Algorithms for Estimating Number of Glitches and Dynamic Power in CMOS Circuits with Delay Variations

Abstract

Dynamic power dissipation of a CMOS VLSI circuit depends on the signal activity at gate outputs. The activity includes the steady-state logic transitions as well as glitches. The latter are a function of gate delays, which, for modern VLSI circuits, have wide process-related variations. Both average and peak power dissipation are useful and are traditionally estimated by Monte Carlo simulation. This is expensive and the accuracy, especially for peak power,depends upon the number of circuit delay samples that are simulated. We present an alternative. We use zero-delay simulation of a vector pair to determine the steady-state logic activity. We derive linear-time algorithms that, using delay bounds for gates, determine the maximum, minimum and average number of transitions that each gate output can produce. From this information, we estimate the average and peak energy consumed by each vector pair in a given vector set. For a set of random vectors applied to c7552 circuit, our analysis determined the per-vector energy consumption as 82.2 picojoules average and 196.3 picojoules peak. In comparison, Monte Carlo simulation of 1,000 circuit samples gave 82.8 picojoules average and 146.1 picojoules peak. The discrepancy of the peak consumption will reduce if more samples were simulated in the Monte Carlo method. Even with 1,000 samples the CPU time of the Monte Carlo analysis was three orders of magnitude greater than the alternative method we offer in this paper.