In this paper, we evaluate the impact on reliability and performance of the selective approximation of Convolutional Neural Networks (CNNs) layers on NVIDIA mixed-precision architectures. We found that, even without affecting accuracy, the approximation from single to half precision of each layer has a different impact on both performance and output error.

Impact of Layers Selective Approximation on CNNs Reliability and Performance / Rech Junior, Rubens Luiz; Rech, Paolo. - (2020), pp. 1-4. (Intervento presentato al convegno DFT 2020 tenutosi a Frascati, Italy (On-line Virtual Event) nel 19th–21st October 2020) [10.1109/DFT50435.2020.9250821].

Impact of Layers Selective Approximation on CNNs Reliability and Performance

Rech, Paolo
Ultimo
2020-01-01

Abstract

In this paper, we evaluate the impact on reliability and performance of the selective approximation of Convolutional Neural Networks (CNNs) layers on NVIDIA mixed-precision architectures. We found that, even without affecting accuracy, the approximation from single to half precision of each layer has a different impact on both performance and output error.
2020
33rd IEEE International Symposium on Defect and Fault Tolerance in VLSI and Nanotechnology Systems
Piscataway, NJ
Institute of Electrical and Electronics Engineers Inc.
978-1-7281-9457-8
Rech Junior, Rubens Luiz; Rech, Paolo
Impact of Layers Selective Approximation on CNNs Reliability and Performance / Rech Junior, Rubens Luiz; Rech, Paolo. - (2020), pp. 1-4. (Intervento presentato al convegno DFT 2020 tenutosi a Frascati, Italy (On-line Virtual Event) nel 19th–21st October 2020) [10.1109/DFT50435.2020.9250821].
File in questo prodotto:
File Dimensione Formato  
DFT_Impact_of_Layers_Selective_Approximation_on_CNNs_Reliability_and_Performance.pdf

Solo gestori archivio

Tipologia: Versione editoriale (Publisher’s layout)
Licenza: Tutti i diritti riservati (All rights reserved)
Dimensione 202.74 kB
Formato Adobe PDF
202.74 kB Adobe PDF   Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11572/346645
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 1
  • ???jsp.display-item.citation.isi??? 1
  • OpenAlex ND
social impact