«Chapter 2 Sources of Variation Variations in process, supply voltage and temperature (PVT) have always been an issue in Integrated Circuit (IC) ...»
Sources of Variation
Variations in process, supply voltage and temperature (PVT) have always been an
issue in Integrated Circuit (IC) Design. In digital circuits, PVT ﬂuctuations affect
the switching speed of the transistors and thus the timing of the logic. To guarantee
fault-free operation for a speciﬁed clock frequency, IC designers have to quantify
these uncertainties and account for them adequately. This is typically done by guardbanding, i.e. adding sufﬁcient voltage safety margin to ensure proper working even under worst-case condition.
At recent technology nodes, transistor characteristics are more and more inﬂuenced also by aging effects. These wear-out effects, namely hot carrier injection (HCI) and bias temperature instability (BTI), degrade the drive current of transistors during use. Hence, further safety margin has to be added, dependent on the speciﬁed lifetime of a product.
The following four sections will give an overview of process, voltage and temperature variations as well as aging (PVTA). The necessary fundamentals are brieﬂy explained and the impact on circuit-level timing is discussed.
2.1 Process Variations As stated in the 2011 International Technology Roadmap for Semiconductors (ITRS) , “One of the key problems that designers face due to further shrinking of feature sizes is the increasing variability of design-related parameters, resulting either from variations of fabrication parameters or from the intrinsic atomistic nature which affects, e.g., channel doping.” The sources for the stated device variability are treated in the following. Afterward, the impact of process variations on digital circuits is discussed.
M. Wirnshofer, Variation-Aware Adaptive Voltage Scaling for Digital CMOS Circuits, 5 Springer Series in Advanced Microelectronics 41, DOI 10.1007/978-94-007-6196-4_2, © Springer Science+Business Media Dordrecht 2013 6 2 Sources of Variation Fig. 2.1 Diffraction causes variations in width and length of transistors. Optical proximity correction (OPC) is used to attenuate these effects  2.1.1 Imperfections of the Manufacturing Process 126.96.36.199 Variations in Critical Dimensions The wavelength of light, used for the lithography process, remained at λ = 193 nm since the 130 nm node . Thus, it is exceeding more and more the minimum feature sizes, i.e. critical dimensions (CD), of transistors and the device structure gets increasingly blurred due to diffraction, see Fig. 2.1(c). Chip manufacturer try to reduce this effect by applying immersion lithography, phase-shift masks (PSM) and optical proximity correction (OPC), see Fig. 2.1(b). However, the relative ﬂuctuations in transistor dimensions are rather increasing, as it is very challenging to reduce the absolute deviations in the same way as the shrinking feature sizes.
Variations of the transistor’s width W and length L directly affect its drive current, which is proportional to W/L. Channel length ﬂuctuations also change the threshold voltage Vt (roll-off effect) and this way additionally shift the drive current.
Note that the channel length also varies randomly across the width of the transistor.
The so-called line edge roughness (LER) is caused by statistical variations of the photon count or imperfections during photoresist removal . LER increases the Ioff /Ion current ratio for short channel devices used in digital circuits .
188.8.131.52 Random Dopant Fluctuation
The channel region of a transistor is doped with impurity atoms. These atoms are randomly placed into the channel—by techniques like dopant implantation— leading to statistical variations in the actual number of implanted impurities. Such a change of the carrier concentration shifts the threshold voltage and thus the drive strength of the transistor. In older technologies, with thousands of dopant atoms per channel region, an absolute deviation by several atoms was negligible. In recent technologies however the nominal number of impurities is only in the range of tens, see Fig. 2.2, leading to increased mismatch due to random dopant ﬂuctuation (RDF).
2.1 Process Variations 7 Fig. 2.2 Number of dopant atoms per channel region over technology nodes  184.108.40.206 Variation of the Gate Oxide Thickness The gate oxide can be grown with an absolute accuracy of 1–2 inter atomic layers.
