Signal and power integrity challenges in VLSI circuits and strategies for their mitigation Independent Study Report (EES837) Under the guidance of: Dr. Kaushik Saha By Pushpak Dagade Department of Electrical Engineering Indian Institute of Technology, Delhi ➞ Pushpak Dagade, 2013. All rights reserved. Contents List of Figures v List of Tables vii 1 Introduction 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Organization of this Book . . . . . . . . . . . . . . . . . . . . . . . . 2 Interconnect: Modeling and Impacts 2.1 Interconnect Modeling . . . . . . . . 2.1.1 Geometry . . . . . . . . . . . 2.1.2 Model . . . . . . . . . . . . . 2.1.2.1 Resistance . . . . . . 2.1.2.2 Capacitance . . . . . 2.1.2.3 Inductance . . . . . 2.2 Interconnect Impact . . . . . . . . . 2.2.1 Delay . . . . . . . . . . . . . 2.2.2 Crosstalk . . . . . . . . . . . 2.2.3 Power . . . . . . . . . . . . . 1 1 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 5 5 6 7 8 9 10 10 11 13 3 Crosstalk: Effects and Mitigation Techniques 3.1 Effects of Crosstalk . . . . . . . . . . . . . . . . . . . . . . 3.1.1 Crosstalk delay . . . . . . . . . . . . . . . . . . . . 3.1.1.1 A switches and B does not . . . . . . . . . 3.1.1.2 A and B switch in same direction . . . . . 3.1.1.3 A and B switch in opposite directions . . 3.1.2 Crosstalk noise . . . . . . . . . . . . . . . . . . . . 3.1.2.1 The victim is floating . . . . . . . . . . . 3.1.2.2 The victim is being driven . . . . . . . . . 3.2 Crosstalk Mitigation . . . . . . . . . . . . . . . . . . . . . 3.2.1 Intelligent Engineering . . . . . . . . . . . . . . . . 3.2.1.1 Increase spacing to adjacent interconnects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 15 15 16 16 16 17 18 18 19 19 19 i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.2 3.2.3 3.2.1.2 Different switching times for neighbouring interconnects Shielding interconnects . . . . . . . . . . . . . . . . . . . . . . 3.2.2.1 Passive Shields . . . . . . . . . . . . . . . . . . . . . 3.2.2.2 Active Shields . . . . . . . . . . . . . . . . . . . . . . Crosstalk cancellation techniques . . . . . . . . . . . . . . . . 3.2.3.1 Staggered repeaters . . . . . . . . . . . . . . . . . . . 3.2.3.2 Charge compensation . . . . . . . . . . . . . . . . . 3.2.3.3 Twisted differential signalling . . . . . . . . . . . . . 20 20 20 21 21 21 23 23 4 Supply Bounce: Causes, Effects and Mitigation Techniques 4.1 L di/dt Noise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.1 Understanding L di/dt Noise . . . . . . . . . . . . . . . . . . . 4.1.2 Causes for L di/dt Noise . . . . . . . . . . . . . . . . . . . . . 4.1.2.1 Number of switching simultaneously outputs (SSO) . 4.1.2.2 Output Load . . . . . . . . . . . . . . . . . . . . . . 4.1.2.3 Location of output pins . . . . . . . . . . . . . . . . 4.1.2.4 Power supply voltage (VDD ) . . . . . . . . . . . . . 4.2 Effects of L di/dt noise . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.1 Noise Margin Reduction . . . . . . . . . . . . . . . . . . . . . 4.2.2 Propagation Delay Degradation . . . . . . . . . . . . . . . . . 4.2.3 Other issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Mitigation Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.1 Intelligent Engineering . . . . . . . . . . . . . . . . . . . . . . 4.3.1.1 Staggered switching of SSO buffers . . . . . . . . . . 4.3.1.2 Increase number of supply pins . . . . . . . . . . . . 4.3.1.3 Output pin positioning . . . . . . . . . . . . . . . . . 4.3.1.4 Separate supply pins for Analog and Digital circuits 4.3.1.5 Slew rate control . . . . . . . . . . . . . . . . . . . . 4.3.2 Guard Rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.2.1 Passive guard rings . . . . . . . . . . . . . . . . . . . 4.3.2.2 Active guard rings . . . . . . . . . . . . . . . . . . . 4.3.3 On chip decoupling or bypass capacitors (decap) . . . . . . . . 25 26 26 28 28 29 31 31 32 32 33 34 34 34 34 35 35 36 36 36 36 36 36 5 Electromagnetic emission and interference in VLSI circuits 5.1 EMC subcomponents . . . . . . . . . . . . . . . . . . . . . . . 5.2 Antenna . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Emissions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.1 Conducted Emission . . . . . . . . . . . . . . . . . . . 5.3.1.1 Through ground and power supply lines . . . 5.3.1.2 Through I/O . . . . . . . . . . . . . . . . . . 5.3.1.3 Radiated emissions from current loops . . . . 5.3.2 Radiated Emission . . . . . . . . . . . . . . . . . . . . 39 39 40 41 41 41 41 41 41 ii . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4 5.5 5.3.2.1 Common mode radiation . . 5.3.2.2 Differential mode radiation Factors affecting EMC . . . . . . . . . . . . 5.4.0.3 Supply voltage . . . . . . . 5.4.0.4 Frequency . . . . . . . . . . 5.4.0.5 Grounding . . . . . . . . . EMC reduction techniques . . . . . . . . . . 5.5.0.6 Device level noise reduction 5.5.0.7 Routing noise reduction . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 42 42 42 42 42 42 42 43 45 iii iv List of Figures 1.1.1 Interconnect Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . 1 2.1.1 Interconnect Geometry . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.2 SEM image of wire cross-sections in Intel’s 45nm process [2] . . . . . 2.1.3 Lumped RC Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.4 Distributed RC Model . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.5 Transmission Line Model . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.6 Effect of fringing fields on capacitance . . . . . . . . . . . . . . . . . 2.1.7 Multilayer capacitance model of an interconnect. The shaded interconnect in the centre is the interconnect of concern for which the total capacitance is being calculated. . . . . . . . . . . . . . . . . . . . . . 2.1.8 Effect of fringing fields on capacitance . . . . . . . . . . . . . . . . . 2.2.1 Crosstalk noise on interconnect B (victim) due to interconnect A (aggressor) due to coupling capacitor C . . . . . . . . . . . . . . . . . . . 2.2.2 Graphical illustration of 1st , 2nd and 3rd order crosstalk . . . . . . . . 5 6 7 7 7 9 3.0.1 Cross sectional view of 2 interconnects A and B with the relevant capacitances shown . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1.1 A typical scenario where interconnect A switches but B does not . . 3.1.2 A best case scenario where both interconnects A and B switch in the same direction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1.3 A worst case scenario where interconnects A and B switch in the opposite direction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1.4 Propagation delay for the 3 cases summarized in Table 3.1. Delay noise is the difference between delays for the best and worst case scenarios. 3.1.5 Coupling to floating victim . . . . . . . . . . . . . . . . . . . . . . . . 3.1.6 Coupling to driven victim . . . . . . . . . . . . . . . . . . . . . . . . 3.1.7 Waveforms of coupling noise . . . . . . . . . . . . . . . . . . . . . . . 3.2.2 Different passive shielding topologies with shields on one or both sides 3.2.1 Interleaving bit lines of buses A and B to avoid crosstalk delay effects 3.2.3 Active shielding topology. . . . . . . . . . . . . . . . . . . . . . . . . 