Search Results (640)

The co-circulation of respiratory viruses, including SARS-CoV-2, influenza A/B, and respiratory syncytial virus (RSV), represents a significant global health challenge that requires rapid, accurate, and differential diagnosis to support infection control and appropriate clinical decision-making. This narrative review summarizes emerging multimodal point-of-care testing (POCT) strategies for the detection and management of these respiratory viruses. Relevant studies were identified through literature searches of major scientific databases, including PubMed, Scopus, and Web of Science, focusing on recent advances in molecular diagnostics, biosensors, microfluidics, and digital health technologies. To improve clinical interpretation and comparative assessment, current POCT platforms were organized into four operational tiers based on infrastructure dependence, degree of portability, and level of decentralization of testing. Tier 1 (Professional Clinical Systems) includes fully integrated automated molecular diagnostic platforms designed for use in hospital and emergency care settings. Tier 2 (Field-Deployable Systems) comprises portable molecular and isothermal amplification technologies designed for use in decentralized or resource-limited environments. Tier 3 (Hardware-Lite Assays) includes simplified diagnostic approaches that minimize instrument requirements and are suitable for near-patient or low-infrastructure settings. Tier 4 (Consumer-Digital Diagnostics) encompasses emerging smartphone- and IoT-integrated diagnostic platforms that support user-driven testing and digital health connectivity. This tier-based framework reflects a proposed stratification of POCT technologies along a decentralization continuum and aims to facilitate comparison and selection of diagnostic strategies across diverse healthcare settings. Full article

Accurate, automated delineation of adult diffuse gliomas from multi-parametric magnetic resonance imaging (mpMRI) is central to quantitative neuro-oncology. Volumetric 3D networks dominate the BraTS leaderboard but require expensive GPUs, long training cycles, and provide diminishing returns relative to their compute budget. Slice-wise 2D models, by contrast, discard inter-slice context that is informative for thin tumor rims and small enhancing foci. We introduce GDNet, a 2.5D multimodal MRI segmentation framework for adult glioma evaluated on the BraTS 2024 cohort. GDNet consumes a stack of three adjacent axial slices from the four standard BraTS modalities (T1, T1ce, T2, FLAIR) as a 12-channel input to a compact U-shaped encoder–decoder with Group Normalization and predicts whole tumor (WT), tumor core (TC), and enhancing tumor (ET) masks for the central slice. The training pipeline pairs the 2.5D backbone with: (i) Exponential Moving Average (EMA) of model weights with decay 0.999, (ii) mixed tumor-aware slice sampling (p_tumor = 0.50), (iii) a compound Cross-Entropy + Soft-Dice loss, and (iv) AdamW with warm-up plus cosine annealing under Automatic Mixed Precision. We performed a systematic, step-by-step ablation covering a 2D baseline, EMA + mixed sampling, tumor-centered crop fine-tuning, a GDNet-inspired architectural integration, a region-aware loss, 3-slice and 5-slice 2.5D inputs, and connected-component post-processing, and we report multi-seed results to quantify reproducibility. On the held-out BraTS 2024 test partition, the final 3-slice 2.5D GDNet achieved positive-only Dice scores of 0.791 ± 0.000 (WT), 0.736 ± 0.003 (TC), 0.654 ± 0.004 (ET), and a mean foreground positive-only Dice of 0.820 ± 0.000 across seeds; the all-slice mean foreground Dice exceeded 0.927 ± 0.000. Validation positive-only scores were 0.805 ± 0.002 (WT), 0.757 ± 0.004 (TC), 0.683 ± 0.009 (ET). The inter-seed standard deviation was small for every region (≤0.01 Dice points), indicating low inter-seed variance across the two seeds evaluated; with only two seeds, we regard this as preliminary evidence of training stability rather than a strong reproducibility claim. The ablation isolated EMA + mixed tumor sampling and the 2.5D context window as the dominant sources of improvement; notably, a GDNet-style architectural integration with a region-aware loss did not outperform the simpler 2.5D U-Net on positive-only WT/TC/ET, and light post-processing improved only all-slice Dice. A failure-mode audit found that the residual catastrophic predictions are concentrated on a small minority of diffuse, infiltrative tumors with mass effect. Conclusions: Carefully engineered training strategies, tumor-aware sampling, EMA stabilization, and a modest 2.5D context window recover a substantial fraction of the accuracy of much heavier 3D networks at a fraction of the compute, are reproducible across seeds, and outperform a heavier GDNet-inspired architectural variant on the same data. GDNet is therefore a practical and, pending external validation, potentially clinically deployable framework for multimodal glioma segmentation on workstation-class GPU hardware. Full article

