Negative Bias Temperature Instability Lifetime Prediction: Problems and Solutions Z. Ji (1) , S. F. W. M. Hatta (1) , J. F. Zhang (1) , J. G. Ma (1) , W. Zhang (1) , N. Soin (2) , B. Kaczer (3) , S. De Gendt (3) , and G. Groeseneken (3) (1) School of Engineering, Liverpool John Moores University, Byrom Street, Liverpool L3 3AF, UK (2) Dept. Elec. Eng, University of Malaya, Malaysia (3) IMEC, Leuven B3001, Belgium Abstract Lifetime of pMOSFETs is limited by NBTI. Conventional slow measurement overestimates lifetime due to recovery. The fast techniques suppress recovery, but cannot give reliable prediction. This work proposes a new lifetime prediction technique that overcomes the shortcomings of both slow and fast methods, based on the A s-grown-G eneration (AG) model. Its advantages over those based on Reaction- Diffusion (RD) and two-stage models include its simple algorithm, only two fitting parameters at a given temperature, and no need for a kinetic model for the as-grown hole traps. This makes it readily implementable in industrial laboratories for process screening. Introduction NBTI is a critical reliability issue for advanced CMOS technology [1-15]. Manufacturers are using NBTI lifetime, i.e. τ, as one of the criteria for determining the operation voltage and a figure-of-merit for process screening [1]. The conventional τ prediction is based on slow measurements that only capture part of the degradation due to the recovery [2-6]. For VLSI circuits, one cannot rule out that some gates rarely flip with little recovery. Fast techniques were used to suppress recovery [2-7], but there is no industry-wide accepted τ prediction technique based on them, because of the problems detailed below. The objective of this work is to propose a new technique for reliably predicting the worst- case lifetime and the maximum operation voltage. Results and Discussions A. Problems with prediction based on fast measurements After suppressing recovery, one would expect that the larger degradation shortens τ. It was reported, however, that the extracted τ can be either similar (Fig. 1a) [6] or even longer (Fig. 1b) [2]. This cannot be true. Moreover, τ prediction requires extrapolation from high stress Vg_st to low operational Vg_op. After suppressing recovery, log(τ) versus log(|Vg_st|) does not always follow a straight line (Figs. 1c&d) [7], making the extrapolation unreliable. Since a reliable prediction is not available without recovery, some researchers [8] purposely inserted a delay between stress and measurement to give a level of recovery. It is not known, however, what level of recovery should be used. B. Two groups of defects NBTI consists of two group of defects: as-grown hole traps (AHT) and generated defects (GD) [9-15]. Fig. 2 shows that, instead of following a power-law, the degradation saturates quickly at short time (symbol ‘□’) [5]. This is a signature of AHT being filled up. Once the AHTs are removed from the total degradation, the remaining part follows a power-law well (symbol ‘●’), which is a feature of generated defects (GD). As a result, the NBTI kinetics under a given stress temperature can be described by an ‘AG’ model [5]: ΔVt=A+Gt n , where ‘A’ and ‘Gt n ’ represent ΔVt(AHT) and ΔVt(GD), the contribution from AHT and GD to NBTI, respectively. ‘A’ will be experimentally determined, so that it does not need a model. ‘G’ and ‘n’ are the only two fitting parameters. C. New method for extracting AHT To reliably predict τ, GD and AHT must be separated. The clear separation and the ‘shoulder’ in Fig. 2 are not always present, so that it cannot be used as a general method for the separation and a new method is needed. Fig. 3a gives the Vg waveform. After stress under Vg_st for a pre-specified time, Vg was stepped towards positive to Vg_op1 to discharge and ∆Vt is monitored (Fig. 3b) during the pulse edges (3 µs) at Id=100nA*W/L. After completing discharge at Vg_op1, Vg was stepped to Vg_op2 and the same procedure is followed until Vg reaches certain pre-defined level [10]. The Vg_st then was re-applied. Fig. 4 shows ΔVt discharged at each Vg_op after stressed for different time. The Vg_op was converted to Ef-Ev at the interface by following the procedure described in [10]. It is clear that below E(AHT=0), the AHTs are fully filled after only 1 sec, since they do not increase further with stress time. This is confirmed by the parallel shift of the three curves, supporting their “as-grown” nature [9-15]. Above E(AHT=0), defects are negligible at 1 sec but increase with stress time, indicating that they are generated defects (GD). To separate ∆Vt(AHT) from ∆Vt(GD), Vg_op(AHT=0) corresponding to E(AHT=0) was first found by stressing the device for 1 sec as shown in Fig. 4. The device was then stressed for longer time and for each stress time, ΔVt was monitored after discharging at different Vg_op until Vg_op(AHT=0) was reached, as shown in Fig. 5a. Since ∆Vt(AHT)=0 here, ΔVt at Vg_op(AHT=0) (‘●’ in Fig. 5a) can originate only from GD, allowing the extraction of ΔVt(GD), which follows a power law well [5,16]. A more negative Vg_op lowers Ef-Ev below E(AHT=0), charging some AHTs, so that ΔVt(AHT) can be evaluated from ΔVt(Vg_op)-ΔVt(GD). If the separation method is correct, one expects that |ΔVt(AHT)| should only depend on |Vg_op|, but not stress time, when the stress time is long enough to fill up the AHTs. Fig. 5b confirms |ΔVt(AHT)| rising with |Vg_op|, but not IEDM13-413 15.6.1 978-1-4799-2306-9/13/$31.00 ©2013 IEEE