3.2.4 Staggered repeaters crosstalk control technique . . . . . . . . . . . . . 3.2.5 Charge compensation crosstalk control technique . . . . . . . . . . . v 10 11 12 12 15 16 16 17 17 18 18 19 20 20 21 22 22 3.2.6 Twisted Differential Signalling crosstalk control technique. a and a are aggressor and its compliment. v and v are victim and its compliment (which is also a victim.) . . . . . . . . . . . . . . . . . . . . . . . . . 4.0.1 A cutaway view of dual-in-line package showing the chip-to-package connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.0.2 Chip package interface parasitics . . . . . . . . . . . . . . . . . . . . 4.1.1 Model to understand ground bounce . . . . . . . . . . . . . . . . . . 4.1.2 V vs. time for a step input to model in Figure 4.1.1 . . . . . . . . . . 4.1.3 I vs. time t for a step input to model in Figure 4.1.1 . . . . . . . . . 4.1.4 VGB vs. time t for a step input to model in Figure 4.1.1 . . . . . . . . 4.1.5 Ground and VDD bounce output waveforms in Figure 4.1.1 . . . . . . 4.1.6 Schematic of 16 output buffers switching simultaneously . . . . . . . 4.1.7 Ground bounce waveforms for varying number of outputs switching simultaneously . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.8 Ground bounce amplitude vs. capacitive loading (on all outputs) . . . 4.1.9 Ground bounce amplitude vs. capacitive loading (only on active outputs) 4.1.10Ground bounce amplitude vs. capacitive loading (only on quiet outputs) 4.1.11Ground bounce waveform for pin farthest to the device ground (worst case output pin) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.12Ground bounce waveform for pin closest to the device ground (best case output pin) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1.13Ground bounce amplitude vs. VDD for varying load capacitances . . . 4.2.1 Noise margin for an inverter with and without ground and power bounce issue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.2 Degradation of propagation delay due to supply bounce . . . . . . . . 4.2.3 Layout and schematic of buffers with multiple fingers . . . . . . . . . 4.3.1 Ground bounce with and without staggering of SSO . . . . . . . . . . 4.3.2 Schematic of chip package interface with multiple ground supply pins 4.3.3 Simultaneous switching noise (SSN) vs. number of drivers for varying number of supply pins . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3.4 Passive guard ring configuration . . . . . . . . . . . . . . . . . . . . . 4.3.5 Active guard ring configuration . . . . . . . . . . . . . . . . . . . . . 4.3.6 On chip decoupling capacitor . . . . . . . . . . . . . . . . . . . . . . 5.1.1 EMC: Radiated Emissions . . . . . . . . . . . . . . 5.1.2 EMC: Radiated Susceptibility . . . . . . . . . . . . 5.1.3 EMC: Conducted Emissions . . . . . . . . . . . . . 5.1.4 EMC: Conducted Susceptibility . . . . . . . . . . . 5.2.1 Schematic of radiated electric field from an antenna 5.4.1 Grounding Topologies depending on frequency . . . vi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 25 26 27 27 27 27 28 29 29 30 30 30 31 31 32 33 34 34 35 35 35 36 37 37 40 40 40 40 41 43 List of Tables 1.1 1.2 ITRS 2007 Predictions [1] . . . . . . . . . . . . . . . . . . . . . . . . Effect of scaling on interconnect parameters . . . . . . . . . . . . . . 1 2 2.1 2.2 Intel 45nm metal stack . . . . . . . . . . . . . . . . . . . . . . . . . . Bulk resistivity of some pure metals (at 22 ➦C) . . . . . . . . . . . . . 6 8 3.1 Effective capacitances for various switching cases vii . . . . . . . . . . . 17 Chapter 1 Introduction 1.1 Introduction As transistors become smaller in size, they switch faster, dissipate lesser power and are cheaper to manufacture. The feature size of transistors has been reducing approximately by 30% every 2-3 years. This has been the trend since 1995 and is still ongoing. An International Technology Roadmap for Semiconductors is prepared and maintained by the Semiconductor Industry Association for predicting future scaling [1]. The ITRS 2007 predictions have been shown in Table 1.1 Along with transistors, interconnects are also scaled accordingly. During the LSI & beginning of VLSI phase, transistors were much slower as compared to interconnect delays. Interconnects were wide and thick and thus had negligible resistance. They could be simply modFigure 1.1.1: Interconnect Geometry eled as ideal equipotential nodes. However, in modern VLSI, the transistors switch faster and the interconnects are becoming slower, to the point that in many cases, the interconnect RC delay is more than the gate delay. Year Feature size (nm) Lgate (nm) VDD (V) Transistors/die (in billions) Wiring levels 2012 34 20 1.0 1.5 12 2012 24 14 0.9 3.1 12 2015 17 10 0.8 6.2 13 2018 12 7 0.7 12.4 14 Table 1.1: ITRS 2007 Predictions [1] 1 2021 8.4 5 0.65 24.7 15 Trend ↓ ↓ ↓ ↑ ↑ 1.1. Introduction Parameters Width Spacing Thickness Interlayer oxide height Die size Resistance/length (Rw ) Fringing plate capacitance/length (Cwf ) Parallel plate capacitance/length (Cwp ) Total wire capacitance/length (Cw ) RC delay/length2 (twu ) Crosstalk noise Parameters (local/semiglobal interconnect) Length Unrepeated RC delay Parameters (global interconnect) Length Unrepeated RC delay 2 Formula Scale factor w s t h 1/wt t/s w/h Cwf + Cwp Rw C w w/h 1/S 1/S 1/S 1/S Dc S2 1 1 1 S2 1 Formula Scale Factor ll 2 ll twu 1/S 1 Formula Scale Factor lg lg2 twu Dc S 2 Dc2 Table 1.2: Effect of scaling on interconnect parameters Schematic of a pair of interconnects is shown in Figure 1.1.1. Assuming that transistors scale by a dimensionless factor S (> 1), the effect of scaling on various interconnect parameters are summarized in Table 1.2. Interconnects can be categorized as local, semiglobal and global. Local interconnects are used to connect individual functional units. They use the bottom most layers of metals. Semiglobal interconnects connects higher level blocks or modules and typically use middle layers of metals. Both these interconnects scale with the feature size (i.e. become smaller). Global interconnects, as the name suggests, run across the entire chip. They use the top metal levels. Global interconnects do not scale with the feature size, but infact increase, since the die size is increasing. In other words, global interconnects scale with the die size (DC ). Most local interconnects are very small. So, their resistances can be ignored. Their RC delay is also remaining constant. So, scaling is not affecting performance of local interconnects. Semiglobal interconnects may have larger lengths so repeaters will be required, as technology scales, to speed them up. This is a minor problem. Global interconnects are the slowest ones, since their lengths are increasing. Even with optimal repeaters, the time required for a signal in the global interconnects to 1.2. Organization of this Book 3 cross the chip can be comparable to a few clock cycles. This is a major problem. Moreover, interconnects are also packed very closer to each other in modern VLSI. Thus a significant fraction of their total capacitance is to their neighbours. As a consequence, when one switches, it affects the other through capacitive coupling. This effect is called as crosstalk and is an important concern. On chip interconnect inductance is usually assumed as negligible, but is an important concern for closely packed buses. Wires also account for a significant portion of switching energy on the chip. Power signal integrity is also an important concern in modern VLSI due to scaling down of interconnects. The faster clock speeds and switching of large number of devices results in a type of noise, called inductive or L di/dt noise, which results in the chip seeing bouncing in ground and power supply voltages. This phenomenon can have a serious impact on the performance of a chip if not handled properly. 