Exploring Fault-Energy Trade-offs in Approximate DNN Hardware Accelerators

Abstract

Systolic array-based deep neural network (DNN) accelerators have recently gained prominence for their low computational cost. However, their high energy consumption poses a bottleneck to their deployment in energy-constrained devices. To address this problem, approximate computing can be employed at the cost of some tolerable accuracy loss. However, such small accuracy variations may increase the sensitivity of DNNs towards undesired subtle disturbances, such as permanent faults. The impact of permanent faults in accurate DNNs has been thoroughly investigated in the literature. Conversely, the impact of permanent faults in approximate DNN accelerators (AxDNNs) is yet under-explored. The impact of such faults may vary with the fault bit positions, activation functions and approximation errors in AxDNN layers. Such dynamacity poses a considerable challenge to exploring the trade-off between their energy efficiency and fault resilience in AxDNNs. Towards this, we present an extensive layer-wise and bit-wise fault resilience and energy analysis of different AxDNNs, using the state-of-the-art Evoapprox8b signed multipliers. In particular, we vary the stuck-at-0, stuck-at-1 fault-bit positions, and activation functions to study their impact using the most widely used MNIST and Fashion-MNIST datasets. Our quantitative analysis shows that the permanent faults exacerbate the accuracy loss in AxDNNs when compared to the accurate DNN accelerators. For instance, a permanent fault in AxDNNs can lead up to 66% accuracy loss, whereas the same faulty bit can lead to only 9% accuracy loss in an accurate DNN accelerator. Our results demonstrate that the fault resilience in AxDNNs is orthogonal to the energy efficiency.

An Optimized Deep-Learning-Based Low Power Approximate Multiplier Design

“…The FPGA implementation is experimented with using this method by achieving minimum power consumption than GPU counterparts. Siddique et al [9] designed an Evoapprox8b signed multipliers which is energy efficient with bit-wise fault resilience and extensive layer-wise. The result shows that the energy efficiency obtained a fault resilience was orthogonal.…”

A design methodology for fault-tolerant computing using astrocyte neural networks

“…Recent efforts to this end include software solutions such as model replication [9] and error prediction coding [7], and hardware solutions such as approximation [12] and redundant mapping [20]. For FPGA-based neuromorphic designs, fault tolerance can also be addressed using periodic scrubbing [11,19].…”

Exposing Reliability Degradation and Mitigation in Approximate DNNs Under Permanent Faults

“…As discussed in this paper, the permanent faults affect the compute units of AxDNN accelerators in every execution cycle and their presence as unmasked faults leads to serious failures in the whole system. Indeed, their impact is stronger in AxDNN accelerators due to self-error-inducing approximate computations compared to the accurate design alternatives, i.e., accurate deep neural network (AccDNN) accelerators [15]. However, the ratio of fault resilience in AxDNNs to that of AccDNNs depends on data precision, location and type of permanent faults, size of accelerators, degree of approximations, activation functions, neural network topology, and accelerators.…”

“…Condia et al studied the effects of faults in critical and user-hidden modules (such as the Warp Scheduler and the Pipeline Registers) for the convolution computations in AccDNNs over GPU [23]. Very recently, our previous work in [15] explored the fault resilience of AxDNNs with their energy trade-offs running on a 256x256 approximate systolic array-based DNN accelerator that is analogous to Google TPU. However, the analysis presented in [15] has several limitations: (i) the analysis is limited to simple approximate multi-layer perceptrons, (ii) only layer-wise permanent faults on TPU-based AxDNNs are analyzed, and non-layer-wise analysis is missing, (iii) only systolic-array based architecture is analyzed and the impact of permanent faults on GPU-based accelerators is ignored, and (iv) the analysis does not provide any insights on the mitigation of the permanent faults.…”

“…Very recently, our previous work in [15] explored the fault resilience of AxDNNs with their energy trade-offs running on a 256x256 approximate systolic array-based DNN accelerator that is analogous to Google TPU. However, the analysis presented in [15] has several limitations: (i) the analysis is limited to simple approximate multi-layer perceptrons, (ii) only layer-wise permanent faults on TPU-based AxDNNs are analyzed, and non-layer-wise analysis is missing, (iii) only systolic-array based architecture is analyzed and the impact of permanent faults on GPU-based accelerators is ignored, and (iv) the analysis does not provide any insights on the mitigation of the permanent faults. It is worth recalling that retraining AxDNNs is a complicated problem; hence, the state-of-the-art retraining-based fault mitigation methods [24]- [26] for AccDNNs cannot be directly applied to AxDNNs.…”