ICCAD99 1

presents

High-Speed High-Accuracy 3D VLSI Extraction

ICCAD99 2

Motivation

�

Interconnects Now Dominate VLSI

"

Timing

"

Signal Integrity

�

Existing Interconnect Extractors Fail

"

Poor Accuracy & High Speed

"

Medium Accuracy & Very Low Speed

�

Need High-Speed High-Accuracy Extractor

0

20

40

60

80

100Gate

Interconnect

0.13 µm0.18 µm0.25 µm0.35 µm0.5 µm

Process Technology

Del

ay [%

]

ICCAD99 3

Overview

�

Company

�

Technology

"

High-speed High-accuracy BEM 3D Field Solver

"

Efficient Net-by-Net RC Extraction

�

Products

"

Extraction of Critical Cells, Blocks, Nets

"

Distributed RC Models

"

IR drop

"

Substrate Coupling

�

Differentiation

"

2D / 2.5D / 3D Tools

�

Summary

ICCAD99 4

Coyote Systems

�

History

"

Privately held, located in San Francisco CA

"

Founded in July 1996 by 3D Field Solver & MEMS experts

"

DARPA contract to develop extremely fast & easy-to-use 3D CAD for MEMS

�

Products

"

First MEMS product AutoMEMS® announced July 1998

"

First VLSI product AutoIC® announced June 1999

"

New Tools

- AutoIC® v2, AutoIC-SMP™, AutoNet™, AutoSubstrate™, AutoIRdrop™, AutoMEMS® v2

�

Customers

"

End-users include Siemens, Monterey, Analog Devices, Motorola, Texas Instruments, Hewlett-Packard, Sandia, MIT, Berkeley, CMU

"

OEM licenses include Monterey Design

ICCAD99 5

Company Mission

�

VLSI

�

MEMS

Provide premium high-speed high-accuracy solutions for extraction, timing, and signal integrity problems for modern VLSI designs.

Provide premium high-speed high-accuracy analysis of large, realistic MEMS devices.

ICCAD99 6

Coyote Products

�

Technology

�

VLSI

�

MEMS

"

AutoBEM

Fastest BEM field solver

"

AutoBEM-SMP

Exploit fine-grained parallelismon SMP computers

"

AutoIC

Capacitance extraction of interconnects

"

AutoIC-SMP

SMP support

"

AutoNet

Distributed RC models from layout

" AutoSubstrate Resistive coupling in substrate

" AutoIRdrop Potential drop in powergrid

" AutoMEMS Coupled electro/mechanical simulator

" AutoMEMS-SMP SMP support

ICCAD99 7

VLSI Geometries� Interconnects dominate performance

� Deep submicron is 3D" Huge layouts

" Intricate topologies

" IBM " AMD

ICCAD99 8

VLSI Extraction� RC Interconnect Extraction Tools

" High-Speed

" High-Throughput

" High-Capacity

" High-Accuracy

� Possible Alternatives

� Identify Advantages/Disadvantages

" 3D Field Solvers " Statistical Methods " Pattern Matchers

- Coyote Systems - QuickCap™ - Simplex

- Raphael™ - Frequency

- FastCap - Mentor

- Ultima

ICCAD99 9

Pattern Matchers� Why they are used

" Mature technology, Very Fast

" Requires 3D field solver to develop pattern library

� 2D/2.5D tools do not capture 3D DSM effects" Fast (<100,000 nets/hour) after

lengthy precharacterization (100-200 hours)

" Increasing difficulties with modern processes

- Many Layers, small feature sizes

- Arbitrary angles & non-manhattan geometries

- PCB-style variable wire width

- Increasing number of patterns needed

" Optimistic “error cancellation” - no bounded errors

" Large errors (>20%)

ICCAD99 10

Accuracy Problem!" Some pattern matching vendors claim <<10% errors

" Simple counterexample...

" 3D simulation reveals 20% errors using superpositioning!

" Geometric non-linearities prevent accurate solutions

≠ + +

3D Field Solver Pattern Matching

ICCAD99 11

Statistical Methods� Why are they sometimes used

" Alternative to FEM, FDM

" Reasonably good for Self-Cap & Small Problems

� Why will they not solve the problem" No Distributed R’s & C’s

" Very long run times with increasing number of nets

" Inaccurate Coupling Capacitance - Cij

" Only supports Manhattan Geometries

" Expensive support for multiple dielectrics

" Solves only part of the problem - Cii

ICCAD99 12

Alternative Solvers� Proposed Technologies

" FEM

" FDM

" Traditional BEM

� Why will they not solve the problem" Not fast enough

" Limited to very small problems

" Huge memory and CPU time requirements

� Computationally not feasible

ICCAD99 13

VLSI Extraction� Coyote combines advantages of all approaches!

" Field solver accuracy

" Pattern matcher speed & capacity

0

2 0

4 0

6 0

8 0

100

Accuracy

SpeedCapacity

Coyote

Field solver

Pattern matcher

ICCAD99 14

VLSI Geometries� Interconnects dominate performance

� Deep submicron is 3D

� 3D problems require 3D solutions!

" IBM " AMD" High aspect ratios

" Extreme Density

" Round Corners

" Cylindrical Vias

" Variable width wires

" Nonideal Cross-Sections

" Coyote can verify the effect of all these features!

