Introduction
As semiconductor manufacturing pushes toward nano-scale CMOS levels, the probability of hardware defects increases significantly. In the context of Deep Neural Network (DNN) accelerators, these defects often manifest as permanent "stuck-at" faults within on-chip weight memory. A single bit frozen at "0" or "1" can catastrophically degrade a model's inference accuracy, leading to reduced chip yield and increased manufacturing costs.
Conventional mitigation strategies, such as Fault-Aware Retraining (FAR), involve retraining the model specifically for each faulty chip. While effective, FAR is computationally prohibitive for large-scale models like ResNet on ImageNet when applied to every individual device. Alternatively, hardware-based solutions like Error Correction Codes (ECC) impose significant area and power overheads. This paper presents Fault-Aware Quantization (FAQ), a lightweight software-based technique that treats hardware memory faults as a specialized quantization problem, restoring accuracy with minimal computational cost.
Methodology: Fault-Aware Quantization (FAQ)
The core innovation of FAQ lies in the realization that a faulty memory location is not entirely useless; it simply has a restricted set of representable values. By identifying the "nearest valid value" that a faulty slot can store, FAQ minimizes the error introduced by the defect without requiring full retraining.
The FAQ pipeline consists of five primary stages:
- Lookup Table (LUT) Generation: The algorithm pre-calculates optimal reproducible values for all possible fault patterns based on the target bitwidth (e.g., 8-bit).
- Error Mask Generation: A hardware fault map is utilized to identify specific stuck-at bits for every weight address in the model.
- Scale Factor Computation: Standard quantization scaling is applied to weights and activations to maintain dynamic range.
- Standard Quantization: Weights are mapped to the target format (e.g., 8-bit signed integers).
- Mapping to Nearest Valid Values: The quantized weights are adjusted using the LUT. If a target weight cannot be represented due to a stuck bit, FAQ replaces it with the closest value the hardware can physically store.
Experimental Setup
The effectiveness of FAQ was evaluated across several standard benchmarks and architectures:
- Datasets: CIFAR-10, CIFAR-100, and the large-scale ImageNet dataset.
- Architectures: A diverse range of models including ResNet-18, VGG11, VGG16, AlexNet, and MobileNetV2.
Key Insights
Faults as Quantization Constraints
By reframing hardware defects as mathematical constraints on available quantization levels, FAQ transforms a complex hardware failure into a simple search-and-replace optimization. This shift allows for rapid deployment across thousands of unique chips.
Synergy with Retraining
While FAQ can function as a standalone solution, it also serves as an excellent "warm start" for retraining. The researchers observed that applying FAQ prior to a retraining cycle allows models to recover accuracy significantly faster than starting from a naive state.
Layer Sensitivity
The study confirms that the first and last layers of a DNN are disproportionately sensitive to memory faults. Protecting these specific layers—perhaps through more robust hardware or ECC—while allowing faults in intermediate layers significantly enhances overall system robustness.
Results
The experimental results demonstrate that FAQ provides a high-performance, low-overhead alternative to traditional methods:
- Accuracy Recovery: In scenarios with a 4% bit-fault rate (0.04), FAQ improved the accuracy of ResNet-18 on CIFAR-10 by an average of 76.38% compared to unmitigated faulty hardware.
- Efficiency: The conversion process is remarkably fast, typically completing in under 3 seconds on a standard CPU. This represents less than 5% of the time required for a single epoch of traditional retraining.
- High-Fault Resilience: At a 10% fault rate (0.1), traditional retraining recovered only 4.36% accuracy, whereas the combination of FAQ and retraining achieved a 66.97% recovery.
Conclusion and Future Work
FAQ offers a commercially viable path to improving chip yield for DNN accelerators by mitigating the impact of permanent stuck-at faults in weight memory. While highly effective, the current scope is limited to static weight memory faults. Future research may extend these principles to dynamic faults (e.g., soft errors from cosmic rays), activation memory protection, and emerging non-volatile memory technologies like ReRAM, which exhibit higher defect rates.
Summary updated