Algorithms for Power Aware Testing of NanometerDigital ICs

At-speed testing of deep-submicron digital very large scale integrated (VLSI) circuits has become mandatory to catch small delay defects. Now, due to continuous shrinking of complementary metal oxide semiconductor (CMOS) transistor feature size, power density grows geometrically with technology scaling. Additionally, power dissipation inside a digital circuit during the testing phase (for test vectors under all fault models (Potluri, 2015)) is several times higher than its power dissipation during the normal functional phase of operation. Due to this, the currents that flow in the power grid during the testing phase, are much higher than what the power grid is designed for (the functional phase of operation). As a result, during at-speed testing, the supply grid experiences unacceptable supply IR-drop, ultimately leading to delay failures during at-speed testing. Since these failures are specific to testing and do not occur during functional phase of operation of the chip, these failures are usually referred to false failures, and they reduce the yield of the chip, which is undesirable. In nanometer regime, process parameter variations has become a major problem. Due to the variation in signalling delays caused by these variations, it is important to perform at-speed testing even for stuck faults, to reduce the test escapes (McCluskey and Tseng, 2000; Vorisek et al., 2004). In this context, the problem of excessive peak power dissipation causing false failures, that was addressed previously in the context of at-speed transition fault testing (Saxena et al., 2003; Devanathan et al., 2007a,b,c), also becomes prominent in the context of at-speed testing of stuck faults (Maxwell et al., 1996; McCluskey and Tseng, 2000; Vorisek et al., 2004; Prabhu and Abraham, 2012; Potluri, 2015; Potluri et al., 2015). It is well known that excessive supply IR-drop during at-speed testing can be kept under control by minimizing switching activity during testing (Saxena et al., 2003). There is a rich collection of techniques proposed in the past for reduction of peak switching activity during at-speed testing of transition/delay faults ii in both combinational and sequential circuits. As far as at-speed testing of stuck faults are concerned, while there were some techniques proposed in the past for combinational circuits (Girard et al., 1998; Dabholkar et al., 1998), there are no techniques concerning the same for sequential circuits. This thesis addresses this open problem. We propose algorithms for minimization of peak switching activity during at-speed testing of stuck faults in sequential digital circuits under the combinational state preservation scan (CSP-scan) architecture (Potluri, 2015; Potluri et al., 2015). First, we show that, under this CSP-scan architecture, when the test set is completely specified, the peak switching activity during testing can be minimized by solving the Bottleneck Traveling Salesman Problem (BTSP). This mapping of peak test switching activity minimization problem to BTSP is novel, and proposed for the first time in the literature. Usually, as circuit size increases, the percentage of don’t cares in the test set increases. As a result, test vector ordering for any arbitrary filling of don’t care bits is insufficient for producing effective reduction in switching activity during testing of large circuits. Since don’t cares dominate the test sets for larger circuits, don’t care filling plays a crucial role in reducing switching activity during testing. Taking this into consideration, we propose an algorithm, XStat, which is capable of performing test vector ordering while preserving don’t care bits in the test vectors, following which, the don’t cares are filled in an intelligent fashion for minimizing input switching activity, which effectively minimizes switching activity inside the circuit (Girard et al., 1998). Through empirical validation on benchmark circuits, we show that XStat minimizes peak switching activity significantly, during testing. Although XStat is a very powerful heuristic for minimizing peak input-switchingactivity, it will not guarantee optimality. To address this issue, we propose an algorithm that uses Dynamic Programming to calculate the lower bound for a given sequence of test vectors, and subsequently uses a greedy strategy for filling don’t cares in this sequence to achieve this lower bound, thereby guaranteeing optimality. This algorithm, which we refer to as DP-fill in this thesis, provides the globally optimal solution for minimizing peak input-switching-activity and also is the best known in the literature for minimizing peak input-switching-activity during testing. The proof of optimality of DP-fill in minimizing peak input-switching-activity is also provided in this thesis.