Hardened-TC: A Low-cost Reliability Solution for CNNs Run by Modern GPUs

Convolutional neural networks (CNNs) have become a compelling solution for various applications such as image classification, object detection, and climate studies. The introduction of tensor cores (TCs) in NVIDIA GPUs targets the acceleration of neural network computations. While there have been numerous studies on the performance and power of TCs, the reliability of TCs received little attention, particularly at the hardware level. Recently, CNNs have been deployed into safety-critical applications such as self-driving cars. Soft errors caused by high-energy particles are concerning as they can lead to catastrophic failures in CNN systems. The high power and area cost of traditional methods for building resilient systems such as triple modular redundancy make selective protection techniques attractive. We propose a hardware-based selective protection mechanism where vulnerable components of a CNN are implemented on resilient TCs. Resizing transistors in hardened TCs offers a low-cost solution to boost the reliability of CNNs. We also propose a precision-aware approach to optimize hardened TCs further. CNNs are resilient to approximation and quite often do not require full precision. By dropping transistor resizing in the least significant bits of network parameters, we are able to offer a low-cost reliability solution for CNNs implemented on TCs while maintaining the accuracy of the original network.