1.2 Organization of this Book Since scaling down of interconnects have lots of consequences, Chapter 2 begins with the modelling of interconnects in order to understand their behaviour. Using the model it then discusses the impacts interconnects can have in modern VLSI. Chapter 3 discusses an important consequence of scaling down interconnect spacing, crosstalk. It first examines the impacts of crosstalk and later on discusses various techniques which can be used to avoid or reduce crosstalk. Chapter 4 discusses ground and power supply bounce noise in great detail. It begins with a simplistic model to develop a fundamental understanding for the cause of this type of noise. It later on discusses the impacts of this type of noise and techniques that can be used to avoid or reduce it in detail. 1.2. Organization of this Book 4 Chapter 2 Interconnect: Modeling and Impacts 2.1 Interconnect Modeling In order to better understand the problems arising due to scaling of interconnects, as mentioned in section 1.1, and to find effective mitigation techniques for the same, it is necessary to model the interconnects. This is done in this section. 2.1.1 Geometry The interconnect geometry for a pair of interconnects is shown again in Figure 2.1.1 for convenience. The interconnects have w width, l length, t thickness, s interconnect-spacing and are h distance away from the conducting layer below. The material in this region acts as a diFigure 2.1.1: Interconnect Geometry electric. Aspect ratio is defined as t/w & pitch is defined as w + s. Earlier CMOS processes had only a couple of metal layers, but with time it became more economical to manufacture many metal layers. Aluminium (Al) interconnects used in older processes were replaced by copper (Cu) around the 180nm - 130 nm node to reduce resistance. Also, the SiO2 insulator between wires was replaced with materials having lower dielectric constants (low-k) to reduce capacitance. A 45nm process typically has 8-10 metal layers. To get an idea of the typical interconnect geometry parameter values, Table 2.1 shows the various interconnect parameter values for an Intel 45nm stack shown in Figure 2.1.2 [2]. 5 2.1. Interconnect Modeling 6 Figure 2.1.2: SEM image of wire cross-sections in Intel’s 45nm process [2] Layer M9 M8 M7 M6 M5 M4 M3 M2 M1 t (nm) 7000 720 504 324 252 216 144 144 144 w (nm) 17500 400 280 180 140 120 80 80 80 s (nm) 13000 410 280 180 140 120 80 80 80 pitch (nm) 30500 810 560 360 280 240 160 160 160 aspect ratio (nm) 0.4 1.8 1.8 1.8 1.8 1.8 1.8 1.8 1.8 Table 2.1: Intel 45nm metal stack 2.1.2 Model Interconnects cannot be modelled as ideal wires anymore, since their propagation delays have now become comparable to the switching speeds of transistors. They need to modelled using passive components like resistors, capacitors and/or inductors. They can be represented using various models ❼ Lumped RC model (Figure 2.1.3) ❼ Distributed RC model (Figure 2.1.4) ❼ Transmission line model (Figure 2.1.5) Of these, the most commonly used is the lumped RC model since it is mathematically simple and easy to use for rough calculation. The lumped RC and distributed RC models model the interconnect using resistors and capacitors. The transmission 2.1. Interconnect Modeling 7 Figure 2.1.3: Lumped RC Model Figure 2.1.4: Distributed RC Model Figure 2.1.5: Transmission Line Model line model uses all the 3 passive components - resistors, capacitors and inductors - to model the interconnects. I have used the lumped RC model throughout this report to represent interconnects. The following sections briefly explain how to compute the equivalent resistance, capacitance and inductance for an interconnect. 2.1.2.1 Resistance By definition, the resistance of the single interconnect for the pair of interconnects shown in Figure 2.1.1 is given as ρ l (2.1.1) tw where, ρ is the resistivity of the material and l, w and t are as shown in Figure 2.1.1. Since the resistivity ρ and thickness t are controlled by the process technology and thus not in the hands of the designer, the resistance of interconnects are more commonly represented in terms of sheet resistance R = ρ/t, as R= 2.1. Interconnect Modeling 8 l w Table 2.2 shows the resistivity of commonly used metals at 22 ❽. R = R Metal Silver (Ag) Copper (Cu) Gold (Au) Aluminium (Al) Tungsten (W) Molybdenum (Mo) Titanium (Ti) (2.1.2) Resistivity (µΩ.cm) 1.6 1.7 2.2 2.8 5.3 5.3 43.0 Table 2.2: Bulk resistivity of some pure metals (at 22 ➦C) 2.1.2.2 Capacitance An isolated interconnect over a substrate can be modeled as a conductor over ground. This capacitance has 2 major components ❼ A parallel plate capacitance formed between the bottom of the interconnect and the ground ❼ A fringing capacitance due to the fringing electric field lines from the edge of the interconnect to the substrate These have been shown graphically in Figure 2.1.6 The parallel plate capacitance is defined as Cpp = ǫox wl h (2.1.3) It is not easy to compute the fringing field capacitance and various approximations for the same have been proposed. A commonly used formula is Cf ringe = 2πǫox l log(h/l) (2.1.4) Therefore, we can write total capacitance for an interconnect as Cinterconnect = Cpp + Cf ringe (2.1.5) 2.1. Interconnect Modeling 9 Figure 2.1.6: Effect of fringing fields on capacitance All this so far has been discussed for an isolated interconnect over substrate. Capacitance interactions with neighbouring interconnects become important in modern CMOS processes since the interconnects and metal lines are very closely spaced. A simplified schematic of a commonly occurring scenario in modern CMOS processes is shown in Figure 2.1.7 The total capacitance of the interconnect of interest (the one which is shaded) is sum of capacitances to the top layer, the bottom layer and adjacent interconnects. Ctotal = Cbot + Ctop + 2Cadj (2.1.6) Theoretically, interconnects will have capacitances to every other conductor in the chip, but they would be extremely small and can be ignored, since most electric field lines terminate at the immediate neighbours. (The effect of switching of neighbours on Ctotal is discussed later in Chapter 3) Depending upon feature size, either capacitance may dominate as shown in Figure 2.1.8. Interwire capacitances start to dominate in multilayer interconnect structures with decreasing feature sizes. 