In former technologies with an oxide thickness (TOX) of tens of inter atomic layers and large gates, TOX induced Vt variations were almost negligible. However, in technologies below 30 nm with oxide thicknesses between 1–3 nm (approx. 5– 15 inter atomic spacings), TOX variations can contribute to the threshold voltage uncertainty as much as RDF .
2.1.2 Global and Local Process Variations Usually, process variations are categorized into global and local variations. For global variations device parameters, such as oxide thickness or dopant concentrations, change equally for all transistors. Wafer-to-wafer or lot-to-lot variations fall into that category. In contrast, for local variations—also known as mismatch or random uncorrelated variations—each transistor is affected differently.
In other words, variations are distinguished by their spatial correlation distance.
For local variations there is no correlation, whereas for global variations the correlation distance is very large. Note that there are also variation effects in between, leading to die-to-die and within-die variations. However, to cope with process ﬂuctuations, it is still common practice to categorize them into global and local variations .
2.1.3 Impact of Process Variations on Digital Circuits Variations of physical parameters (e.g. oxide thickness or doping concentration) lead to variations of electrical parameters, like threshold voltage or gate capacitance. In turn, this affects the performance of digital circuits as it changes gate delays or leakage currents.
8 2 Sources of Variation Fig. 2.3 Effect of global variations on the path delay. The depicted histogram results from MC simulations. The arrows represent the simulation results for fast and slow corner 220.127.116.11 Global Variations By performing Monte-Carlo (MC) simulations on transistor level, the effect of global variations on the circuit behavior can be explored. Probability distributions for the varying device parameters serve as input for MC simulations. The distributions are based on measured statistics of the manufactured transistors. Therefore, IC manufacturers add Process Control Monitoring (PCM) structures on the scribe-line of the wafers.
The histogram in Fig. 2.3 illustrates the delay of an inverter chain in 65 nm technology under global variations. The delay values of 2000 MC-runs are divided into bins of 12 ps width. As MC simulations are computationally intensive and hence hardly feasible for large designs, corner simulations are typically used to evaluate the impact of global variations.
The corners represent extreme cases, where the devices ultimately diverge from their nominal characteristics. For the fast corner, all process ﬂuctuations increase the drive current of a transistor, leading to maximum speed. At slow corner, a device is ultimately slowed down by the process variations. The simulation results for fast and slow corner are also drawn into Fig. 2.3.
Fast and slow corner are typically deﬁned by multiples of σglobal (e.g. three or six sigma) or founded on measured performance statistics of so-called split lots. For split lots, the manufacturing process (doping concentration, oxide thickness etc.) is intentionally biased to obtain extremely slow and fast chips. Besides fast and slow corner, also cross corners exist with maximum p-FET and minimum n-FET speed and vice versa. Cross corners are often critical in analog circuits, but are of minor importance in digital designs.
2.1 Process Variations 9 Fig. 2.4 Effect of local process variations on the path delay 18.104.22.168 Local Variations Local variations, which are by nature uncorrelated, are increasing due to the scaling in CMOS technologies. This comes from the fact that with decreasing transistor dimensions, the standard deviation of threshold voltage σ V t and current factor σ k/k (k = μCox W/L) are increasing, since they are proportional to the inverse of the square root of the active device area 
Fig. 2.5 Clock signal, current proﬁle and resulting supply voltage of a 32-bit multiplier circuit. Close to the clock edges, a high current ﬂow can be seen due to the switching activity of the circuit and ﬂip-ﬂops. The combination of IR-drop and di/dt noise in turn leads to voltage drops and—if resonance occurs—also to overshoots It is a common pitfall, to think that the absolute uncertainty σt,d decreases for long paths. However, only the relative uncertainty σt,d /td decreases with the path length, whereas the absolute uncertainty σt,d naturally increases, see Fig. 2.4.
2.2 Voltage Variations 2.2.1 Origin of Voltage Fluctuations Supply voltage ﬂuctuations are mainly caused by IR drop and di/dt noise. IR drop is caused by the current ﬂow over the parasitic resistance of the power grid, whereas di/dt noise is due to the parasitic inductance in combination with capacitance and resistance of power grid and package. These fast changing effects—also called power noise—typically have time constants in the range of nano- to microseconds. Figure 2.5 shows an example of a supply current proﬁle together with the resulting supply voltage sequence for a 32-bit multiplier circuit.
