Abstract Efficient data ingestion and online data augmentation remain challenges in deep learning workflows, particularly when dealing with datasets containing non-standard formats or massive multidimensional arrays that natively optimised functions cannot fully manage. This work presents a parallel framework that integrates and global shared memory through a ring buffer architecture, enabling high-throughput data loading and flexible on-the-fly augmentation. The framework decouples data production from consumption, allowing multiple CPU workers to load and preprocess batches in parallel while completely bypassing the Python GIL and memory bottlenecks. Crucially, the framework supports both CPU-side and GPU-side augmentation strategies, adapting to whether complex conditional transformations or framework-native operations are required. The proposed approach was validated on two representative tasks: (i) sign language recognition from human pose CSV sequences, and (ii) hyperspectral image classification using massive arrays. Relative to standard sequential baselines, the proposed framework achieved up to $$27\times $$ 27 × acceleration in isolated data ingestion and up to $$28\times $$ 28 × in end-to-end training. Importantly, even against natively optimised parallel TensorFlow and PyTorch pipelines, it still delivered up to $$8\times $$ 8 × faster data loading and up to $$7\times $$ 7 × faster full training in memory-intensive scenarios. Overall, the proposed framework provides a scalable, multi-GPU compatible solution for deep learning pipelines, showing robust performance across both I/O-bound and memory-constrained scenarios in TensorFlow and PyTorch while alleviating memory fragmentation and allocation constraints.

Abstract This paper presents a novel composite architecture for high-precision sign language gesture recognition referred to as ViTSLA. The model integrates convolutional inductive bias with transformer-based global attention to improve predictive accuracy and computational efficiency. Extensive experiments are conducted on three datasets using five-fold cross-validation, ablation analysis, and comparative evaluation against standard vision transformer (ViT) variants. Experimental results demonstrate consistently high classification accuracy exceeding 99.5% with minimal variance across folds, confirming robustness and balanced class-wise performance. Beyond predictive performance, this work addresses the computational demands of training and deploying hybrid deep architectures. Transformer-based attention mechanisms and high-resolution tokenization increase memory consumption and computational complexity. Therefore, efficient training requires parallel processing, GPU acceleration, and high-performance computing (HPC) environments to handle large-scale tensor operations and multi-fold cross-validation experiments. Furthermore, real-time inference capability is maintained at over 300 FPS, which is essential for deployment in latency-sensitive applications such as interactive systems and intelligent human–computer interfaces. Distributed computation further enables scalable training and stable convergence across datasets. Moreover, the low mean latency, stable P95 behavior, and high throughput confirm real-time deployment capability. These results demonstrate that the proposed architecture achieves high predictive performance while maintaining computational characteristics aligned with supercomputing-driven intelligent vision systems. The developed model achieves a favorable balance between accuracy, parameter efficiency, and computational cost. The results demonstrate that hybrid convolutional-attention architectures benefit from HPC-enabled parallelism, leading to scalable optimization and deployment-ready performance. These findings establish the relevance of the proposed approach within the domain of supercomputing-driven intelligent vision systems.

Abstract Although optimization algorithms have made significant progress in a wide range of power system applications, many research areas and case studies remain unexplored and require additional investigation. Therefore, this paper tackles the coordination process of overcurrent relays (OCRs) in large-scale and complicated networks comprising hundreds of decision variables and constraints. In this context, savannah bengal tiger optimizer (SBTO) is employed to solve the OCR’s coordination of the 42-bus and 30-bus power grids by optimizing pickup current, time dial, and operating curve. It is worth noting that the 42-bus network comprises 291 decision variables and 713 constraints while the 30-bus incorporates 333 decision variables and 1119 constraints: indicating challenging problem formulation to be optimized. Thus, high-performance computing is necessary to provide the necessary computational efficiency and real-time results since the suggested method requires parallel processing capabilities for large-scale simulations and high-resolution models. The objective function (OF) is adapted to have a minimal total operating time (TOT) along with fulfilling problem constraints. Exceptional results of the utilized SBTO are validated against other benchmark algorithms such as genetic algorithm, particle swarm optimizer, water cycle algorithm, and gray wolf optimizer. Obviously, SBTO effectively meets the coordination criteria by achieving a minimal OF and incurring only a limited number of constraint violations. Furthermore, SBTO attains the minimum possible TOT of 19.66 s among other competitors for the 42-bus study case. This approach highlights SBTO’s capability in effectively addressing highly complex optimization problems with numerous constraints.