This paper proposes the Continuous Operational Feedback Loop (COFL) architecture, a fully localized, event-driven operational monitoring and response system for Direct-to-Consumer (D2C) commerce. The architecture integrates the n8n workflow engine with on-premise large language model (LLM) inference via the Ollama framework, forming a containerized stack deployable on commodity CPU-only edge hardware (~USD 1640). Using a multi-source dataset of 1800 records constructed from publicly available e-commerce corpora and evaluated with a silver-standard automated labeling protocol, empirical validation demonstrates an end-to-end latency of 3.22 s and a macro-F1 sentiment classification score of 0.836—representing 98.2% of the full-precision baseline and 94.0% of cloud GPT-4o API generation quality measured by ROUGE-L—at approximately 1/200th of the per-request inference cost. A systematic quantization ablation study across six model-quantization configurations establishes LLaMA 3 8B Q4_K_M as the Pareto-optimal selection for the target hardware. An Analytic Hierarchy Process (AHP) multi-criteria framework with criterion weights derived from published literature confirms the COFL implementation achieves a higher composite score than cloud API deployment under the stated evaluation assumptions. Failure mode and effects analysis (FMEA) is summarized to characterize system reliability under identified failure scenarios. Full article

Stereo vision is essential for passive three-dimensional perception in resource-constrained applications that require low power consumption, predictable latency, and explainable geometry. Although deep learning architectures dominate recent benchmarks, the classical block-matching pipeline remains a foundational approach. Optimizing this pipeline involves navigating complex trade-offs among matching robustness, map density, and computational efficiency. This study systematically surveys and physically validates the classical stereo framework. After revisiting geometric first principles, three matching costs (SAD, NCC, ZNCC) are benchmarked alongside Sobel preprocessing and structural refinements, with subsequent validation using a calibrated consumer webcam rig. Middlebury benchmarks (2001–2021) indicate that while SAD fails under complex radiometric distortion, NCC consistently achieves superior quantitative metrics, incurring only a 1.2-fold computational overhead. Extending the disparity search range improves foreground localization, while block size imposes a trade-off between resolving the aperture problem and preserving fine geometric detail. To bridge theoretical analysis and practical deployment, the pipeline is validated using a custom-calibrated consumer stereo rig. The optimized Sobel-NCC architecture is then evaluated for real-time edge deployment on constrained hardware (NVIDIA Jetson Nano) and narrow-baseline sensors (OAK-D SR) in the context of agricultural robotic manipulation. By prioritizing metric precision over dense prediction, the classical pipeline reconstructs target surfaces with approximately 1 cm depth accuracy at 21 frames per second. These results demonstrate that optimized local algorithms offer deterministic and reliable geometric foundations for real-time edge-computed robotics. Although neural networks are essential for dense reconstructions in ill-posed regions, the foundational principles established here remain indispensable for advanced stereo vision system deployment. Full article

Despite recent advances, medical question answering systems still struggle with domain-specific reasoning and data efficiency. This paper presents Med-LLaMA3, a family of medical large language models developed by parameter-efficient fine-tuning of the LLaMA-3.1 (8 billion) and LLaMA-3.2 (1 and 3 billion) architectures using quantized low-rank adaptation (QLoRA) and low-rank adaptation (LoRA) with 4-bit quantization. Beyond model training, this work contributes the following: (1) a formalized dataset curation taxonomy (source type × clinical granularity × task format) with a source-category ablation confirming that the multi-source combination drives benchmark gains beyond any single category; (2) a systematic characterization of low-rank-adaptation rank-scaling behavior for the LLaMA-3 family in the medical domain (monotonic improvement up to rank 128, with no observed plateau); and (3) statistically validated comparisons using McNemar’s test and 95% bootstrap confidence intervals. We curated a medical instruction dataset of over 1.5 million samples spanning medical examinations, clinical dialogues, and biomedical literature. Our approach trains only ∼4% of the base model’s parameters and, consistent with prior studies of parameter-efficient methods in the medical domain, achieves performance comparable to full fine-tuning at a fraction of the memory footprint. Evaluated with five in-context examples per prompt, the 8-billion-parameter model attains a mean accuracy of 75.71% across the eight medical-domain subsets of the Massive Multitask Language Understanding benchmark; improvements over the unmodified LLaMA-3.1-8B-Instruct baseline are statistically significant on the medical multiple-choice benchmark MedMCQA and, after Bonferroni correction across the eight subsets, on three subsets (Clinical Knowledge, Medical Genetics, and Nutrition), with two further subsets being significant only before correction. A structured named-entity-recognition evaluation on 100 hospital discharge summaries (macro-averaged F1 0.94; dual-annotator agreement $\kappa =0.87$) provides complementary evidence of clinical-text utility. A safety mitigation pilot shows that context-disambiguation preprocessing reduces the highest-severity abbreviation-ambiguity error rate from 30% to 10% on a 30-case held-out set. These results show that parameter-efficient fine-tuning can deliver high-performance medical large language models while training only ∼4% of the model’s parameters and reducing memory use by roughly 75%, enabling development on low-cost consumer-grade hardware. Full article

