A Hybrid JavaScript–WebAssembly Framework for Efficient Deep Learning Inference in Web Browsers

• Chaeeun Lee,
• Sanghoon Jeon,

Vol. 20, No. 1, January 31, 2026
10.3837/tiis.2026.01.011, Download Paper (Free):

• Web-based Deep Learning
• Javascript
• WebAssembly
• Hybrid Execution Framework
• Real-Time Inference

Abstract

As web applications increasingly demand real-time AI inference, optimizing deep-learning execution in browser environments has become critical. This paper evaluates the performance of deep learning models executed using JavaScript, WebAssembly, and the proposed Hybrid Framework that combines both approaches. Using the ResNet-18, ResNet-50, and ResNeXt-50 architectures across the CIFAR-10 and STL-10 datasets, we assessed four key metrics: inference time, memory usage, CPU utilization, and GPU utilization. Experimental results show that WebAssembly generally achieves faster inference and higher CPU/GPU utilization than JavaScript but at the cost of significantly increased memory usage. In contrast, JavaScript demonstrates better memory efficiency and lower system overhead. The Hybrid Framework attempts to balance these trade-offs by dynamically selecting the execution mode based on input resolution and performance metrics. However, in high-resolution tasks such as STL-10, the Hybrid Framework exhibited the worst inference time and highest memory consumption among all the methods, indicating that its optimization logic favored stable CPU and GPU utilization at the expense of latency and memory efficiency. These findings suggest that although hybrid execution can offer performance benefits under moderate conditions, its effectiveness under complex workloads requires careful calibration. This paper highlights the need for metric-aware and resource-adaptive execution strategies to realize efficient and scalable web-based deep learning services.

Statistics

Cite this article