2.1.2.3 Inductance The inductance of a conductor (shown in Figure 1.1.1) can be defined approximately as 8h w µ0 (2.1.7) L = l ln( + ) 2π w 4h 2.2. Interconnect Impact 10 Figure 2.1.7: Multilayer capacitance model of an interconnect. The shaded interconnect in the centre is the interconnect of concern for which the total capacitance is being calculated. Inductance is difficult to extract and model since inductances are highly dependent on the 3-dimensional interconnect geometry. We usually try to model the circuit in terms of resistances and capacitances such that the effect of inductances is negligible. Inductances however cannot be ignored in high speed designs for power and clock buses. 2.2 Interconnect Impact Since interconnects are no more ideal equipotential wires in modern VLSI circuits, their presence affects the circuits in various aspects. These have been discussed briefly below using the lumped RC model defined in Section 2.1.2. 2.2.1 Delay Interconnects increase the circuit delay since it adds its own resistance as well as loads the gate with its own capacitance (as the lumped RC model of Figure 2.1.3 indicates) thus contributing to the overall RC delay. This delay is more pronounced for long interconnects which have significant resistance and capacitance which scale with interconnect length. Inorder to reduce delay, wider and thicker wires can be used (for the upper level interconnects, which are most affected by interconnect delay) made up of low resistivity materials like copper to help reduce the interconnect resistance. Using lower dielectric constant materials will help reduce interconnect capacitance and thus the delay. These are however not always effective, since there is a limit to the extent at which these parameters can be tweaked. Appropriate repeaters (buffers made using 2.2. Interconnect Impact 11 Figure 2.1.8: Effect of fringing fields on capacitance inverters) can be used to break a long interconnect into smaller lengths of interconnect such that the overall delay can become a linear function of interconnect length (as against to a quadratic function of length which comes if a single long interconnect is used). 2.2.2 Crosstalk As shown in Figure 2.1.7, interconnects have capacitances to both ground and neighbouring interconnects. As the interconnect switches, it tries to bring its neighbours along with it due to the capacitive coupling (Cadj ) between the 2 interconnects. This phenomenon is known as crosstalk. If the neighbouring interconnect is also supposed to switch at the same time, then, due to crosstalk, the switching delay of both the interconnects may either increase or decrease. On the other hand, if the neighbour is not supposed to switch, then crosstalk will cause noise on the neighbour. This is shown graphically in Figure 2.2.1. The type of crosstalk mentioned above with reference to Figure 2.1.7 is called as 1st order crosstalk, since it concerns with the switching of only the immediate neighbouring interconnects. If we consider both the first and second immediate neighbouring interconnects, this is called as 2nd order crosstalk and so on. Figure 2.2.2 illustrates graphically 1st , 2nd and 3rd order crosstalk. Higher order crosstalk effect is usually very small and can be neglected in most cases. Crosstalk is discussed in detail in Chapter 3 along with mitigation techniques for 2.2. Interconnect Impact 12 Figure 2.2.1: Crosstalk noise on interconnect B (victim) due to interconnect A (aggressor) due to coupling capacitor C Figure 2.2.2: Graphical illustration of 1st , 2nd and 3rd order crosstalk 2.2. Interconnect Impact 13 the same. 2.2.3 Power Switching power consumption is dependent on the capacitance at the switching node. Since interconnects add to the gate capacitance, more power is required to switch the nodes at the same frequency (as compared to the case if the wire is treated as ideal). 2.2. Interconnect Impact 14 Chapter 3 Crosstalk: Effects and Mitigation Techniques Crosstalk was discussed briefly in Section 2.2.2 where a scenario involving multilayer capacitances to an interconnect was shown in Figure 2.1.7. It was also discussed that crosstalk may cause delay effects or noise effects or both. These effects are discussed in detail in the next section (Section 3.1). In this chapter, only 1st order Figure 3.0.1: Cross sectional view of 2 incrosstalk is considered. Cross sectional terconnects A and B with the relevant caview of a simplified scenario involving pacitances shown 2 neighbouring interconnects A and B coupled capacitively is shown in Figure 3.0.1. This figure will be used frequently for discussion in the next section. 3.1 3.1.1 Effects of Crosstalk Crosstalk delay The relative directions of switching of interconnects A and B affects the amount of charge that must be delivered to the coupling capacitor Cadj and thus the delay. The more the charge required, the more is the delay involved. The charge delivered is given by Qrequired = Cadj ∆V (3.1.1) where ∆V is the difference in voltages of A and B. We also define the Miller Coupling Factor (MCF) which is the ratio of effective adjacent capacitance to Cadj . Deviation of 15 3.1. Effects of Crosstalk 16 this factor from 1 helps in getting an idea of how much crosstalk affects the switching delay of interconnects. Three distinct cases are possible depending upon the relative directions of switching of interconnects A and B for which ∆V is different [9]. These have been discussed below. 3.1.1.1 A switches and B does not Figure 3.1.1: A typical scenario where interconnect A switches but B does not This is a typical case delay scenario as shown in Figure 3.1.1. In this case, ∆V = VDD and accordingly Equation 3.1.1 becomes Q = Cadj ∆V . Thus, the total capacitance seen by A is Cgnd + Cadj . The MCF in this case is 1. 3.1.1.2 A and B switch in same direction Figure 3.1.2: A best case scenario where both interconnects A and B switch in the same direction This is a best case delay scenario as shown in Figure 3.1.2. In this case, ∆V = 0 and accordingly Equation 3.1.1 becomes Q = 0. Thus, the total capacitance seen by A is simply Cgnd which is the least possible capacitance. The MCF in this case is 0 (minimum possible). 3.1.1.3 A and B switch in opposite directions This is a worst case delay scenario as shown in Figure 3.1.3. In this case, ∆V = 2VDD and accordingly Equation 3.1.1 becomes Q = 2Cadj VDD . Thus, the total capacitance 3.1. Effects of Crosstalk 17 Figure 3.1.3: A worst case scenario where interconnects A and B switch in the opposite direction Switching Direction of B Not switching (Typical case) Same as A (Best case) Opposite to A (Worst case) Qreq Cadj VDD 0 2Cadj VDD Ceff (A) Cgnd + Cadj Cgnd Cgnd + 2Cadj MCF 1 0 2 Relative Delay 1 ↓ ↑ Table 3.1: Effective capacitances for various switching cases seen by A is Cgnd + 2Cadj . The MCF in this case is 2 (maximum). The above discussion has been summarized in Table 3.1 and is shown graphically in Figure 3.1.4. Figure 3.1.4: Propagation delay for the 3 cases summarized in Table 3.1. Delay noise is the difference between delays for the best and worst case scenarios. 3.1.2 Crosstalk noise In the typical case we discussed in Section 3.1.1.1, when interconnect A switches and interconnect B remains constant, noise is induced on B as it tries to switch partially. This si known as interconnect noise. In this case, A is called as the aggressor or perpetrator and B is called as the victim of crosstalk noise. 