High dynamic range (HDR) image reconstruction from a single low dynamic range (LDR) input remains an important problem for computational photography, particularly when practical deployment on consumer-grade hardware is considered. With the increasing availability of hardware supporting HDR, public demand for capturing and viewing HDR images has grown significantly. Recent research has explored deep learning-based approaches to reconstruct HDR images from low dynamic range (LDR) inputs by extracting regional pixel features or leveraging the camera response function (CRF) for model training. Many of these approaches employ Convolutional Neural Network (CNN) architectures and utilize skip connections to preserve learned information. Nevertheless, the configuration-level effects of data augmentation in HDR reconstruction remain insufficiently discussed. Existing CNN-based approaches, such as HDRCNN, HDRUNet, and ExpandNet, have demonstrated promising reconstruction ability, but they may involve a heavy backbone architecture, a long training time, or a limited discussion of how preprocessing configurations affect reconstruction performance. This study presents an engineering-oriented HDR reconstruction framework derived from HDRCNN, focusing on practical efficiency, structural fidelity, and training feasibility. The proposed framework introduces three modifications: (1) a configuration-level comparison of composite data augmentation settings, including unsharp masking, denoising, Gaussian blur, and brightness–contrast adjustment; (2) the replacement of the original VGG16 backbone with a ResNet50-based encoder enhanced with attention blocks and squeeze-and-excitation (SE) blocks for improved multi-scale feature extraction and channel-wise recalibration; and (3) the integration of mixed-precision training with cosine annealing learning-rate scheduling to reduce computational cost. Experimental results on the SI-HDR dataset show that the best composite augmentation configuration improves PSNR from 19.05 dB to 22.10 dB and SSIM from 0.6444 to 0.7714 without increasing the training time. Compared with the original VGG16-based HDRCNN setting, the ResNet50-based model reduces training time while improving SSIM from 0.2705 to 0.8512. Under the adopted comparison protocol, the proposed model achieves the shortest training time and slightly higher PSNR than HDRUNet, while HDRUNet retains a higher SSIM. This indicates a trade-off among pixel-wise fidelity, structural similarity, and computational efficiency. The current evaluation is limited by a small test setting, composite rather than operation-level augmentation analysis, and the use of PSNR and SSIM only; therefore, future work should include full benchmark evaluation, additional perceptual/HDR-specific metrics, and controlled component-level ablation studies. Full article

Image sensors produce high-dimensional visual data for classification algorithms. Deep Neural Networks (DNNs) achieve high accuracy but require large labeled datasets and computational and energy resources, limiting their use in embedded systems. Active Learning (ALrn) can reduce labeling effort by selecting samples based on informativeness scores, but it remains computationally expensive, especially for high-dimensional images. This work presents a hardware-accelerated approach for the instance selection stage based on a query strategy in uncertainty-based ALrn for image classification using a novel in-line top-k selection algorithm that avoids conventional sorting and reduces memory and computational requirements. The algorithm is implemented on an Xilinx ZYNQ-7000 System on Chip (SoC) using a Field Programmable Gate Array (FPGA)-based accelerator operating at 110 MHz, interfacing with an embedded Advanced RISC Machine (ARM) processor for data acquisition and communication via the Python Productivity for Zynq (PYNQ) framework. Experiments on diverse multiclass datasets demonstrate correctness within an ALrn setting, showing negligible performance deviation in the learning curves compared to software baselines. The accelerator achieves speedup of 231.7× and 22.9× over software baseline and optimized software implementation of the proposed algorithm, respectively, in query-strategy computation while consuming only 0.473 W, substantially lower than conventional Central Processing Unit (CPU)- and Graphics Processing Unit (GPU)-based platforms. These results demonstrate the efficiency and extensibility of the proposed accelerator across alternative ALrn designs and hardware platforms, where the computational cost of instance selection scales with the size of the unlabeled pool. Full article

