TechRxiv
TCASII_ClusterOfFPAAs.pdf (17.69 MB)

A Cluster of FPAAs to Recognize Images using Neural Networks

Download (17.69 MB)
preprint
posted on 24.05.2021, 18:00 by Daniel García MorenoDaniel García Moreno, Alberto A Del Barrio, Guillermo Botella, Jennifer Hasler
Analog computing has been recovering its relevance in the recent years. FPAAs are the equivalent to FPGAs but in the analog domain. The main drawback of FPAAs is their reduced integration capacity. In order to increase the amount of analog resources, in this paper a cluster of 40 FPAAs is proposed. As a use case, a 19-8-6-4 feedforward Neural Network has been implemented on such cluster. With the help of a DCT-based software framework, this NN is able to classify 28x28 MNIST images. Results show that the analog network is able to obtain almost the same results as the software baseline network.

Funding

ConvergenciaA Big dAta-Hpc: de Los sensores a las Aplicaciones (S2018/TCS-4423)

Detección temprana de desarrollo tumoral mediante computación de altas prestaciones (PR26/16-20B-1)

cHIMERA: Heterogeneity and specialization in the Post-Moore ERA (RTI2018-093684-BI00)

History

Email Address of Submitting Author

daniel10@ucm.es

ORCID of Submitting Author

0000-0003-1707-106X

Submitting Author's Institution

Complutense University of Madrid

Submitting Author's Country

Spain