Lecture 14: Wires
Outline Introduction Interconnect Modeling Wire Resistance Wire Capacitance Wire RC Delay Crosstalk Wire Engineering Repeaters 14: Wires
Introduction Chips are mostly made of wires called interconnect In stick diagram, wires set size Transistors are little things under the wires Many layers of wires Wires are as important as transistors Speed Power Noise Alternating layers run orthogonally 14: Wires
Wire Geometry Pitch = w + s Aspect ratio: AR = t/w Old processes had AR << 1 Modern processes have AR 2 Pack in many skinny wires 14: Wires
Layer Stack AMI 0.6 mm process has 3 metal layers M1 for within-cell routing M2 for vertical routing between cells M3 for horizontal routing between cells Modern processes use 6-10+ metal layers M1: thin, narrow (< 3l) High density cells Mid layers Thicker and wider, (density vs. speed) Top layers: thickest For VDD, GND, clk 14: Wires
Example Intel 90 nm Stack Intel 45 nm Stack 14: Wires [Thompson02] [Moon08] 14: Wires
Intel 45nm Stack Layer t (nm) w (nm) s (nm) Pitch (nm) M9 7000 17500 13000 30500 M8 720 400 410 810 M7 504 280 560 M6 324 180 360 M5 252 140 M4 216 120 240 M3 144 80 100 160 M2 M1 14: Wires
Intel 45nm Stack 14: Wires
Interconnect Modeling Current in a wire is analogous to current in a pipe Resistance: narrow size impedes flow Capacitance: trough under the leaky pipe must fill first Inductance: paddle wheel inertia opposes changes in flow rate Negligible for most wires 14: Wires
Impact of Interconnect Reduce reliability Affect performance Increase tp Increase energy dissipation Cause the introduction of extra noise sources Inductive effects usually ignored Resistive effects ignored if wire is short Interwire capacitance usually ignored if overlap is small Wire capacitance is dominant 14: Wires
Wire Models Capacitance-only All-inclusive model
Capacitance of Wire Interconnect
Capacitance: The Parallel Plate Model
Permittivity
Fringing Capacitance
Fringing Capacitance 14: Wires
Fringing Capacitance Some other formulas This empirical formula is accurate to 6% for AR < 3.3 14: Wires
Fringing versus Parallel Plate
Wire Capacitance Wire has capacitance per unit length To neighbors To layers above and below Ctotal = Ctop + Cbot + 2Cadj 14: Wires
Interwire Capacitance
Capacitance Trends Parallel plate equation: C = eoxA/d Wires are not parallel plates, but obey trends Increasing area (W, t) increases capacitance Increasing distance (s, h) decreases capacitance Dielectric constant eox = ke0 e0 = 8.85 x 10-14 F/cm k = 3.9 for SiO2 Processes are starting to use low-k dielectrics k 3 (or less) as dielectrics use air pockets 14: Wires
M2 Capacitance Data Typical dense wires have ~ 0.2 fF/mm Compare to 1-2 fF/mm for gate capacitance 14: Wires
Impact of Interwire Capacitance
Capacitance of Dense Wires An empirical equation is Also, floating capacitors occur, which Create noise Affect performance 14: Wires
Wiring Capacitances We typically use simple models for capacitance, given by 14: Wires
Wiring Capacitances (0.25 mm CMOS)
Diffusion & Polysilicon Diffusion capacitance is very high (1-2 fF/mm) Comparable to gate capacitance Diffusion also has high resistance Avoid using diffusion runners for wires! Polysilicon has lower C but high R Use for transistor gates Occasionally for very short wires between gates 14: Wires
Wire Resistance r = resistivity (W*m) R = sheet resistance (W/) is a dimensionless unit(!) Count number of squares R = R * (# of squares) 14: Wires
Sheet Resistance
Choice of Metals Until 180 nm generation, most wires were aluminum Contemporary processes normally use copper Cu atoms diffuse into silicon and damage FETs Must be surrounded by a diffusion barrier Metal Bulk resistivity (mW • cm) Silver (Ag) 1.6 Copper (Cu) 1.7 Gold (Au) 2.2 Aluminum (Al) 2.8 Tungsten (W) 5.3 Titanium (Ti) 43.0 14: Wires
Contacts Resistance Contacts and vias also have 2-20 W Use many contacts for lower R Many small contacts for current crowding around periphery 14: Wires
Copper Issues Copper wires diffusion barrier has high resistance Copper is also prone to dishing during polishing Effective resistance is higher 14: Wires
Example Compute the sheet resistance of a 0.22 mm thick Cu wire in a 65 nm process. Ignore dishing. Find the total resistance if the wire is 0.125 mm wide and 1 mm long. Ignore the barrier layer. 14: Wires
Skin Effect 14: Wires
Skin Effect Define a skin depth, d, where the current falls to 1/e of its nominal value. Here, m is the permeability of the surrounding dielectric and has a typical value of approximately 4p X 10-7 for all dielectrics. For Al at 1 GHz, d = 2.6mm. To see the effect, assume a rectangular wire. Assume that the current flows only in the skin as defined above 14: Wires
