Publication Details

Exploiting Quantization and Mapping Synergy in Hardware-Aware Deep Neural Network Accelerators

KLHŮFEK Jan, ŠAFÁŘ Miroslav, MRÁZEK Vojtěch, VAŠÍČEK Zdeněk and SEKANINA Lukáš. Exploiting Quantization and Mapping Synergy in Hardware-Aware Deep Neural Network Accelerators. In: 2024 27th International Symposium on Design & Diagnostics of Electronic Circuits & Systems (DDECS). Kielce: Institute of Electrical and Electronics Engineers, 2024, pp. 1-6. ISBN 979-8-3503-5934-3. Available from: https://arxiv.org/abs/2404.05368
Czech title
Výzkum synergie kvantizace a mapování v oblasti hardwarových akcelerátorů hlubokých neuronových sítí
Type
conference paper
Language
english
Authors
Klhůfek Jan, Ing. (DCSY FIT BUT)
Šafář Miroslav, Bc. (FIT BUT)
Mrázek Vojtěch, Ing., Ph.D. (DCSY FIT BUT)
Vašíček Zdeněk, doc. Ing., Ph.D. (DCSY FIT BUT)
Sekanina Lukáš, prof. Ing., Ph.D. (DCSY FIT BUT)
URL
Keywords

Quantization, Neural networks, Hardware accelerator

Abstract

Energy efficiency and memory footprint of a convolutional neural network (CNN) implemented on a CNN inference accelerator depend on many factors, including a weight quantization strategy (i.e., data types and bit-widths) and mapping (i.e., placement and scheduling of DNN elementary operations on hardware units of the accelerator). We show that enabling rich mixed quantization schemes during the implementation can open a previously hidden space of mappings that utilize the hardware resources more effectively. CNNs utilizing quantized weights and activations and suitable mappings can significantly improve trade-offs among the accuracy, energy, and memory requirements compared to less carefully optimized CNN implementations. To find, analyze, and exploit these mappings, we: (i) extend a general-purpose state-of-the-art mapping tool (Timeloop) to support mixed quantization, which is not currently available; (ii) propose an efficient multi-objective optimization algorithm to find the most suitable bit-widths and mapping for each DNN layer executed on the accelerator; and (iii) conduct a detailed experimental evaluation to validate the proposed method. On two CNNs (MobileNetV1 and MobileNetV2) and two accelerators (Eyeriss and Simba) we show that for a given quality metric (such as the accuracy on ImageNet), energy savings are up to 37% without any accuracy drop. 

Published
2024
Pages
1-6
Proceedings
2024 27th International Symposium on Design & Diagnostics of Electronic Circuits & Systems (DDECS)
Conference
International Symposium on Design and Diagnostics of Electronic Circuits and Systems, Kielce, PL
ISBN
979-8-3503-5934-3
Publisher
Institute of Electrical and Electronics Engineers
Place
Kielce, PL
DOI
EID Scopus
BibTeX
@INPROCEEDINGS{FITPUB13071,
   author = "Jan Klh\r{u}fek and Miroslav \v{S}af\'{a}\v{r} and Vojt\v{e}ch Mr\'{a}zek and Zden\v{e}k Va\v{s}\'{i}\v{c}ek and Luk\'{a}\v{s} Sekanina",
   title = "Exploiting Quantization and Mapping Synergy in Hardware-Aware Deep Neural Network Accelerators",
   pages = "1--6",
   booktitle = "2024 27th International Symposium on Design \& Diagnostics of Electronic Circuits \& Systems (DDECS)",
   year = 2024,
   location = "Kielce, PL",
   publisher = "Institute of Electrical and Electronics Engineers",
   ISBN = "979-8-3503-5934-3",
   doi = "10.1109/DDECS60919.2024.10508920",
   language = "english",
   url = "https://www.fit.vut.cz/research/publication/13071"
}
Back to top