Head-mounted eye-tracking systems play a critical role in virtual reality, human–computer interaction, and clinical applications, yet achieving both high angular accuracy and precise 3D gaze position estimation with low-cost hardware remains challenging. This paper proposes a lightweight, training-free geometric 3D gaze tracking framework for binocular 3D gaze tracking using consumer-grade hardware, which leverages stereo geometric triangulation and a simplified physiological eye model to achieve robust 3D gaze estimation, requiring only standard infrared cameras and dichroic mirrors without additional specialized hardware. The method was evaluated in controlled indoor conditions with 30 participants, where it achieved an angular error ranging from 1.1° to 2.82° and a 3D gaze position error below 13.24 mm. Compared to two state-of-the-art academic non-deep-learning methods, the proposed framework delivers competitive angular accuracy while significantly reducing 3D position error, outperforming the baselines by 34% to 56% in depth estimation precision. These results demonstrates that the proposed geometric framework is a practical and effective solution for high-precision 3D gaze tracking on low-cost hardware, suitable for both research and consumer applications. Full article

Access to operational information in industrial plants forces operators to interrupt their tasks, walk to the human–machine interface (HMI) terminals, and navigate heterogeneous platforms—namely programmable logic controllers (PLC), supervisory control and data acquisition (SCADA) systems, manufacturing execution systems (MES), and enterprise resource planning (ERP) systems—consuming 15–30 min per query. Previous work integrated local large language models (LLMs) into a five-layer cognitive architecture deployed in a precast concrete plant, reducing that time to 14–23 s through voice-based conversational queries; however, model inference accounted for 55.3% of total latency and the system remained reactive. This work incorporates the event-driven paradigm as a non-intrusive augmentation layer that keeps the operational context permanently updated, continuously monitoring the process and refreshing knowledge only when significant changes occur. The architecture is fully local, cloud-independent, graphics processing unit (GPU)-free, and containerized via Docker Compose. Experimental results demonstrate a 26–31% reduction in response times (means of 9.84 s, 11.23 s, and 16.47 s for simple, moderate, and complex queries), an 8.4 °C reduction in peak hardware temperature (from 79.6 °C to 71.2 °C), a 41.6% decrease in thermal variability, and an expansion of the safety margin before central processing unit (CPU) throttling from 5.4 °C to 13.8 °C. The system achieved 100% success rate and availability over 30 min of autonomous operation, validated in a real industrial environment. Full article

The emerging Kolmogorov–Arnold networks (KANs) have set a new standard in machine learning (ML) tasks by prevailing over traditionally deployed multilayer perceptrons (MLPs) thanks to their enhanced interpretability through activation function learning, while they require increased computational complexity and memory footprint. Radial-basis function (RBF)-based KAN models maintain high performance over other variants of KANs with considerable size reduction and consequently more efficient execution. Aiming at effectively supporting the inference of RBF-KANs on Internet-of-Things (IoT) devices, this paper focuses on edge-oriented computing and introduces a soft intellectual property (IP) core, written in hardware description language (HDL), targeting the execution of such networks on all-programmable systems-on-chip (APSoC). The proposed design is fully pipelined and runtime configurable, allowing for real-time inference and latency-sensitive neural network deployment on-the-fly. A testbench reveals up to $43.6\times$ speedup when compared with a commercial edge central processing unit (CPU) and consumes considerably less power. The core’s adaptable design enables efficient allocation of resources and meets diverse throughput demands, making it well-suited for a broad range of IoT applications. Full article