Considering only IR drop, the voltage bounce VIR drop is given by Ohms law
Fig. 2.6 Relation between supply voltage and logic delay, illustrated by the most critical path of a multiplier circuit in 65 nm technology.
The simulation results can be accurately ﬁtted by Eq. 2.7 2.2.2 Impact of Voltage Changes on the Path Delay
with path-speciﬁc parameters a, b and Vt. Figure 2.6 shows the simulated path delay of a multiplier circuit for different supply voltages. By applying the appropriate input pattern, the most critical path was sensitized in the simulation. The solid line demonstrates the ﬁt by Eq. 2.7.
2.3 Temperature Variations Dependent on the thermal conductivity, the dissipated power affects the temperature of a chip. Power dissipation hence leads to global temperature variations as well as local ﬂuctuations in regions of high-activity, so called hot-spots. Additionally, ambient temperature changes lead to global shifts in chip temperature.
Temperature ﬂuctuations typically have time constants in the range of milliseconds to seconds . An increase in temperature typically causes a circuit to slow down due to reduced carrier mobility and increased interconnect resistance, see 12 2 Sources of Variation
Fig. 2.7 Inﬂuence of temperature on device characteristic and path delay
Fig. 2.7(b) (VDD = 1.2 V) and . However, for low VDD the circuit is operated in temperature inversion. Here, the effect of decreasing threshold voltage with temperature exceeds the mobility degradation, see Fig. 2.7(a). Consequently, the circuit exhibits an inverted temperature characteristic, as it speeds up with increased temperature and vice versa, see Fig. 2.7(b) (VDD = 0.8 V and VDD = 0.9 V).
2.4 Aging Due to increasing electrical ﬁelds and new materials, transistor wear-out is of increasing concern in recent technologies. BTI- as well as HCI-effects degrade the speed of transistors during their lifetime and demand for additional safety margin.
2.4.1 Hot Carrier Injection (HCI) Hot Carrier Injection mainly occurs during switching of logic gates. Carriers are accelerated in the lateral ﬁeld under the oxide and gain sufﬁcient kinetic energy to be injected into the gate dielectric. This effect is illustrated for an n-MOSFET in Fig. 2.8. The trapped charge increases the threshold voltage of the device and reduces its current drivability. Recent research on HCI is presented in .
2.4.2 Bias Temperature Instability (BTI)
Fig. 2.8 Hot Carrier Injection (HCI) for an n-MOSFET Fig. 2.9 Bias Temperature Instability (BTI) for an n-MOSFET Temperature Instability (PBTI). Accordingly, the term NBTI is used in the case of a p-FET. PBTI and NBTI increase the threshold voltage of the device and slow down the switching speed .
Note that BTI-aging is caused by charge trapping and detrapping with a wide range of capture and emission times . Therefore, small Vt -shifts can be observed already after very short stress times down to microseconds . However, due to the distribution of the capture and emission time constants, considerable Vt -shifts arise only after days, weeks or even years. More details on BTI can be found in [28–31].
2.5 Summary Figure 2.10 categorizes the discussed variations according to their time constants.
As the dynamics of variation effects depend on various circumstances, the shown classiﬁcation is rather qualitatively.
Process variations occur during fabrication and lead to ﬁxed changes in device parameters. During operation, device characteristics can, however, still be affected by aging induced wear-out. As mentioned before, the trapping events leading to device aging have a wide variety of time constants. However, signiﬁcant degradation of CMOS logic delays mainly develops in the long run. The temperature, also inﬂuencing circuit speed, changes rather slowly, whereas power noise has normally very short time constants.
Fig. 2.10 Temporal classiﬁcation of variations 14 2 Sources of Variation The shorter the time constant of a variation effect, the more challenging becomes the voltage adaptation to it. In the following chapter, it is explained how Pre-Error AVS effectively handles slow as well as fast changing effects compared to state-ofthe-art approaches.