Reliability-Aware Performance Optimization of DNN HW Accelerators Through Heterogeneous Quantization

The paper introduces a framework for implementing heterogeneous quantization tailored to individual layers of DNN accelerators. The proposed method, evaluated across DNN benchmarks AlexNet, VGG-11 and ResNet-18, enhances both accuracy drop and performance. On average, the accuracy drop at a BER of 10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−4</sup> was reduced by nearly 4 times compared to homogeneous quantization, demonstrating a considerable improvement in reliability under fault-prone conditions. Furthermore, the memory overhead introduced by the method for protected heterogeneous networks is less than 0.1%, highlighting that these reliability gains come with negligible resource costs.