durum wheat, a task otherwise complicated by the species' large genome size. This illustrates its potential applications in various biological research contexts.

Accurate numerical prototypes of microfluidic droplet generators with open source tools

Droplet-based microfluidics gained significant attention for its high technological impact in various fields like (bio)analysis and (bio)synthesis. Precise and controlled droplet size is critical, for the encapsulated products, or the yield of chemical reactions. In a broad range of experimental parameters, the understanding of how droplets form, interact and move with accurate predictive models is crucial. In this work, numerical prototypes of droplet generators were made with Basilisk, an open source software for solving partial differential equations on adaptive Cartesian meshes including grid adaptation and scalability for High-Performance Computing (HPC). This research aims to analyze and compare the obtained droplets against existing experimental data. The evaluation involves qualitative and quantitative comparisons, considering various channel geometries, flow rates, and rheological conditions. The validation of the proposed tool in terms of accuracy and computational performance, enable us to offer to the microfluidics community a reliable tool to design and optimize droplet generators.

Censored Least Squares for Imputing Missing Values in PARAFAC Tensor Factorization.

Tensor factorization is a dimensionality reduction method applied to multidimensional arrays. These methods are useful for identifying patterns within a variety of biomedical datasets due to their ability to preserve the organizational structure of experiments and therefore aid in generating meaningful insights. However, missing data in the datasets being analyzed can impose challenges. Tensor factorization can be performed with some level of missing data and reconstruct a complete tensor. However, while tensor methods may impute these missing values, the choice of fitting algorithm may influence the fidelity of these imputations. Previous approaches, based on alternating least squares with prefilled values or direct optimization, suffer from introduced bias or slow computational performance. In this study, we propose that censored least squares can better handle missing values with data structured in tensor form. We ran censored least squares on four different biological datasets and compared its performance against alternating least squares with prefilled values and direct optimization. We used the error of imputation and the ability to infer masked values to benchmark their missing data performance. Censored least squares appeared best suited for the analysis of high-dimensional biological data by accuracy and convergence metrics across several studies.

Portability and scalability evaluation of large-scale statistical modeling and prediction software through HPC-ready containers

HPC-based applications often have complex workflows with many software dependencies that hinder their portability on contemporary HPC architectures. In addition, these applications often require extraordinary efforts to deploy and execute at performance potential on new HPC systems, while the users expert in these applications generally have less expertise in HPC and related technologies. This paper provides a dynamic solution that facilitates containerization for transferring HPC software onto diverse parallel systems. The study relies on the HPC Workflow as a Service (HPCWaaS) paradigm proposed by the EuroHPC eFlows4HPC project. It offers to deploy workflows through containers tailored for any of a number of specific HPC systems. Traditional container image creation tools rely on OS system packages compiled for generic architecture families (x86_64, amd64, ppc64, …) and specific MPI or GPU runtime library versions. The containerization solution proposed in this paper leverages HPC Builders such as Spack or Easybuild and multi-platform builders such as buildx to create a service for automating the creation of container images for the software specific to each hardware architecture, aiming to sustain the overall performance of the software. We assess the efficiency of our proposed solution for porting the geostatistics ExaGeoStat software on various parallel systems while preserving the computational performance. The results show that the performance of the generated images is comparable with the native execution of the software on the same architectures. On the distributed-memory system, the containerized version can scale up to 256 nodes without impacting performance.

RoGeR v3.0.5 – a process-based hydrological toolbox model in Python

Abstract. Although water availability and water quality are equally important for effective water resources management at various spatial and temporal scales, to date, a combined representation of soil water balance components and water quality components in Python is not available. The new RoGeR toolbox contains models that can be used for not only the quantification of hydrological processes, fluxes and stores, but also solute transport processes based on StorAge selection (SAS). This study presents the code structure and functionalities of RoGeR developed as a scientific model toolbox following defined open-source software guidelines. RoGeR uses five different computational backends covering just-in-time compilation, parallelism and graphical processing units (GPUs) that might be used for optimizing computational performance. We show that graphical processing unit computing has the greatest potential to improve computation time and energy usage, especially for large modelling experiments. A simple modelling experiment highlights the capabilities of the new RoGeR model toolbox. We simulated the soil water balance, stable water isotope (18O) transport, and corresponding travel time distributions of the Eberbaechle catchment, Germany, for a 3-year period. Due to the current limitations of a variety of process components, further development of RoGeR as a scientific software is needed. Future modifications are easily possible due to the open software architecture of RoGeR.

Academic achievement in Ugandan children with sickle cell anaemia: A cross-sectional study.