With the development of human–computer interaction technology, non-contact interaction based on gaze tracking and facial movements has become a research hotspot. Traditional mouse-and-keyboard methods pose challenges for people with disabilities or limited hand movements, while existing gaze-tracking systems often rely on expensive hardware or lack sufficient accuracy. This paper designs and implements a real-time system using ordinary cameras, achieving natural, efficient interaction via multimodal input combination. The system uses an improved MobileNetV2 backbone to construct GazeTrackNet for gaze estimation. It adopts MediaPipe Face Mesh to detect facial landmarks. Meanwhile, it applies geometric feature analysis, including eye aspect ratio and mouth aspect ratio, to identify actions such as blinking and mouth opening. It adopts a hybrid control strategy that combines gaze jumping and head fine-tuning, using mouth state as the main control switch. Key contributions include a lightweight gaze-tracking algorithm that enables stable and efficient gaze detection on consumer-grade hardware, a multimodal interaction strategy based on facial movement that improves system stability and ease of use, and a complete prototype system that achieves real-time performance on standard laptops. Experimental results show an average gaze average angle error of 3.0°, 97% eye state recognition accuracy, and end-to-end latency below 70 ms. The system can satisfy the requirements of daily desktop interaction under normal indoor lighting, and shows potential for future barrier-free interaction applications after further validation with target users. Existing gaze-tracking methods either suffer from low precision on lightweight devices or bring heavy computational overhead. Common facial recognition approaches also face frequent false trigger interference. Compared with them, our scheme achieves balanced accuracy and real-time performance via an attention-enhanced structure, and the designed dual anti-shake mechanism effectively suppresses misjudgment, delivering a more stable hands-free interaction experience. Full article

This work presents the development and validation of an automated 3D scanning device based on two opposed consumer-grade Apple TrueDepth sensors integrated into a controlled rotational architecture, designed for the digitization of complex freeform surfaces such as the external cranial geometry. The system design was guided by a prior metrological characterisation of the sensor’s distance-dependent behaviour and complemented by an additional study of the influence of surface orientation, from which a suitable operating window for complete head acquisition was derived. On this basis, a mechatronic system was implemented comprising a mechanical structure, electronic hardware, a control architecture, and a calibration procedure that registers the local point clouds from both sensors into a common global coordinate system. Geometric validation was performed using symmetric and asymmetric cranial phantoms digitized with both the proposed device and a professional reference scanner. Surface comparison revealed localized discrepancies concentrated in fine anatomical details, while the cranial vault showed good overall agreement, with RMS deviations of 0.314 mm and 0.286 mm for the symmetric and asymmetric phantoms, respectively. Morphometric consistency was assessed through the cranial vault asymmetry index (CVAI), for which both systems produced the same general trend with a maximum difference of 0.2%. These results demonstrate the feasibility of the proposed system as a geometrically consistent and morphometrically reliable instrument for head surface digitization under controlled laboratory conditions. Full article

Many machine learning and deep learning algorithms based on Artificial Neural Networks (ANNs) have been implemented on software platforms for handwritten digit and character recognition. However, an ANN is difficult to deploy on an embedded platform based on a Central Processing Unit (CPU) because of its large computation, complex structure, and frequent memory access. However, Field Programmable Gate Array (FPGA) devices facilitate this task and offer the capability to design fully customizable hardware architectures. Additionally, they provide high flexibility and high parallel computations based on parallel processing techniques, and they contain sufficient on-chip Digital Signal Processing (DSP) blocks useful for complicated multiplications. In this paper, we present a detailed FPGA-based implementation of a handwritten digit recognition system based on a Multi-Layer Perceptron (MLP) model. The internal modules of the network are designed using the VHSIC Hardware Description Language (VHDL) to achieve a high-level optimization on the hardware platform, and the functionality is simulated and tested using Vivado ISIM Tools. The system has been characterized to reach acceptable performance compared to previous approaches. After implementing the whole neural network on a Xilinx Pynq-Z2 board, it occupies in the device 20758 LUTs, 4426 FFs, 3.50 blocks of random-access memory (BRAM), and 42 DSPs. It reaches an execution time of 2.192 µs to recognize a handwritten number, while consuming only 0.36 Watts, and it achieves a classification accuracy of 97%. Additionally, the proposed architecture can be easily scaled on different FPGA devices thanks to its regularity. Therefore, it offers more portability of the architecture and can be used on different real embedded applications. Full article