Skin Effect The cross section is given by H W 14: Wires
Skin Effect Using this cross sectional area, We can define a frequency fs where the skin depth is half the highest dimension of the conductor. It is not meaningful to increase the dimensions beyond that point for that frequency. 14: Wires
Skin Effect For Al in SiO2, at 1GHz fs, the largest dimension should be 5.2mm. Actual results show 30% increase in R due to skin effect for a 20mm wire and 2% for a 1mm wire. One other thing to note is that the actual frequency of the square wave should not be used. A sine wave whose rise and fall times equal to the rise and fall times of the square wave will give more accurate results. For 20% - 80% rise and fall, the equivalent frequency is given by 14: Wires
Skin Effect As another example, choose copper in SiO2 with 20ps edge rates. f=5.8GHz, d = 0.99mm Note that the resistivity of metals drops at very low temperatures. For example, an order of magnitude improvement at 77K (liquid nitrogen). 14: Wires
Inductance We will ignore inductance in this course Inductance causes voltage variations Inductance causes extra impedance. Remember where c: capacitance per unit length l: inductance per unit length e: permittivity of the surrounding dielectric m: permeability of the surrounding dielectric 14: Wires
Inductance Also, remember that Dielectric er Prop. Speed (cm/ns) Vacuum 1 30 SiO2 3.9 15 PC Board (epoxy glass) 5.0 13 Alumina (ceramic packages) 9.5 10 14: Wires
Inductance How do we use this information? From a previous table, c (aF/mm) l (pH/mm) W = 0.4mm 92 0.47 W = 1mm 110 0.39 W = 10mm 380 0.11 14: Wires
Inductance Using Equating the impedances, Z = wl For a 1mm wide wire, r = Z at 30GHz. Inductance is not an issue for now. Typically, lower level metals are microstrips whose inductances are given by 14: Wires
Inductance On-chip inductance is important for wires where the speed of light flight time is longer than either the rise times of the circuits or the RC delay of the wire. This can be expressed as 14: Wires
Inductance 14: Wires
Inductance 14: Wires
Lumped Element Models Wires are a distributed system Approximate with lumped element models 3-segment p-model is accurate to 3% in simulation L-model needs 100 segments for same accuracy! Use single segment p-model for Elmore delay 14: Wires
The Lumped Model
Wire RC Delay Estimate the delay of a 10x inverter driving a 2x inverter at the end of the 1 mm wire. Assume wire capacitance is 0.2 fF/mm and that a unit-sized inverter has R = 10 KW and C = 0.1 fF. tpd = (1000 W)(100 fF) + (1000 + 800 W)(100 + 0.6 fF) = 281 ps 14: Wires
Wire Energy Estimate the energy per unit length to send a bit of information (one rising and one falling transition) in a CMOS process. E = (0.2 pF/mm)(1.0 V)2 = 0.2 pJ/bit/mm = 0.2 mW/Gbps 14: Wires
Elmore Delay 14: Wires
Elmore Delay 14: Wires
The Elmore Delay - RC Chain
Wire Model Assume: Wire modeled by N equal-length segments For large values of N:
The Distributed RC Line For a distributed line, we have the diffusion equation This equation has a solution in the s domain Vout(t) cannot be solved in closed form. It can be approximated by 14: Wires
Step-response as a function of time and space
RC Wires – Lumped vs Distributed Value of Interest Lumped RC Distributed RC 0 -> 50% (tp) 0.69RC 0.38RC 0 -> 63% (t) RC 0.5RC 10% -> 90% (tr) 2.2RC 0.9RC 0 -> 10% 0.1RC 0 -> 90% 2.3RC 14: Wires
Distributed RC Lines 14: Wires
Distributed RC Lines 14: Wires
The Transmission Line 14: Wires
The Transmission Line The diffusion equations are First, ignore r. This is a lossless transmission line. 14: Wires
The Transmission Line A step input applied to a lossless transmission line propagates through the line with speed v. Z0 is the characteristic impedance and is independent of length. Z0 is between 100W and 500W for typical wires. 14: Wires
The Transmission Line Define a wave reflection coefficient as When a transmission line is being driven by an ideal source, the termination is important. For a termination of Z0, no reflection. For a short circuit termination, r = -1 For an open circuit termination, r = 1 See the site http://www.williamson-labs.com/xmission.htm 14: Wires
The Transmission Line Case 1: Large source resistance and infinite load resistance. Take RS = 5Z0 Cycle 1: When this wave reaches the destination, it is fully reflected -> 1.67V. It comes back to the source. The source has become the load. 14: Wires
