E = B; H = J + D D = ρ ; B = 0 D = ρ ; B = 0 Yehia Massoud ECE Department Rice University Grasping The Deep Sub-Micron Challenge in POWERFUL Integrated Circuits ECE Affiliates 10/8/2003 Background: Integrated Circuits D = ρ ; B = 0 IC Transistors G Wires (Interconnects) S D Transistor is a switch (0 or 1) Wires transmit logic between transistors For faster chips Shorter gate length ( t=l/v ) Shorter gate length (0.25 Micron) Faster wires ( less R) Use Copper Shorter gate length (< 100nm)
Technology scaling D = ρ ; B = 0 Faster and smaller devices Wire delay has already exceeded Gate delay SM DSM Delay Gate Delay Wire Delay Last Century 2003 2007 1000 800 500 300 100 70 Technology (nm) 50 Interconnect Modeling D = ρ ; B = 0 Model interconnects as short-circuit (Gate delay was dominant) Interconnect resistance became comparable to gate resistance Use of Copper, R is smaller Faster clock speeds and faster transition times Inductive impedance became comparable to interconnect resistance ωl R Need Accurate and fast modeling of RC and L
3D Interconnect world D = ρ ; B = 0 Courtesy Mike Chou 3D Interconnect Extraction (Field Solvers) D = ρ ; B = 0 Finite Element Method (FEM) Mature subject (many codes available) Solve for the field contained in a finite volume Volume must be big enough to assume no fields outside Produces many unknowns Very Slow Cannot be run across chip
3D Interconnect Extractors (Field Solvers) D = ρ ; B = 0 Boundary Element Method Solve for the charges on the boundaries Fewer # of unknowns than FEM but produces dense system Need fast methods to solve dense system Direct Gaussian Elimination, O(n^3), way too slow Iterative methods, O(n^2), still too slow. Multipole Accelerated iterative method O(n ) FFT O(n log(n)) Faster than FEM but still too slow to be run for the whole chip 3D Interconnect Extraction (Pattern Matching) D = ρ ; B = 0 Build a library of representative geometric shapes Use field solvers to extract these shapes Generate library Parameterized equations Train a Neural Net? Extraction Decompose the layout into a set of these shapes Evaluate interconnect values library generation takes long time May loose accuracy when breaking into shapes
Optimal Interconnect Extraction D = ρ ; B = 0 But how would you extract thousands of interconnects? What you Want to do Use Super fast 2D globally Use 3D libraries/ local 3D Solvers 2D model breakdown Detailed analysis Critical, Clock, Power nets Super fast extractors Use 2D extraction ( has been done of R and C) 2D inductance extraction ( Problematic? ) What is inductance? D = ρ ; B = 0 Inductance is a loop quantity ( Flux Linkage to a loop) Knowledge of return path is required ( not necessary for 3D solvers) Signal Line Return Path
Current Return Path Frequency Dependence D = ρ ; B = 0 Low Frequency High Frequency G S S G G G G G G G G G G G Inductance vs Frequency D = ρ ; B = 0 Inductance(H) x 10-10 7 6 5 4 2 Returns 4 Returns 6 Returns 10 Returns 16 Returns 32 Returns 3 2 10 6 10 8 10 10 10 12 Frequency(Hz)
Current 2D Inductance extraction D = ρ ; B = 0 l s G G G d d S G G G d=2(w+s) d l Gnd Signal Gnd s H Return is the nearest line Most common assumption on the current return path Formula-Based Inductance Extraction D = ρ ; B = 0 µ l d a0 log a 2 π w + H d + a l L loop = 1 2 1 a0 = 1+ 2N 5N 1 1 a1 = + 2 6 3N 2N 3 3 log( N + 1) 1 a2 = + + 2 4N 2N N N 1 1 + 2N i( 2log( i) log( i + 1) log(2n + 1 i) ) 2 i= 1 N i= 1 logi + log N l s G G G S G G G d d d=2(w+s)
Formula Vs Field Solvers D = ρ ; B = 0 6.5 x 10-10 6 Analytical Equation FASTHENRY Inductance(H) 5.5 5 4.5 4 w=1.6um d=4.8um w=1.2um d=3.6um 3.5 3 2.5 w=0.8um d=2.4um 2 4 6 8 10 12 14 Number of returns Predictions for future technology D = ρ ; B = 0 Underestimation(%) 250 200 150 100 50 w=s=1.0um(180nm) w=s=0.8um(130nm) w=s=0.5um(100nm) w=s=0.4um(70nm) w=s=0.3um(50nm) 50nm 70nm 100nm 130nm 180nm 0 0 5 10 15 20 25 30 35 Number of returns
Interconnect Challenge D = ρ ; B = 0 Need fast and accurate extractors But What do you do with all these RLC elements? (Spice definitely can t handle that! ) Need to represent these big circuits using reduced order models that are fast to build Spice can handle them Can be used in design optimization Deep Sub-Micron Challenges D = ρ ; B = 0 C Interconnect Extraction and Modeling Power reduction Signal integrity (noise, IR drop, ringing) System-on-chip integration Educational
