Reliability Assessment of QNN Functional Units deployed in RISC-V based systems

The reliability of Quantized Neural Networks (QNNs) on RISC-V hardware is a critical concern for safety-critical edge AI. This paper presents a systematic gate-level fault criticality assessment of permanent stuck-at faults within the accumulator and multiplier of a MAC unit, leveraging the ALU and MUL of an Ibex RISC-V core. Gathered on four QNN models (MobileNetV1/V2, ResNet18/34) and three datasets (CIFAR-10, GTSRB, ImageNet2012), our results quantify a significant disparity in hardware criticality: at parity of model and dataset, an accumulator fault can lead to a Top-1 accuracy degradation up to 22.18\% greater than a multiplier fault. While multiplication stages introduce substantial noise, accumulation defects trigger catastrophic collapses, with Top-1 accuracy dropping to 0.07\% on ImageNet2012. Consequently, hardening both critical MAC components is essential for resilient RISC-V accelerators, as the accumulation stage is especially prone to degrading final accuracy.