2 distinct cases are possible in this case - 3.1. Effects of Crosstalk 3.1.2.1 18 The victim is floating The victim can be modelled as shown in Figure 3.1.5. The victim noise in this case can be obtained easily since it is just a capacitive voltage divider circuit. The victim noise is given by Cadj ∆Vaggressor Cgnd−v + Cadj (3.1.2) where ∆Vaggressor is the change in voltage of the aggressor due to switching. It is usually equal to VDD . Note that since the victim is floating, the noise will remain indefinitely. ∆Vvictim = 3.1.2.2 Figure 3.1.5: Coupling to floating victim The victim is being driven The victim in this case can be modeled as shown in Figure 3.1.6 where the resistors represent the drivers. Since the victim is being driven, its driver will provide the necessary current to oppose and reduce the victim noise. So, the noise won’t remain indefinitely in this case. It would rise to a peak and then go down to 0 again. The peak noise in this case can Figure 3.1.6: Coupling to driven victim be given as [4] 1 Cadj ∆Vaggressor (3.1.3) ∆Vvictim = Cgnd−v + Cadj 1 + k where k= τaggressor Raggressor (Cgnd−a + Cadj ) = τvictim Rvictim (Cgnd−v + Cadj ) (3.1.4) From Equation 3.1.3 it is clear that the faster the victim driver (ie smaller Rvictim ), smaller is the crosstalk noise induced on the victim, which is expected. Equation 3.1.3 reduces to Equation 3.1.2 when the victim is floating ie when Rvictim → ∞, which is also as expected. Figure 3.1.7 illustrates all the above discussed cases graphically (for Cadj = Cgnd−a = Cgnd−v ). Equations 3.1.2 and 3.1.3 indicate that the effect of crosstalk depends on the ratio of Cadj to Ctotal . Since the load capacitance is included in the total capacitance, 3.2. Crosstalk Mitigation 19 Figure 3.1.7: Waveforms of coupling noise crosstalk is unimportant for short interconnects (since they have smaller capacitances) with large loads. The load capacitance dominates in this case. Accordingly, crosstalk is very important for long interconnects. 3.2 Crosstalk Mitigation In modern VLSI, as interconnects scale down, coupling capacitors can account for a significant percent of the total capacitance and can thus (as per Equations 3.1.2 and 3.1.3) create large amounts of noise and data dependent delay variations. It is therefore necessary to adopt crosstalk mitigation techniques. Some of the commonly used effective techniques have been discussed below. 3.2.1 Intelligent Engineering These are the simplest techniques that require no additional circuitry. They simply exploit the existing scenario cleverly to avoid or reduce crosstalk. 3.2.1.1 Increase spacing to adjacent interconnects This is the easiest approach to fix a minor crosstalk. Increasing the distance between interconnects decreases coupling capacitance (Equation 2.1.4) and thus decreases crosstalk noise (Equation 3.1.3). This method, however, cannot be used for fixing 3.2. Crosstalk Mitigation 20 Figure 3.2.2: Different passive shielding topologies with shields on one or both sides severe crosstalk issues since in that case the spacing may have to be increased significantly, which is not acceptable in modern VLSI as area is an important concern. In such cases we need to resort to other methods. 3.2.1.2 Different switching times for neighbouring interconnects If we have prior knowledge about the timings of signals in 2 interconnects, they may be arranged in such a manner so as to guarantee different switching times for neighbouring interconnects. As an example, consider 2 buses A and B such that bus A switches only on rising edge of clock and bus B switches on falling edges only. If we interleave the individual bit lines of the 2 buses, as shown in Figure 3.2.1, we can guarantee that for every interconnect, while it switches, its 2 immediate neighbours remain constant. Thus, Figure 3.2.1: Interleaving bit lines this can avoid the delay impact of crosstalk (Sec- of buses A and B to avoid tion 3.1.1). However, we must ensure that the crosstalk delay effects crosstalk noise is within the acceptable noise margins. 3.2.2 Shielding interconnects Another scheme of techniques to avoid crosstalk is to shield interconnects on one or both sides so as to avoid direct coupling with the immediate neighbouring interconnects and thus prevent crosstalk noise. 3.2.2.1 Passive Shields In this technique, as shown in Figure 3.2.2 [9], power and ground supplies are used as shields on one or both sides. Critical signals like clock can be shielded even on top 3.2. Crosstalk Mitigation 21 Figure 3.2.3: Active shielding topology. and bottom so that the switching activity of the neighbours do not affect the delay of the clock wire and introduce clock jitter. 3.2.2.2 Active Shields In this technique, as shown in Figure 3.2.3 [6], every interconnect is divided into 3 interconnects, which are in parallel. The central interconnect is made wide. The two interconnects on the either side of the central one are floating and have smaller widths. They themselves act as shields for the central interconnect. 3.2.3 Crosstalk cancellation techniques These techniques try to create equal amounts of positive and negative impact of crosstalk on the victim. These impacts cancel each other and effectively no net crosstalk impact is produced on the victim. These techniques may require additional circuitry and may thus result in additional power dissipation and area consumption. 3.2.3.1 Staggered repeaters Figure 3.2.4 shows two interconnects with staggered repeaters (inverters) [3]. Due to the staggered arrangement of inverters, each segment of the victim interconnect sees one half of a rising aggressor segment and one half of a falling aggressor segment, as indicated in Figure 3.2.4. Both these aggressor half-segments almost cancel each other’s noise on the victim. This is a very effective technique for crosstalk noise cancellation. Note that this technique does not require extra circuitry, since inverters (buffers) are usually used to break up a long interconnect into smaller segments for reducing the interconnect delay. This technique just proposes a clever arrangement of these inverters. 3.2. Crosstalk Mitigation 22 Figure 3.2.4: Staggered repeaters crosstalk control technique Figure 3.2.5: Charge compensation crosstalk control technique Figure 3.2.6: Twisted Differential Signalling crosstalk control technique. a and a are aggressor and its compliment. v and v are victim and its compliment (which is also a victim.) 3.2. Crosstalk Mitigation 3.2.3.2 23 Charge compensation This technique is shown schematically in Figure 3.2.5 [3]. An extra circuitry consisting of a transistor and an inverter is needed between the aggressor and the victim interconnects for this technique. The transistor’s source, drain and body terminals are tied together and it thus behaves as a capacitor (with a capacitance of Cgs + Cgd ). When the aggressor tries to rise, it tries to pull the victim upwards along with it. At the same time the extra added inverter tries to fall and tries to pull the victim downwards. By appropriately choosing the compensation transistor, most of the noise induced on the victim can be cancelled. This is also a very effective technique with the only disadvantage being the extra area and power required for the circuitry. 