Quantized AVF : A Means of Capturing Vulnerability Variations over Small Windows of Time

Architectural vulnerability factor (AVF) is the probability that a transient fault in a bit, gate, or transistor becomes a user-visible error. AVFs vary widely across time, applications, and bits. Usually AVFs averaged over time and across applications are used to compute the overall soft error rate of a processor. Average AVFs, however, cannot express the short-term vulnerability variations of a bit as they tend to settle down to a fixed value over time. To quantify the vulnerability of bits over short durations, we introduce the concept of Quantized AVF (Q-AVF). Q-AVF expresses the vulnerability of a bit to soft errors over short intervals of time. The average AVF of a bit for a specific interval can be computed as a weighted average of Q-AVFs of all the quanta in that interval. Our analysis of Q-AVF shows significant run-time variation—as much as 80% or more for certain applications. By capturing vulnerability variations over short windows of time, Q-AVFs provide better opportunities for reducing the performance and power overhead of reliability solutions at run-time. To compute Q-AVF in hardware, linear regression analysis is used to create highly accurate equations that can be implemented with eight simple parameters. These parameters can accurately track Q-AVFs of various structures throughout the processor pipeline. Implementing these equations with fewer parameters is critical to reduce the complexity of run-time Q-AVF tracking, thereby making Q-AVF estimation in hardware practical.