Introduction
As hardware accelerators become integral to data centers and safety-critical applications like autonomous driving, ensuring their reliability is paramount. A significant threat to these systems is the "soft error"—transient faults typically caused by cosmic radiation or electrical noise that result in bit-flips within registers. When these occur in control logic, they can lead to Silent Data Corruption (SDC) or catastrophic system hangs.
While traditional solutions like Triple Modular Redundancy (TMR) provide high reliability, they incur substantial silicon area overhead. Moreover, TMR is often blind to errors originating from faulty upstream inputs, as all redundant modules process the same corrupted data. This research proposes a more area-efficient, "in-situ" monitoring approach capable of detecting both internal bit-flips and propagated input errors.
Methodology: On-Chip Monitoring
The researchers developed a three-step framework for implementing hardware monitors, focusing on two distinct but complementary approaches: Specification-Derived Petri Nets and Behavior-Derived State Sequences.
1. Specification-Derived Petri Nets
Petri nets are formal mathematical models used to describe discrete event systems. In this approach, "event sets" are manually extracted from hardware specifications (e.g., the sequence of "Initiate processing" to "Complete calculation"). These are then translated into small, dedicated hardware circuits that monitor control signals in real-time.
If the monitored hardware attempts a transition that violates the logic defined by the Petri net, a fault flag is immediately raised. This method is highly precise for well-documented control flows and extremely area-efficient.
2. Behavior-Derived State Sequences
For complex designs where formal specifications may be incomplete, the authors propose a behavior-learned approach. This method "learns" valid operational sequences from error-free RTL simulations, recording valid combinations of control signals. A "sequence checker" circuit then compares live hardware transitions against a lookup table of known-good sequences. This approach is more automated and adaptable to non-deterministic behaviors.
Experimental Evaluation
The proposed monitors were validated across four hardware benchmarks:
- CNN Accelerator: A quantized LeNet-5 convolutional layer.
- Gaussian Blur: An image processing pipeline.
- AES Encryption: A standard security primitive.
- Network-on-Chip (NoC) Router: A 4x4 mesh router handling complex traffic patterns.
The evaluation involved thousands of fault injection simulations, targeting both internal control registers and primary control inputs.
Key Results and Insights
The study yielded several critical findings regarding the efficiency and efficacy of in-situ monitoring:
- Detection Performance: The monitors achieved error detection rates between 48% and 100%. Notably, in the CNN accelerator, a mere 1% area overhead was sufficient to reach a 93.7% detection rate for internal faults.
- Low Latency: Errors were detected rapidly, often within 1 to 4 clock cycles for the AES implementation.
- Input Vulnerability: A major advantage of these sequence-based monitors is their ability to catch errors stemming from faulty inputs—a scenario where traditional TMR often fails. By checking if the behavior is logical, the monitors remain effective regardless of the error's source.
- Scalability and Complexity: While Petri nets excel in area-constrained environments with clear specs, state sequences are better suited for complex behaviors, such as the multicast patterns in NoC routers, which required a higher area overhead (~12.5%) to achieve a 94.5% detection rate.
Conclusion
This work demonstrates that specification-derived and behavior-derived monitors provide a robust, low-overhead alternative to traditional redundancy for hardware error detection. By focusing on control-flow logic and hierarchical monitoring, designers can achieve high reliability with minimal silicon cost. Future work will look toward automating the Petri net generation process using Large Language Models and exploring hybrid monitoring systems.
Summary updated