3.2.3.3 Twisted differential signalling This technique is illustrated in Figure 3.2.6 [3]. In this technique the signal interconnects are routed in differential (complimentary) pairs and in a twisted fashion. As a result, the victim and its compliment each are affected equally by the aggressor and its compliment, resulting in no net noise on either of them. This technique is used circuits such as memory, where signals naturally occur in differential (complimentary) fashion. If used otherwise, it would prove to be costly in terms of the wiring resources. 3.2. Crosstalk Mitigation 24 Chapter 4 Supply Bounce: Causes, Effects and Mitigation Techniques Figure 4.0.1 shows a cutaway view of a dual-in line package showing the chip-topackage connections. The chip (die) is connected to the package through bond wires bonded to metal pads on the chip pad frame. These bond wires connect to a metal lead frame in the package which provides I/O connections to the chip. One of the important things we expect from an ideal chip package is to provide connections between chip and board with little delay and distortion. However, practically, this does not happen. The bond wires and lead frame introduce Figure 4.0.1: A cutaway view of dual-inparasitics between the chip (die) and the line package showing the chip-to-package package. Figure 4.0.2 shows schemati- connections cally the parasitics introduced between the chip and package I/O pins [9]. Notably, the VDD and GN D wires have inductance from both the bond wires and lead frame as shown in the figure. The non-zero resistance also is important, especially for chips drawing large amounts of currents. These parasitics cause signal integrity problems, which is a major concern in modern VLSI. When large currents with large slew rates flow through the bonding wires and package pins, the ground and VDD supply voltages seen by the chip are not the actual ground and VDD supply voltages. They experience fluctuations due to the presence of the chip-to-package interface parasitics shown in Figure 4.0.2. This bouncing is even more enhanced if multiple drivers switch simultaneously. These fluctuations 25 4.1. L di/dt Noise 26 Figure 4.0.2: Chip package interface parasitics in ground and VDD supply voltages are called as Ground Bounce and VDD Bounce respectively. These fluctuations are caused due to two reasons, the IR drop primarily due to lead frame and bond wire resistances and the L di/dt noise due to lead frame, bond wire, etc. inductances. Of these, the L di/di noise is a major component of circuit noise in modern VLSI and its causes have been discussed in detail in the next section. 4.1 L di/dt Noise This is also called as Simultaneous Switching Noise since this is most pronounced in the case of a large number of I/O drivers switching simultaneously. 4.1.1 Understanding L di/dt Noise Figure 4.1.1 shows a simple model to understand ground bounce [8]. The inverter represents a driver in the lead frame of the package driving a load CL . The inductances L1, L2, L3 represent the intrinsic inductances in the ground lead, power lead and output lead of the package receptively. R1 represents the output impedance of the driver. Note that ideally we would want inductances L1, L2, L3 to be absent. The driver is now given a rising step input, so that its output falls gradually. Figures 4.1.2, 4.1.3 and 4.1.4 show output waveforms for V , I and VGB against time (symbols are marked in Figure 4.1.1). The output voltage V slew rate is dependent on CL , L3, L1 and the pull down transistor. For simplicity, the output voltage V is shown to fall at a constant rate in Figure 4.1.2. As the output voltage V falls, 4.1. L di/dt Noise 27 Figure 4.1.1: Model to understand ground bounce Figure 4.1.2: V vs. time for a step input to model in Figure 4.1.1 Figure 4.1.3: I vs. time t for a step input to model in Figure 4.1.1 Figure 4.1.4: VGB vs. time t for a step input to model in Figure 4.1.1 4.1. L di/dt Noise 28 the capacitor starts to discharge and generates a current I as shown in Figure 4.1.3. This discharging current goes to the ground through the inductances L3 and L1. The voltage waveform across the ground lead inductance VGB is shown in Figure 4.1.4. Figure 4.1.4 clearly shows that the internal device ground is not same as the external system ground at all times or in other words, the internal device ground bounces. A very similar analysis can be done to illustrate VDD Bounce where, instead of a rising step input, a falling step input would be given. In this case the power lead inductance L2 would come into picture instead of the ground lead inductance L1. The output waveforms of V and VGB for ground and VDD Figure 4.1.5: Ground and VDD bounce outbounce have been summarized in Figure put waveforms in Figure 4.1.1 4.1.5. VDD bounce is simply the inverse of ground bounce. Henceforth, all the analysis in this chapter would be done for ground bounce which would be equally applicable for VDD bounce. All the analysis done in this section were for ideal conditions. Practically, there would be 2nd and higher order effects as well due to which we would not get such perfectly linear output waveforms. Infact, ground bounce is generated not just on high to low transitions of output voltage V , but also on low to high transitions due to the large gate capacitance of the output transistors on the die, which was neglected here. Nevertheless, the purpose of this section was just to illustrate the fundamental origin of ground and VDD bounce. 4.1.2 Causes for L di/dt Noise There can be several reasons for ground bounce, some of which have been explained below. 4.1.2.1 Number of switching simultaneously outputs (SSO) The number of simultaneously switching outputs significantly affects the ground bounce amplitude [8]. As an example, consider a commonly occurring scenario involving a 16 bit bus as shown in Figure 4.1.6. Suppose that all the 16 drivers switch at the same time. Let the effective parasitic inductance between the chip and the system ground be 10nH, which is reasonable, and the rate of change of current dI/dt be 5mA/nsec which is also reasonable. If we calculate the peak value of ground 4.1. L di/dt Noise 29 Figure 4.1.6: Schematic of 16 output buffers switching simultaneously Figure 4.1.7: Ground bounce waveforms for varying number of outputs switching simultaneously bounce roughly using dI (4.1.1) dt we get a peak bounce of 0.8V , which is very high for modern VLSI systems where the power supply voltage itself is 1.2V or less. Figure 4.1.7 shows the output waveforms of ground bounce (ie of VGB defined in Figure 4.1.1) for varying number of outputs switching simultaneously. It is clear, from Equation 4.1.1 given above and waveforms in Figure 4.1.7, that as the number of outputs switching simultaneously increases, the amplitude and duration of ground bounce increases. VGB = N L 4.1.2.2 Output Load The amplitude of ground bounce is significantly affected by the type and value of output load [8]. Figure 4.1.8 shows ground bounce amplitude for varying output load capacitances on all the outputs. It is seen that initially the amplitude peaks, but then falls down as the capacitance is increased. This fall down is caused by the filtering action of large capacitors. Figure 4.1.9 shows ground bounce amplitude for varying the output load capacitances only on the active outputs, keeping the load on the quiet outputs as constant. It can be seen that as the capacitive load increases the ground bounce increases. 