Impact on the Accuracy of Aggressive Voltage Underscaling in CNN Accelerators

Abstract

Chip designers usually rely on conservative supply voltage (Vdd) guardbands to prevent permanent
faults as a consequence of CMOS process variations. On the other hand, aggressively undervolting
below the safe voltage margin leads to huge energy savings since energy scales quadratically with
Vdd. Convolutional Neural Networks (CNNs) can be, to some extent, resilient to faults since they
usually include significant amounts of data redundancy. This paper shows that the accuracy of large
CNNs, like Alexnet and Squeezenet, is severely compromised when the Vdd of a CNN accelerator is
reduced down to 0.54 V and 0.58 V, respectively.