On the impact of OxRAM-based synapses variability on convolutional neural networks performance

D. Garbin (1, 2) , E. Vianello (1) , O. Bichler (3) , M. Azzaz (4) , Q. Rafhay (2) , P. Candelier (4) , C. Gamrat (3) , G. Ghibaudo (2) , B. Desalvo (1) , L. Perniola (1)

1 CEA-LETI - Commissariat à l'énergie atomique et aux énergies alternatives - Laboratoire d'Electronique et de Technologie de l'Information
2 IMEP-LAHC - Institut de Microélectronique, Electromagnétisme et Photonique - Laboratoire d'Hyperfréquences et Caractérisation
3 DACLE-LIST - Département d'Architectures, Conception et Logiciels Embarqués-LIST
4 ST-CROLLES - STMicroelectronics [Crolles]

In this work, both temporal (cycle-to-cycle) and spatial (device-to-device) variability of hafnium oxide based OxRAM cells are investigated at array level. The impact of the resistance variability on OxRAM-based convolutional neural network is then evaluated. Two different types of neurons, analog and digital, are considered. Results show that the studied architecture is strongly immune to both temporal and spatial variability.

Mots clés: variability; Convolution; Convolutional neural network; Resistive memory; Network architecture; OxRAM; Nanotechnology; Neural networks; synapse; Hafnium oxides

Domaines: Sciences de l'ingénieur [physics]

D. Garbin, E. Vianello, O. Bichler, M. Azzaz, Q. Rafhay, et al.. On the impact of OxRAM-based synapses variability on convolutional neural networks performance. 2015 IEEE/ACM International Symposium on Nanoscale Architectures (NANOARCH´15), Jul 2015, Boston, MA, United States. pp.193-198, ⟨10.1109/NANOARCH.2015.7180611⟩. ⟨cea-01839851⟩