Academic achievement in school-age children is crucial for advancing learning goals. Children with sickle cell anaemia (SCA) in Sub-Saharan Africa may be at risk of disease-associated school difficulties. Limited data exist on the academic achievement of children with SCA in the region. This study aimed to assess academic achievement of children with SCA in Uganda compared to siblings without SCA. A cross-sectional study conducted at Mulago Hospital SCA Clinic in Uganda. School-going children (6-12 years) with SCA and age-matched sibling controls without SCA. Academic achievement was tested using the Wide Range Achievement Test, Fourth Edition (WRAT4). Outcome measures were spelling, mathematical computation, word reading, and sentence comprehension by age-normalized Z-scores on the WRAT4 test. Among 68 SCA and 69 control, the mean age (standard deviation) was 9.44 (2.04) and 9.42 (2.02) years and males were 55.9% and 46.4% respectively. Mean haemoglobin was 7.9 (SD 0.89)g/dL in the SCA group versus 12.8 (SD 0.89)g/dL in the controls, (p<0.001). Children with SCA scored lower in spelling, (mean difference [95% confidence interval] - 0.36 [-0.02 to -0.69], p=0.04) and mathematical computation, (mean difference [95% confidence interval] -0.51 [-0.17 to -0.85], p=0.003) than the controls. In the SCA group, lower scores in spelling correlated with age, while males performed better than females in mathematical computation. School-aged children with SCA are at risk of poor performance in spelling and mathematical computation. Our findings support the need for educational evaluation and possible support, especially in these two areas.

Enhancing computational efficiency in 3-D seismic modelling with half-precision floating-point numbers based on the curvilinear grid finite-difference method

SUMMARY Large-scale and high-resolution seismic modelling are very significant to simulating seismic waves, evaluating earthquake hazards and advancing exploration seismology. However, achieving high-resolution seismic modelling requires substantial computing and storage resources, resulting in a considerable computational cost. To enhance computational efficiency and performance, recent heterogeneous computing platforms, such as Nvidia Graphics Processing Units (GPUs), natively support half-precision floating-point numbers (FP16). FP16 operations can provide faster calculation speed, lower storage requirements and greater performance enhancement over single-precision floating-point numbers (FP32), thus providing significant benefits for seismic modelling. Nevertheless, the inherent limitation of fewer 16-bit representations in FP16 may lead to severe numerical overflow, underflow and floating-point errors during computation. In this study, to ensure stable wave equation solutions and minimize the floating-point errors, we use a scaling strategy to adjust the computation of FP16 arithmetic operations. For optimal GPU floating-point performance, we implement a 2-way single instruction multiple data (SIMD) within the floating-point units (FPUs) of CUDA cores. Moreover, we implement an earthquake simulation solver for FP16 operations based on curvilinear grid finite-difference method (CGFDM) and perform several earthquake simulations. Comparing the results of wavefield data with the standard CGFDM using FP32, the errors introduced by FP16 are minimal, demonstrating excellent consistency with the FP32 results. Performance analysis indicates that FP16 seismic modelling exhibits a remarkable improvement in computational efficiency, achieving a speedup of approximately 1.75 and reducing memory usage by half compared to the FP32 version.

Rapid Alzheimer's Disease Diagnosis Using Advanced Artificial Intelligence Algorithms

Alzheimer's disease (AD) is a leading cause of dementia, predominantly impacting the elderly and characterized by progressive cognitive decline. Early and precise detection is critical for effective management and improved patient outcomes. Traditional diagnostic methods such as neuroimaging and cerebrospinal fluid analysis are often invasive, expensive, and time- consuming. Advances in artificial intelligence (AI) and machine learning (ML) provide promising alternatives that are non-invasive, efficient, and cost-effective. This study explores the application of various ML algorithms to predict Alzheimer's disease. The methodology involved data preprocessing and feature selection using the Spearman algorithm to enhance computational efficiency and model performance. We evaluated k-Nearest Neighbors (k-NN), Naive Bayes (NB), Decision Trees (DT), and Ensemble methods. Results indicate that the Ensemble method achieved a predictive accuracy of 94.07% using only 13 features. These results demonstrate the potential of ML algorithms in revolutionizing AD diagnostics, offering scalable and accurate solutions for early detection.

