Impact of device level faults in a digital avionic processor

SeungJun Kim, Ravishankar K. IyerDigital 1988 · DOI

Abstract

This study presents an experimental analysis of gate and device-level faults within the processor of a Bendix BDX-930 flight control system. Utilizing mixed-mode simulation, faults were injected at both the gate (stuck-at) and transistor levels to observe their propagation to output pins. The findings indicate a significant discrepancy between the traditional stuck-at fault model and actual device-level behavior regarding error detection and activity. While the models align in terms of error activity outside the injected unit, the stuck-at model overestimates fault propagation probability to output pins by over 100%. Furthermore, it underestimates the impact of internal chip faults on output pins, specifically regarding Mean Error Durations (MED). This research also quantifies the impact of device faults based on internal location and output pin manifestation.

Introduction

In the design of flight-critical systems, such as the Bendix BDX-930 flight control system, hardware reliability is paramount. Engineers traditionally rely on "stuck-at" fault models—where a logic gate is simulated as permanently fixed at a '0' or '1' state—to validate fault tolerance. However, physical failures in the real world often occur at the transistor level, involving complex phenomena like oxide breakdowns or metal shorts.

This study investigates whether the computationally efficient stuck-at model serves as a reliable proxy for these intricate physical failures. Inaccuracies in these models could lead to over-engineering for irrelevant failure modes or, more critically, overlooking vulnerabilities in flight hardware.

Methodology

The research employed a mixed-mode simulation approach to achieve a balance between computational speed and physical accuracy. The target hardware was the AMD 2901 four-bit microprocessor slice, a fundamental component of the Bendix BDX-930 CPU.

The simulation environment integrated an event-driven gate-level logic simulator with SPICE2, a device-level circuit simulator. The study involved the injection of 700 device-level faults (comprising 68% oxide and 32% metal faults) and 300 gate-level stuck-at faults. During fault activity, the processor executed a standard CPU self-test program, which included cyclic RAM tests, ALU tests, and memory address processor tests. Key metrics tracked included error propagation across twelve functional units, Mean Time Between Errors (MTBE), and Mean Error Durations (MED) at the output pins.

Results and Discussion

The experimental results reveal several critical insights into the limitations of standard fault modeling:

Internal vs. External Activity

There is minimal correspondence between stuck-at and device-level models concerning activity within a functional unit. Device-level faults were detected more frequently within their own unit (5.1%) compared to stuck-at faults (1.5%). This is attributed to the fact that device faults must propagate through multiple physical transistor layers before exiting the functional unit.

Propagation Discrepancies

The stuck-at model proved to be highly pessimistic regarding fault propagation. It predicted a pin detection rate of 26.9%, which is more than double the actual 12.7% observed at the device level. Conversely, the model was overly optimistic regarding the severity of errors. It underestimated the Mean Error Duration (MED) by 62%, failing to account for the tendency of physical faults to "latch" or sustain errors on output pins significantly longer than gate-level models suggest.

Fault Characteristics and Instruction Sensitivity

While oxide faults are more frequent, faults occurring in metal layers were found to affect the system more actively. Additionally, fault detection was highly dependent on the software context. For instance, "Jump" instructions showed the highest probability of fault detection, whereas "Store" instructions showed the lowest. The study also identified a 15% probability of "near-coincident errors," where multiple errors occur within seven clock cycles of a single device-level fault.

Conclusion

The study demonstrates that while the stuck-at fault model is a staple of hardware testing, it fails to accurately reflect the nuances of physical device-level failures in avionic processors. It simultaneously overestimates propagation probability and underestimates error duration. Future work should focus on developing more realistic "pin error" injection profiles. By mimicking the measured MTBE and MED of actual device faults at the chip pins, engineers can perform more accurate system-level testing without the prohibitive cost of full transistor-level simulation.

Summary updated

Abstract

This study describes an experimental analysis of the impact of gate and device-level faults in the processor of a Bendix BDX-930 flight control system. Via mixed mode simulation, faults were injected at the gate (stuck-at) and at the transistor levels and, their propagation through the chip to the output pins was measured. The results show that there is little correspondence between a stuck-at and a device-level fault model, as far as error activity or detection within a functional unit is concerned. In so far as error activity outside the injected unit and at the output pins are concerned, the stuck-at and device models track each other. The stuck-at model, however, overestimates, by over 100 percent, the probability of fault propagation to the output pins. An evaluation of the Mean Error Durations and the Mean Time Between Errors at the output pins shows that the stuck-at model significantly underestimates (by 62 percent) the impact of an internal chip fault on the output pins. Finally, the study also quantifies the impact of device fault by location, both internally and at the output pins.
BibTeX
@inproceedings{kim1988impact,
  author = {SeungJun Kim and Ravishankar K. Iyer},
  title = {{Impact of device level faults in a digital avionic processor}},
  booktitle = {Digital},
  year = {1988},
  doi = {10.2514/6.1988-3904},
}
Figure