4.1. L di/dt Noise 30 Figure 4.1.8: Ground bounce amplitude vs. capacitive loading (on all outputs) Figure 4.1.9: Ground bounce amplitude vs. capacitive loading (only on active outputs) Figure 4.1.10: Ground bounce amplitude vs. capacitive loading (only on quiet outputs) 4.1. L di/dt Noise 31 Figure 4.1.10 shows ground bounce amplitude for varying the output load capacitances only on the quiet outputs, keeping the load on the active outputs as constant. It can be seen that as the capacitive load increases the ground bounce decreases. This is due to the filtering effect of the load capacitances. All these figures also indicate that as the value of VDD increases, the ground bounce increases. This is explained in Section 4.1.2.4. 4.1.2.3 Location of output pins Ground bounce amplitude also depends on the location of the output pins with respect to the device ground [8]. The closer is the pin, the smaller is the ground bounce noise. Figures 4.1.11 and 4.1.12 depict the waveforms of ground bounce for the pin farthest and closest from the device ground respectively. As the figure indicates, by choosing outputs closer to the device ground, the ground bounce noise may be reduced by nearly 50%. Figure 4.1.11: Ground bounce waveform for pin farthest to the device ground (worst case output pin) 4.1.2.4 Figure 4.1.12: Ground bounce waveform for pin closest to the device ground (best case output pin) Power supply voltage (VDD ) The amplitude of ground bounce is also affected by the power supply voltage VDD . [8] Reducing the value of VDD reduces both the output voltage swing and the current delivering ability of the driver. Both the factors contribute positively towards the 4.2. Effects of L di/dt noise 32 Figure 4.1.13: Ground bounce amplitude vs. VDD for varying load capacitances reduction of ground bounce noise. Thus reducing the power supply voltage VDD reduces ground bounce noise. This is illustrated in Figure 4.1.13 for different load capacitances. This phenomenon was also observed previously in Section 4.1.2.2 (Figures 4.1.8, 4.1.10 and 4.1.9). 4.2 Effects of L di/dt noise Some of the important consequences of L di/dt noise have been discussed briefly in this section. 4.2.1 Noise Margin Reduction This is an important consequence of L di/dt noise [6]. The ground and power supply seen by the chip may be either more and less than the expected values. Since for noise margin calculation we consider the worst case scenarios (which in this case is a smaller VDD supply and larger ground supply) the noise margin reduces. The definition of noise margins can be modified to take into account ground and power supply bounce as follows N ML = (VOH − VLDD ) − VIH (4.2.1) 4.2. Effects of L di/dt noise 33 Figure 4.2.1: Noise margin for an inverter with and without ground and power bounce issue N MH = VIL − (VOL + VLSS ) (4.2.2) This has been shown graphically in Figure 4.2.1. An immediate consequence of reduction in noise margin can be alteration in states of digital circuits. 4.2.2 Propagation Delay Degradation When outputs switch, ground bounce occurs and this causes an increase in the propagation delay [8]. This can be explained as follows, with reference to Figure 4.1.1. When the output switches, voltage appears across L1. Any voltage appearing across L1 causes a reduction in the voltage across the pull down transistor which in turn causes the current flowing through it to reduce. Thus the discharging of the capacitor slows down and hence the propagation delay of the output waveform increases. This effect is even more pronounced when more number of outputs switch at the same time. This has been illustrated in Figure 4.2.2. 4.3. Mitigation Techniques Figure 4.2.2: Degradation of propagation delay due to supply bounce 4.2.3 34 Figure 4.2.3: Layout and schematic of buffers with multiple fingers Other issues There are various other consequences of supply bounce like ground noise undershoots may cause unexpected leakage currents due to turning on (forward biasing) of the source-substrate junctions, change in DC operating points of analog circuits, etc. 4.3 4.3.1 Mitigation Techniques Intelligent Engineering Some of the commonly used strategies for reducing ground bounce have been discussed briefly below. 4.3.1.1 Staggered switching of SSO buffers In Section 4.1.2.1 it was seen that larger the number of SSO, more is the ground and power supply bounce. To avoid this, we can use buffers with multi finger layout as shown in Figure 4.2.3 [6]. The smaller individual buffers in the multi finger layout would switch at slightly different times and would thus reduce the amplitude of the overall ground bounce. The output waveforms of SSO with and without staggering are shown in Figure 4.3.1. The drawback of this technique is the reduced operating speed. 4.3. Mitigation Techniques 4.3.1.2 35 Increase number of supply pins When the number of supply pins are increased, as shown in Figure 4.3.2, the parasitic inductances become parallel to each other and the effective inductance decreases [6]. Thus the supply bounce noise also decreases. This has been illustrated graphically in Figure 4.3.3 Figure 4.3.1: Ground bounce with and without staggering of SSO Figure 4.3.2: Schematic of chip package interface with multiple ground supply pins 4.3.1.3 Figure 4.3.3: Simultaneous switching noise (SSN) vs. number of drivers for varying number of supply pins Output pin positioning It was seen in Section 4.1.2.3 that the location of pin positions with respect to the power supply pins also affect the amplitude of the supply bounce. Hence, high drive outputs & SSO must be placed as close as possible to the ground supply pins. Critical signals like clock must be kept as far as possible from high drive outputs & SSO and 4.3. Mitigation Techniques 36 near ground supply pin to have as small ground bounce noise on these critical signals as possible. 4.3.1.4 Separate supply pins for Analog and Digital circuits It is important to keep separate power supply pins and pads for analog and digital components of a chip. Since switching activity occurs mostly in digital circuits, keeping separate pins and pads would help avoid coupling of supply bounce noise generated due to digital components with the analog components of the chip. 4.3.1.5 Slew rate control If the chip is not meant for high speed operations, the slew rate should be controlled. Lower slew rate would mean smaller di/dt and thus smaller power supply bounce. 4.3.2 Guard Rings 4.3.2.1 Passive guard rings Figure 4.3.4: Passive guard ring configuration Guard rings help in reducing power supply bounce by efficiently absorbing noise signals [6]. Passive guard rings can be placed close to noise sources in a configuration shown in Figure 4.3.4 to reduce power supply bounce noise efficiently. 