Power Challenge D = ρ ; B = 0 Why should we reduce power? Everything that uses battery! It is becoming hard to cool chips down! ( heat sinks, fans,?) POWER= P(dynamic) + P (leakage) + P(dynamic) has been dominating 2 P(dynamic)= C L V dd So reduce Vdd! ( went down 5 Volts to sub 1V now) Problem [ Delay increases as Vdd decreases] Research Dual Vdd Use high Vdd on part where you need speed Use low Vdd otherwise Optimal regions for power and delay [ power-delay product] Power Challenge D = ρ ; B = 0 What about leakage power? Important whenever you use batteries ( Telecomm; Medical; mobile electronics; ) Inversely proportional to threshold voltage ( exp. relation) Problem [ Delay increases as Vt increases] To enable low Vdd logic, we have to use low Vt a 100mv drop in Vt increases leakage by an ORDER of magnitude So leakage power has been increasing with scaling Leakage power will be soon as big as dynamic Research Dual Vt (Multi Vt) Use low Vt on part where you need speed Use high Vt otherwise Dual Vdd- Dual Vt
Signal Integrity (Coupling Noise) D = ρ ; B = 0 Vdd D D D Q D W=3um, S=1.5um, Metal 6 IN1 IN2 IN3 IN4 IN5 W S quiet line D D D OP1 Gnd OUT1 OUT2 OUT3 OUT4 OUT5 IN6 IN7 IN8 Eight-Bit Buss Cross-Talk Example OUT6 OUT7 OUT8 Tech=0.18um,Vdd=1.8v, Wire length=3000um, Noise induced on the quiet line? D = ρ ; B = 0 Peak Noise including L = 1.17V (Vdd=1.8V) Peak Noise ignoring L = 0.54V 0.18um tech Do not model L, miss the glitch! 2.5 Voltage (V) 2.0 1.5 RLC RC 1.0 0.0 0.5 1.0 1.5 2.0 2.5 3.0 Time (ns)
Noise Reduction remedies D = ρ ; B = 0 Research on Noise Reduction Architectural Controlling bus width and spacing can be good enough if inductance is not an issue Shielding Excellent for RC problems Good if L is a factor but may not be enough because of the inductance long range effects Noise Reduction remedies D = ρ ; B = 0 Circuit techniques repeaters Repeaters work at a cost additional area, delay, and power Can not use aggressive repeaters (have inferior noise properties) differential signaling Less noise production can route over sensitive areas more noise immune Useful for global nets Fast but consumes static power?
System-on-Chip Integration (Model substrate coupling) D = ρ ; B = 0 Model substrate as resistive network D = ρ ; B = 0
Problem description D = ρ ; B = 0 Basic procedure of BEM approach D = ρ ; B = 0 Assume one contact has charge, it induces potential at the other contact Maxwell in integral form In matrix form Charge at Contact i Φ ( x, y, z ) = V ' Φ Maxwell eq. Potential at Contact j ρ ( x ', y ', z ' ) G ( x, y, z, x ', y ', z ' ) dv 1 = PQ P p ( i, j ) = V iv j G ( x, y, z, x ', y ', z ' ) dv i dv j Formulate Green function G Enforce boundary conditions (Floating or Grounded substrate) Construct potential coefficient matrix P solve the linear system and compute the conductance matrix '
Effects of distance between contacts D = ρ ; B = 0 Size of substrate: a=b=1600um Depth of substrate: d=50um Effects of distance between contacts D = ρ ; B = 0
Educational Challenge D = ρ ; B = 0 As feature size decreases, more EM effects need to understood and modeled Ground?, L, inductive coupling, electromagnetic interference, transmission line effects, Digital Vs analog? Digital signals are looking (behaving) like analog signals Integration of digital and analog parts on same chip Numerical Analysis Needed to solve these complicated integral/differential equations! Amazing Similarities D = ρ ; B = 0 Maxwell s Equations RULES! EM simulations VLSI Modeling MEMS Modeling [ sensors, Actuators, Micro-motors] Applications [ Automotive, Biomedical, Defense] RF and wireless Modeling Nanotechnology [ Talk by N. Halas on Nanoshells! ]
(EM) Charge Distribution D = ρ ; B = 0 Charge distribution (High density at corners) Within 1% from theory MEMS D = ρ ; B = 0 Different types of Inductors Ref: M. Allen from GTECH High Permeability materials are often used in MEMS applications to increase inductance or magnetic force while keeping small device size [Inductors, actuators, sensors, micro-motors, transformers]
MEMS D = ρ ; B = 0 Multi layer Micro-fabricated Inductor [ L. Daniel et.al.] Extracted inductance is within 3% from measured inductance RF Modeling D = ρ ; B = 0 Spiral inductor Our Method Theory