Tomato leaf diseases substantially reduce tomato yields and quality and remain a persistent challenge for efficient crop management. Although deep learning-based detectors have achieved strong accuracy in controlled benchmarks, many existing solutions are still difficult to transfer to resource-constrained agricultural systems because they rely on high-end GPUs, consume considerable power, and often lose performance after deployment on embedded devices. To address this practical gap, this study proposes HGS-YOLO, a system-oriented deployable lightweight adaptation of YOLOv11 for leaf-level tomato disease detection, together with an end-to-end edge sensing pipeline for low-power agricultural deployment. The main contribution lies in the coordinated system-level co-design of model structure, optimization, and deployment rather than in a novel detector architecture. Specifically, YOLOv11 is adapted through three coordinated modifications: an HGNetV2 backbone for efficient feature extraction, an HS-FPN neck with channel attention for lightweight multi-scale fusion, and an MPDIoU loss function for more stable localization optimization. Beyond the model architecture, the study establishes a complete engineering pipeline that includes training, optimization, post-training quantization, and hardware deployment with BPU acceleration on a D-Robotics RDK X5 handheld platform. Comprehensive benchmark experiments indicate that HGS-YOLO achieves 93.6% ${\mathrm{mAP}}_{50}$ and 72.1% mAP@[0.5:0.95] with 86.5% recall, only 1.3 M parameters, and a 3.1 MB model size, substantially reducing the model complexity and storage cost relative to the YOLOv11 baseline. A three-seed retraining comparison shows that HGS-YOLO trades roughly 0.5 ${\mathrm{mAP}}_{50}$ points for this compactness (a statistically significant but small concession) and recovers the cost on the deployment side: on the RDK X5 chip, HGS-YOLO is the fastest, most memory-efficient, and lowest-power model among all compared detectors. Indoor deployment tests using separately collected tomato leaf samples further achieve 90.3% ${\mathrm{mAP}}_{50}$, 82.3% recall, 89.0% precision, 25.0 ± 0.4 ms end-to-end latency, 40.0 ± 0.6 FPS, and 9.8 ± 0.4 W average system power. After PTQ, the ${\mathrm{mAP}}_{50}$ drops from 93.6% to 93.0% on the same benchmark; because this figure was measured under controlled imaging conditions, it is presented as an in-distribution reference point rather than as evidence of robustness in the open field. We also took the handheld system into a working tomato greenhouse for a small outdoor field round, where it ran end-to-end and produced on-device disease detections under natural sunlight, specular highlights, partial occlusion, background clutter, and handheld motion blur. These results show that HGS-YOLO reaches a good balance of accuracy, efficiency, and deployability and that it works in the field on an independent small-scale test; validating it more widely across sites, seasons, and weather is left to future work. Full article

Polymer-augmented waterflooding is a key enhanced oil recovery technique whose simulation remains computationally demanding at a high spatial resolution. This paper presents a fully GPU-resident parallel solver for the one-dimensional Buckley–Leverett polymer-flooding problem within an Implicit-Pressure–Explicit-Saturation framework. The solver combines Jacobi iteration for pressure, first-order upwind flux splitting for saturation, and a first-order upwind flux-splitting update for polymer mass with explicit concentration recovery inside a coupled Picard–IMPES iteration. Two CUDA implementations are compared: a global-memory baseline and a shared-memory variant that stages a per-block pressure tile with halo cells on chip. Both kernels were profiled on an NVIDIA GeForce RTX 2080 Ti over problem sizes from $N=65,536$ to $N=67,108,864$ and block sizes 128, 256, 512, and 1024. The two GPU implementations match the serial reference within $2\times {10}^{-8}$, and peak speed-ups are $20.2\times$ (global) and $20.1\times$ (shared). Per-kernel Nsight Compute profiling classifies every kernel in both builds as compute-bound: SM throughput is 54–83% of peak and DRAM throughput 3–29% of peak. The bottleneck is the FP64 pipeline of consumer Turing hardware (FP64 throughput is one thirty-second of FP32); three FP64 divisions per cell, from inline polymer-modified mobility recomputation, saturate the FP64 unit. Shared-memory tiling cannot improve performance because it acts on memory traffic rather than on compute throughput. The result therefore characterizes a specific regime, namely FP64 one-dimensional, low-reuse transport stencils on consumer-class NVIDIA GPUs with reduced FP64 throughput, and is not a universal property of CUDA shared memory. Full article