FedSHE: privacy preserving and efficient federated learning with adaptive segmented CKKS homomorphic encryption

Unprotected gradient exchange in federated learning (FL) systems may lead to gradient leakage-related attacks. CKKS is a promising approximate homomorphic encryption scheme to protect gradients, owing to its unique capability of performing operations directly on ciphertexts. However, configuring CKKS security parameters involves a trade-off between correctness, efficiency, and security. An evaluation gap exists regarding how these parameters impact computational performance. Additionally, the maximum vector length that CKKS can once encrypt, recommended by Homomorphic Encryption Standardization, is 16384, hampers its widespread adoption in FL when encrypting layers with numerous neurons. To protect gradients’ privacy in FL systems while maintaining practical performance, we comprehensively analyze the influence of security parameters such as polynomial modulus degree and coefficient modulus on homomorphic operations. Derived from our evaluation findings, we provide a method for selecting the optimal multiplication depth while meeting operational requirements. Then, we introduce an adaptive segmented encryption method tailored for CKKS, circumventing its encryption length constraint and enhancing its processing ability to encrypt neural network models. Finally, we present FedSHE, a privacy-preserving and efficient Federated learning scheme with adaptive Segmented CKKS Homomorphic Encryption. FedSHE is implemented on top of the federated averaging (FedAvg) algorithm and is available at https://github.com/yooopan/FedSHE. Our evaluation results affirm the correctness and effectiveness of our proposed method, demonstrating that FedSHE outperforms existing homomorphic encryption-based federated learning research efforts in terms of model accuracy, computational efficiency, communication cost, and security level.

DEVELOPMENT OF MODELS COMPLEX OSCILLATIONS FOR AUTOMATION SYSTEMS OF VIBROACOUSTIC OPERATIONAL SAFETY CONTROL OF BUILDINGS AND STRUCTURES

Problem statement. Long-term safe operation of buildings and structures is impossible without determining the stability of their load-bearing elements, especially after shocks and dynamic loads under the influence of explosions and fires. In such conditions, reliable and productive monitoring of the reliability of buildings and structures is a critical element in increasing human safety. Vibroacoustic is an effective and informative method that allows for non-destructive assessment of the condition of concrete, reinforced concrete, brick, multi-layer walls and other types of structures. However, to increase the speed and quality of determining the state of objects, maximum automation of the vibroacoustic method is required. Purpose of the study. Increasing the productivity and reliability of vibroacoustic monitoring of buildings and structures by automating the impact action with specified parameters based on the development of dynamic models of complex oscillations. Methods. Analytical studies of dynamic and kinematic models of the exciter of mechanical vibrations, computer modeling, laboratory tests of the vibration generator control system. Research results. Models of complex oscillations have been developed to improve reliability and implement new control laws that are impossible for classical vibration systems and are necessary for automated vibroacoustic control of buildings and structures and ensure the safety of their operation. The use of reduced bit depth coefficients in the calculation scheme using the Jack W. Crenshaw algorithm is justified in a wider range of argument values than with initial coefficient values for controlling vibration systems with limited computational performance. This is necessary when creating executive bodies of modern systems for automated vibroacoustic monitoring of the cracks and violations of the homogeneity of load-bearing structures of buildings and structures. The oscillation exciter has been tested in laboratory conditions. Scientific novelty. New methods and algorithms for automatic control of vibration exciters have been developed to obtain polyfrequency oscillations, linear waves and wave fields with specified amplitude and frequency characteristics. Practical significance. An oscillation exciter has been developed for a system of automated vibroacoustic control of the buildings and structures operation safety. This provides a necessary and sufficient basis for improving methods for assessing changes in the environment as a result of the appearance of cracks and loss of homogeneity.

Effect of impeller rotational phase on the FDA blood pump velocity fields.

The Food and Drug Administration (FDA) blood pump is an open-source benchmark cardiovascular device introduced for validating computational and experimental performance analysis tools. The time-resolved velocity field for the whole impeller has not been established, as is undertaken in this particle image velocimetry (PIV) study. The level of instantaneous velocity fluctuations is important, to assess the flow-induced rotor vibrations which may contribute to the total blood damage. To document these factors, time-resolved two-dimensional PIV experiments were performed that were precisely phase-locked with the impeller rotation angle. The velocity fields in the impeller and in the volute conformed with the previous single blade passage experiments of literature. Depending on the impeller orientation, present experiments showed that volute outlet nozzle flow can fluctuate up to 34% during impeller rotation, with a maximum standard experimental uncertainty of 2.2%. Likewise, the flow fields in each impeller passage also altered in average 33.5%. Considerably different vortex patterns were observed for different blade passages, with the largest vortical structures reaching an average core radii of 7 mm. The constant volute area employed in the FDA pump design contributes to the observed velocity imbalance, as illustrated in our velocity measurements. By introducing the impeller orientation parameter for the nozzle flow, this study considers the possible uncertainties influencing pump flow. Expanding the available literature data, analysis of inter-blade relative velocity fields is provided here for the first-time to the best of our knowledge. Consequently, our research fills a critical knowledge gap in the understanding of the flow dynamics of an important benchmark cardiovascular device. This study prompts the need for improved hydrodynamic designs and optimized devices to be used as benchmark test devices, to build more confidence and safety in future ventricular assist device performance assessment studies.

