Anatomizing Deep Learning Inference in Web Browsers DOI Open Access
Qipeng Wang, Shiqi Jiang, Zhenpeng Chen

et al.

ACM Transactions on Software Engineering and Methodology, Journal Year: 2024, Volume and Issue: unknown

Published: Aug. 14, 2024

Web applications have increasingly adopted Deep Learning (DL) through in-browser inference , wherein DL performs directly within browsers. The actual performance of and its impacts on the quality experience ( QoE ) remain unexplored, urgently require new measurements beyond traditional ones, e.g., mainly focusing page load time. To bridge this gap, we make first comprehensive measurement to date. Our approach proposes metrics measure inference: responsiveness, smoothness, accuracy. extensive analysis involves 9 representative models across browsers 50 popular PC devices 20 mobile devices. results reveal that exhibits a substantial latency averaging 16.9 times slower CPU 4.9 GPU compared native gap is 15.8 7.8 times, respectively. Furthermore, identify contributing factors such including underutilized hardware instruction sets, inherent overhead in runtime environment, resource contention browser, inefficiencies software libraries abstractions. Additionally, imposes significant memory demands, at exceeding 334.6 size themselves, partly attributable suboptimal management. We also observe leads 67.2% increase time it takes for GUI components render browsers, significantly affecting overall user reliant technology.

Language: Английский

GL2GPU: Accelerating WebGL Applications via Dynamic API Translation to WebGPU DOI
Yudong Han, Weichen Bi,

R. An

et al.

Published: April 22, 2025

Language: Английский

Citations

0
WeInfer: Unleashing the Power of WebGPU on LLM Inference in Web Browsers DOI

Z.M. Chen,

Yun Ma,

Haiyang Shen

et al.

Published: April 22, 2025

Language: Английский

Citations

0
Anatomizing Deep Learning Inference in Web Browsers DOI Open Access
Qipeng Wang, Shiqi Jiang, Zhenpeng Chen

et al.

ACM Transactions on Software Engineering and Methodology, Journal Year: 2024, Volume and Issue: unknown

Published: Aug. 14, 2024

Web applications have increasingly adopted Deep Learning (DL) through in-browser inference , wherein DL performs directly within browsers. The actual performance of and its impacts on the quality experience ( QoE ) remain unexplored, urgently require new measurements beyond traditional ones, e.g., mainly focusing page load time. To bridge this gap, we make first comprehensive measurement to date. Our approach proposes metrics measure inference: responsiveness, smoothness, accuracy. extensive analysis involves 9 representative models across browsers 50 popular PC devices 20 mobile devices. results reveal that exhibits a substantial latency averaging 16.9 times slower CPU 4.9 GPU compared native gap is 15.8 7.8 times, respectively. Furthermore, identify contributing factors such including underutilized hardware instruction sets, inherent overhead in runtime environment, resource contention browser, inefficiencies software libraries abstractions. Additionally, imposes significant memory demands, at exceeding 334.6 size themselves, partly attributable suboptimal management. We also observe leads 67.2% increase time it takes for GUI components render browsers, significantly affecting overall user reliant technology.

Language: Английский

Citations

0