The Transmission Line Case 1 continued The new source voltage becomes We restart the analysis with 1.1 V for cycle 2. It is obvious that the signal builds up. However, the rise time is not determined by any RC constant. It is in terms of number of reflection cycles given by length/velocity. See the site http://www.eecs.tufts.edu/~alanh/simulation.html 14: Wires
The Transmission Line Case 2: Small source resistance, infinite load resistance. Almost all the signal injected into the transmission line. Reflected from the load. Almost doubles by the time it comes back to the source. The signal is phase reversed at the source as 14: Wires
The Transmission Line Case 2: The signal exhibits severe ringing. It takes many cycles before it settles to its final value. Case 3: Matched source resistance. Half the signal is injected into the line Doubles at the termination. Final value is reached within length/velocity. Capacitive termination (our case) No overshoot, behavior asymptotic to t = Z0CL Interesting behavior observed only at source. 14: Wires
A Look into the Future Ideal Scaling Typically, S = 1.15, SC = 0.94 per year. Delay of global wires increases 50% per year. Parameter Relation Local Wire Constant Length Global Wire W, H, t 1/S L 1 1/SC C LW/t R L/WH S S2 S2/SC RC L2/Ht S2/S2C 14: Wires
A Look into the Future Constant R scaling eC is introduced to model the extra fringing capacitance effects. As long as eC < S, not too bad. Parameter Relation Local Wire Constant Length Global Wire W,t 1/S H 1 L 1/SC C eCLW/t eC/S eC eC/SC R LW/H S S/SC RC eCL2/Ht eCS eCS/S2C 14: Wires
Crosstalk A capacitor does not like to change its voltage instantaneously. A wire has high capacitance to its neighbor. When the neighbor switches from 1-> 0 or 0->1, the wire tends to switch too. Called capacitive coupling or crosstalk. Crosstalk effects Noise on nonswitching wires Increased delay on switching wires 14: Wires
Crosstalk Delay Assume layers above and below on average are quiet Second terminal of capacitor can be ignored Model as Cgnd = Ctop + Cbot Effective Cadj depends on behavior of neighbors Miller effect B DV Ceff(A) MCF Constant VDD Cgnd + Cadj 1 Switching with A Cgnd Switching opposite A 2VDD Cgnd + 2 Cadj 2 14: Wires
Crosstalk Noise Crosstalk causes noise on nonswitching wires If victim is floating: model as capacitive voltage divider 14: Wires
Driven Victims Usually victim is driven by a gate that fights noise Noise depends on relative resistances Victim driver is in linear region, agg. in saturation If sizes are same, Raggressor = 2-4 x Rvictim 14: Wires
Coupling Waveforms Simulated coupling for Cadj = Cvictim 14: Wires
Noise Implications So what if we have noise? If the noise is less than the noise margin, nothing happens Static CMOS logic will eventually settle to correct output even if disturbed by large noise spikes But glitches cause extra delay Also cause extra power from false transitions Dynamic logic never recovers from glitches Memories and other sensitive circuits also can produce the wrong answer 14: Wires
Wire Engineering Goal: achieve delay, area, power goals with acceptable noise Degrees of freedom: Width Spacing Layer Shielding 14: Wires
Repeaters R and C are proportional to l RC delay is proportional to l2 Unacceptably great for long wires Break long wires into N shorter segments Drive each one with an inverter or buffer 14: Wires
Repeater Design How many repeaters should we use? How large should each one be? Equivalent Circuit (Using the PI model) Wire length l/N Wire Capacitance Cw*l/N, Resistance Rw*l/N Inverter width W (nMOS = W, pMOS = 2W) Gate Capacitance C’*W, Resistance R/W 14: Wires
Repeater Results Write equation for Elmore Delay Differentiate with respect to W and N Set equal to 0, solve ~40 ps/mm in 65 nm process 14: Wires
Repeater Energy Energy / length ≈ 1.87CwVDD2 87% premium over unrepeated wires The extra power is consumed in the large repeaters If the repeaters are downsized for minimum EDP: Energy premium is only 30% Delay increases by 14% from min delay 14: Wires
Repeaters Let us detail the derivation. It is simpler if distributed analysis is used. Using Elmore formulation, If there are k identical inverters, 14: Wires
Repeaters Make some approximations to make problem simpler. Ignore Cint and collect terms Take the derivative to find the optimum 14: Wires
Repeaters Now, let us try to size the repeaters as well. Let us use h to scale them. Now, taking two partial derivatives, 14: Wires