Computational imaging-based single-lens imaging systems and performance evaluation

The minimalist optical system has a simple structure, small size, and lightweight, but the low optical complexity will produce optical aberration. Addressing the significant aberration degradation in minimalist systems, we propose a high-quality computational optical framework. This framework integrates a global point spread function (PSF) change imaging model with a transformer-based U-Net deep learning algorithm to achieve high-quality imaging in minimalist systems. Additionally, we introduce an imaging performance evaluation method based on the modulation transfer degree of resolution (MTR). We addressed severe chromatic and spherical aberrations in single-lens systems, a typical example of minimalist optical systems, by simulating the degradation process and reconstructing the imaging effects. This approach demonstrated significant improvements, thus validating the feasibility of our method. Specifically, our technique calculated the MTR values in real images captured with the GCL010109 single lens at 0.8085, and with the GCL010110 single lens at 0.8055. Our method enhanced the imaging performance of minimalist systems by 4 times, upgrading minimalist system capabilities from poor to good lens grade. This work can provide reference for wavefront coding, matelens, diffraction optical systems, and other computational imaging work. It can also promote the application of miniaturization of medical, aerospace, and head-mounted optical systems.

Optimal design of electromagnetic repulsion mechanism based on hybrid infill criterion and kriging surrogate model

Abstract In the optimization design of the electromagnetic repulsion mechanism in finite element simulation, there are problems of high computational cost and long optimization time. To shorten the development cycle of the prototype and improve the efficiency of optimization, an optimization algorithm based on the Hybrid Infill Criterion and Kriging Surrogate Model (HIC-KSM) is proposed by organically combining the Expected Improvement (EI) infill criterion and the Pareto front infill criterion to make full use of the performance of parallel computation and to accelerate the convergence speed of the algorithm. The proposed algorithm is compared with the other optimization algorithms, the optimization results of five multi-objective test functions are compared to verify the efficiency of the proposed algorithm, and the algorithm is applied to the parameter optimization of the electromagnetic repulsion mechanism.

Multi-layer perceptron hyperparameter optimization using Jaya algorithm for disease classification

This study introduces an innovative hyperparameter optimization approach for enhancing multilayer perceptrons (MLP) using the Jaya algorithm. Addressing the crucial role of hyperparameter tuning in MLP’s performance, the Jaya algorithm, inspired by social behavior, emerges as a promising optimization technique without algorithm-specific parameters. Systematic application of Jaya dynamically adjusts hyperparameter values, leading to notable improvements in convergence speeds and model generalization. Quantitatively, the Jaya algorithm consistently achieves convergences at first iteration, faster convergence compared to conventional methods, resulting in 7% higher accuracy levels on several datasets. This research contributes to hyperparameter optimization, offering a practical and effective solution for optimizing MLP in diverse applications, with implications for improved computational efficiency and model performance.

Energy-Performance Assessment of Oscillatory Neural Networks Based on VO2 Devices for Future Edge AI Computing.

Oscillatory neural network (ONN) is an emerging neuromorphic architecture composed of oscillators that implement neurons and are coupled by synapses. ONNs exhibit rich dynamics and associative properties, which can be used to solve problems in the analog domain according to the paradigm let physics compute. For example, compact oscillators made of VO2 material are good candidates for building low-power ONN architectures dedicated to AI applications at the edge, like pattern recognition. However, little is known about the ONN scalability and its performance when implemented in hardware. Before deploying ONN, it is necessary to assess its computation time, energy consumption, performance, and accuracy for a given application. Here, we consider a VO2-oscillator as an ONN building block and perform circuit-level simulations to evaluate the ONN performances at the architecture level. Notably, we investigate how the ONN computation time, energy, and memory capacity scale with the number of oscillators. It appears that the ONN energy grows linearly when scaling up the network, making it suitable for large-scale integration at the edge. Furthermore, we investigate the design knobs for minimizing the ONN energy. Assisted by technology computer-aided design (TCAD) simulations, we report on scaling down the dimensions of VO2 devices in crossbar (CB) geometry to decrease the oscillator voltage and energy. We benchmark ONN versus state-of-the-art architectures and observe that the ONN paradigm is a competitive energy-efficient solution for scaled VO2 devices oscillating above 100 MHz. Finally, we present how ONN can efficiently detect edges in images captured on low-power edge devices and compare the results with Sobel and Canny edge detectors.