4.3.2.2 Active guard rings Active guards consist of an amplifier, an active component. A typical active guard ring configuration is shown in Figure 4.3.5. Active guard rings help reduce power supply bounce noise more efficiently as compared to passive guard rings, but at the cost of the extra power and area. 4.3.3 On chip decoupling or bypass capacitors (decap) 4.3. Mitigation Techniques 37 The use of a decoupling capacitor on chip is an efficient technique to reduce supply bounce [6]. A typical scenario schematic is shown in Figure 4.3.6. Decoupling capacitors are charged during steady supply voltages. Whenever there is a local spike in current, the decap provides the instantaneous charge requirements of the chip, reducing the di/dt drawn from the chip and thus the supply bounce. Decoupling capacitors are usually distributed Figure 4.3.5: Active guard ring configuraacross the chip to meet the instantaneous tion current requirements across the chip. Figure 4.3.6: On chip decoupling capacitor Active decap technique can be used to get larger effective capacitances from same amount of capacitance, with he help of opamps (active elements). This technique is used when chip area is a critical concern and higher decoupling capacitors are required. 4.3. Mitigation Techniques 38 Chapter 5 Electromagnetic emission and interference in VLSI circuits Electromagnetic emissions can cause interference with electronic devices. Interference is said to occur when the energy received causes the receptor to behave in undesired manner. The interference can either occur directly (ex:common impedance coupling) or indirectly (ex: crosstalk, radiation coupling). It is important to design systems which are electromagnetically compatible with the environment i.e. it is important to design a system which satisfy the following criteria [7] ❼ It does not cause interference with other systems ❼ It is not susceptible to emissions from other systems ❼ It does not cause interference with itself In order to prevent interference, we can either [7] ❼ Suppress the emission at its source, or ❼ Try to make the coupling as inefficient as possible, or ❼ Make the receptor less susceptible to the emissions 5.1 EMC subcomponents Electromagnetic energy can be divided into four categories as shown below ❼ Radiated emissions (Figure 5.1.1) ❼ Radiated susceptibility (Figure 5.1.2) 39 5.2. Antenna 40 ❼ Conducted emissions (Figure 5.1.3) ❼ Conducted susceptibility (Figure 5.1.4) Figure 5.1.1: EMC: Radiated Emissions Figure 5.1.2: EMC: Radiated Susceptibility Figure 5.1.3: EMC: Conducted Emissions Figure 5.1.4: EMC: Conducted Susceptibility 5.2 Antenna The electric field strength, E, at a given r from an antenna, shown in Figure 5.2.1, can be calculated as [5] k×I ×A E= sinν (5.2.1) r 5.3. Emissions 41 where, I is the current through the antenna, A is the area of the antenna, nu is angle of the point with the plane of the antenna and k is the antenna constant. Antennas are indirectly created within and outside the chip due to current loops which result in electromagnetic emissions. 5.3 5.3.1 Emissions Conducted Emission This type of emission can be further categorized as given below [6] 5.3.1.1 Through ground and power supply lines Current spikes at ground and power supFigure 5.2.1: Schematic of radiated electric ply lines result in conducted emissions. field from an antenna Although presence of chip decoupling capacitances reduces current spikes at the ground and power supply lines, they cannot fully eliminate the supply bounce. 5.3.1.2 Through I/O As discussed in Chapter 4, simultaneous switching of I/O ports can result in significant current being drawn from the power supply. The chip package interface parasitics can act as antennas and propagate electromagnetic emissions. 5.3.1.3 Radiated emissions from current loops Low impedance loops ca sustain large currents which can generate significant electromagnetic fields. 5.3.2 Radiated Emission This type of emission can be further categorized as given below [6] 5.3.2.1 Common mode radiation This type of radiation is also called as monopole antenna radiation. This is caused due an unintential ground bounce which raises all ground connections of a chip above 5.4. Factors affecting EMC 42 the system ground. Roughly, it can be calculated as ECM = 4 × 10−7 × L × f × If d (5.3.1) where, f is frequency of operation, L is cable length, d is distance from cable and If is the common mode current in the cable at the frequency f . (All are in SI units.) 5.3.2.2 Differential mode radiation This occurs when an AC current passes in small loops. Roughly, it can be calculated as A × f 2 × If (5.3.2) EDM = 265 × 10−16 × d where A, f , If and d are as defined in Section 5.3.2.1. 5.4 Factors affecting EMC The three main factors affecting EMC are supply voltage, frequency and grounding [6]. 5.4.0.3 Supply voltage Higher supply voltages result in large voltage swings, which result in larger currents and thus more emissions. On the other hand, lower supply voltages can affect susceptibility of the device. 5.4.0.4 Frequency When the chip operates at higher frequencies, digital components in it create current spikes due to switching activity which result in emissions. 5.4.0.5 Grounding The type of grounding also affects EMC. Hence, the topologies shown in Figure 5.4.1 are adopted for based on the range of operating frequency. 5.5 5.5.0.6 EMC reduction techniques Device level noise reduction ❼ Various techniques discussed in Section 4.3 can be used to reduce device level noise and thus EMC. Some of those techniques are using multiple ground and 5.5. EMC reduction techniques 43 Figure 5.4.1: Grounding Topologies depending on frequency power supply pins, ❼ Fewer clocks should be off and should turned off when not required (effectively reducing the f of the chip). ❼ Minimizing areas of loops within the chip can also help in reduction of EMC as given by Equation 5.2.1. 5.5.0.7 Routing noise reduction ❼ Spacing between adjacent active traces should be kept greater than trace width. This would help in reducing crosstalk noise. ❼ Clock signal loop areas should be kept as small as possible. ❼ High speed lines and clock signals should be kept short and direct and loops should be avoided. ❼ There should be no floating outputs to avoid necessary switching and thus noise generation of the floating node. ❼ Critical signals like clock signals, chip enables should be kept separate from I/O ports. 5.5. EMC reduction techniques 44 References [1] Semiconductor Industry Association. Semiconductors, 2007. International Technology Roadmap for [2] P. Moon et al. Process and electrical results for the on-die interconnect stack for intel’s 45nm process generation. Intel Technology Journal, June 2008. [3] R. Ho, Ken Mai, and M. Horowitz. Managing wire scaling: a circuit perspective. pages 177–179, 2003. [4] R. Ho, K.W. Mai, and M.A. Horowitz. The future of wires. Proceedings of the IEEE, 89(4):490–504, 2001. [5] Texas Instruments. Electromagnetic emission from logic circuits. Technical report, 1999. [6] Sanjeev Kumar Jain. Isssues in deep-submicron cmos technology. Technical report, Indian Institute of Technology Delhi. [7] Clayton Paul. Introduction to Electromagnetic Compatibility. Wiley, 2006. [8] Fairchild Semiconductor. Understanding and minimizing ground bounce. Technical report. [9] Neil Weste and David Harris. CMOS VLSI Design: A Circuits and Systems Perspective. Addison-Wesley, 2011. 45