Large-scale statistical analysis of early failures in Cu electromigration, Part I: Dominating mechanisms

Abstract

With continuing scaling of Cu-based metallization, the electromigration ͑EM͒ failure risk has remained one of the most important reliability concerns for advanced process technologies. The main factors requiring attention are the activation energy related to the dominating diffusion mechanism, the current exponent as well as the median lifetimes and lognormal standard deviation values of experimentally acquired failure time distributions. In general, the origin and scaling behavior of these parameters are relatively well understood. However, the observation of strong bimodality for the electron up-flow direction in dual-inlaid Cu interconnects has added complexity. The failure voids can occur both within the via ͑"early" mode͒ or within the trench ͑"late" mode͒. Over the last few years, bimodality has been reported also in down-flow EM, leading to very short lifetimes due to small, slit-shaped voids under vias. These voids, requiring only a very limited amount of mass movement, are generally causing concerns with respect to long-term, reliable chip operation at elevated temperatures. For a more thorough investigation of the aforementioned early failure phenomena, specific test structures were designed based on the Wheatstone Bridge ͑WSB͒ technique. The use of these structures enabled an increase in the tested sample size past 800 000 for the 90 nm technology node, allowing a direct analysis of EM failure mechanisms at the single-digit ppm regime. Results indicate that down-flow EM can exhibit bimodality at very small percentage levels, not readily identifiable with standard testing methods. The activation energy for the down-flow early failure mechanism was determined to be 0.83Ϯ 0.01 eV. Within the small error bounds of this large-scale statistical experiment, this value is deemed to be significantly lower than the usually reported activation energy of 0.90 eV for EM-induced diffusion along Cu/SiCN interfaces. Due to the advantages of the WSB technique, we were also able to expand the experimental temperature range down to 150°C, coming quite close to typical operating conditions up to 125°C. As a result of the lowered activation energy, we conclude that the down-flow early failure mode may control the chip lifetime at operating conditions. This publication contains the first part of our large-scale statistical analysis of early failures in Cu EM. In the second part of this study, we will discuss the EM scaling behavior across 90, 65, and 45 nm technologies. In addition, short-length effects will be evaluated using our large-scale, statistical approach. Utilizing the advantages of the WSB technique, the total sample size will be increased past 1.2 million.

Electromigration void nucleation and growth analysis using large-scale early failure statistics

“…the nonfailed links in the EM failure distribution have been included in the data analysis as censored values to enable a direct comparison to the singlelink data. Details of this calculation were reported earlier [9,11]. The lifetime distributions of differently sized test structures line up well.…”

“…Besides significantly reducing testing time and thus enabling the evaluation of more testing conditions, the use of WSB structures provides data for the early failure mode, which defines the chip lifetime. Details describing this methodology were published previously [9,11]. The EM test matrices using WSB devices are shown in Fig.…”

“…Furthermore, the already mentioned bimodality causes additional concern whether the generally used values are appropriate. It was already shown that the activation energy for the early and late failure modes can be significantly different with the downflow early mode showing a slight, but significant reduction to 0.83 ± 0.01 eV [9]. Also, another study seemed to indicate that the n values for early and late modes can differ significantly [10].…”

Stress relaxation in pulsed DC electromigration measurements

“…Note that due to differences of scale, the x-axes vary between plots. The data appear to follow a bimodal distribution such as that seen by Hauschildt et al, 5 which disrupts the fit to the lognormal distribution shown in the plot and thereby increases the apparent standard deviation when plotted as a single mode lognormal. The assumption was made that either failure mode would react in the same way to any possible relaxation, so that separating the modes would not be necessary for this study.…”

Electromigration degradation mechanism analysis of SnAgCu interconnects for eWLB package

“…It is therefore relevant to use directly the failure data of the whole chain. In the case of computing lifetime for a single link, failure data extracted directly from the whole chain should have been deconvoluted through conditional reliabilities (in order to include non-failed links) [10]. Stress conditions and sampling are resumed in Table 1.…”