This paper presents FORTUNE, a hardware-agnostic fault tolerance technique for DNNs that leverages quantization to enhance reliability without significant performance overhead. Unlike conventional methods like Triple Modular Redundancy (TMR), which are computationally expensive, the proposed approach uses memory savings from quantization to protect the critical Most Significant Bit, improving fault tolerance in Deep Neural Networks (DNNs). Memory utilization has been reduced by 37.5% across all networks, with vulnerability in AlexNet reduced by 56% compared to the 8-bit version and 84% compared to the unprotected 3-bit version. These improvements come with only a minor increase in execution time of less than 3%. Using AlexNet as an example demonstrates how our approach effectively enhances memory utilization and resilience while causing only a minimal increase in execution time.

Reliability assessment is mandatory to guarantee the correct behavior of Deep Neural Network (DNN) hardware accelerators in safety-critical applications. While fault injection stands out as a well-established, practical and robust method for reliability assessment, it is still a very time-consuming process. This paper contributes with three recipes for optimizing the efficiency of the reliability assessment: a) hybrid analytical and hierarchical FI-based reliability assessment for systolic-array-based DNN accelerators; b) mixing techniques for the reliability assessment of in-chip AI accelerators in GPUs; c) reliability assessment of DNN hardware accelerators through physical fault injection. The experimental results demonstrate the efficiency of the proposed methods applied to their target DNN HW accelerator platforms.  

Deep Neural Network (DNN) hardware accelerators are essential in a spectrum of safety-critical edge-AI applications with stringent reliability, energy efficiency, and latency requirements. Multiplication is the most resource-hungry operation in the neural network's processing elements. This paper proposes a scalable adaptive fault-tolerant approximate multiplier (AdAM) tailored for ASIC-based DNN accelerators at the algorithm and circuit levels. AdAM employs an adaptive adder that relies on an unconventional use of input Leading One Detector (LOD) values for fault detection by optimizing unutilized adder resources. A gate-level optimized LOD design and a hybrid adder design are also proposed as a part of the adaptive multiplier to improve the hardware performance. The proposed architecture uses a lightweight fault mitigation technique that sets the detected faulty bits to zero. The hardware resource utilization and the DNN accelerator's reliability metrics are used to compare the proposed solution against the Triple Modular Redundancy (TMR) in multiplication, unprotected exact multiplication, and unprotected approximate multiplication. It is demonstrated that the proposed architecture enables a multiplication with a reliability level close to the multipliers protected by TMR while at the same time utilizing 2.74 × less area and with 39.06% less power-delay product compared to the exact multiplier. Moreover, it has similar area, delay, and power consumption parameters compared to the state-of-the-art approximate multipliers with similar accuracy while providing fault detection and mitigation capability.