- Simplex, Frequency, Avanti, Mentor, QuickCap™ cannot

ICCAD99 15

3D Field Solvers� Arbitrary Geometries

" PCB-style, manhattan & non-manhattan

� Arbitrary Materials" Planar & conformal dielectrics

� Arbitrary Boundary Conditions" Neumann, Dirichlet, mixed, floating

� Only Coyote’s AutoIC® solves everything!" AutoIC is faster, more accurate, and easier to use!

ICCAD99 16

3D Field Solvers� If field solver, then pick best field solver

" Simplest, smallest model

" Most accurate (states & gradients)

" Fastest runtime

" Lowest memory requirements

" Coyote uses Multipole Accelerated BEM

" Computational Scaling N^2 ➔ NlogN

BEM FEM

Direct Iterative Multipole Direct Iterative

Nr. nodes N^2 N^2 N^2 N^3 N^3

Memory N^4 N^4 (NlogN)^2 N^6 N^4.5

CPUtime N^6 N^4 (NlogN)^2 N^9 N^4.5

ICCAD99 17

AutoIC Field Solver� VLSI Interconnect Extraction

" Accelerated Boundary Element Method

" Automatic meshed 3D model from 2D layout

" Adaptive mesh refinement

" Robust automation

" Scriptable batch-mode or interactive GUI-mode operation

" Superset of Sematech extraction API

" Cluster & symmetric multiprocessing (SMP) support

ICCAD99 18

AutoIC Speed� AutoIC™ 3D Field Solver

" 2,000,000 BEM elements/hour per cpu

" 1,500 nets/hour per cpu (user selected <2% error)

� AutoIC vs. Field Solvers" 10-50x faster than QuickCap™

" 100-500x faster than Raphael™

� AutoIC vs. Library-based "Pattern Matchers"" Similar speed to Frequency

— 5 — 5/26/99 Copyright © 1999 by Frequency Technology, Inc. All rights reserved.

� Accuracy� +2.93% to -4.93% with no outliers

� Performance� Full accuracy / GDS mode : 500 nets /hour:

� Full accuracy / LEF- DEF mode : 2500 nets/hour

� “-fast” option (+/- 15% accuracy) : 6500 nets/hour

ICCAD99 19

AutoIC� Best Performance

" Faster than FastCap, QuickCap™, Raphael™, Space, etc.

" As fast as pattern matchers!

" Solves many more problems

" Large savings replacing multiple licenses

" Large engineering time savings

ICCAD99 20

Customer Tests� Motorola SRAM Evaluation

" Self-cap Extraction

" Cross-cap Extraction

" Convergence

" Time & Memory

" Model Generation

" Ease-of-Use

� Motorola Compared" AutoIC™

" QuickCap™

" Raphael™

ICCAD99 21

Motorola SRAM Cell

�

Convergence of bit-line self-capacitance

"

Require better than ±5% accuracy

�

AutoIC

"

10-100x faster!

"

Best convergence!

"

Tight error bounds!

0.00E+00

5.00E-16

1.00E-15

1.50E-15

2.00E-15

2.50E-15

1 1 0 1 0 0 1000 10000

cputime [sec] (AutoIC on 500MHz PentiumIII, others on 333MHz UltraII)

Cap

acit

ance

[F

]

AutoIC-constant

AutoIC-linear

Raphael

QC 0.01u

QC 0.005u

105%

95%

ICCAD99 22

Siemens Block� Extract Critical Nets

" 500 nets

" 1x1mm die

" 0.35um process

" 3 metal layers

" Multiple dielectrics

� AutoIC results" <2% Errors

" Excellent scaling

" 8mm/min extraction

" All nets extracted in 20 min using 1 cpu = 1,500 nets/hour

" All nets extracted in 1 min using 20 cpus = 30k nets/hour

0

2 0

4 0

6 0

0 2000 4000 6000 8000

Critical Net Length [um]

Ext

ract

ion

Tim

e [s

ec]

7700um/min

ICCAD99 23

Design Flow� Industry Std Layout

" GDSII, annotated GDSII, CIF, Sematech API

� 3D Model Generator" Flexible process descriptor

" API-mode or Batch-mode or GUI-mode

� 3D RC Extraction" Lumped, distributed

" API-mode or Batch-mode or GUI-mode

� Industry Std Output" Spice, Sematech API

" 3D visualization

ICCAD99 24

Model Generation� Input

" Layout

" Process Description

- Arbitrary layers

- Interlayer dielectrics

- Conformal dielectrics

� Output" Meshed 3D Model

ICCAD99 25

Applications� Critical Cells

" Coefficient generation

" Cell optimization

� Critical Blocks" IP characterization

" Routing

� Critical Nets" Clock trees, control lines

" User-selected nets

� Substrate Coupling

� IR drop" Power grid effects delays

ICCAD99 26

Critical Cells� AutoIC simulates 10,000 cells in 2 hours

" 1 AutoIC license replaces 50-200 Raphael™ licenses

" Adaptive meshing on SRAM bit-line shown

ICCAD99 27

Cell Optimization� AutoIC identifies where interactions occur

� Modify layout to optimize performance

" Magnitude of electrostatic flux on SRAM bit-line shown

ICCAD99 28

Critical Nets� Handles largest global nets

" Automatic tunneling around aggressor net

" 3D visualization indicates crosstalk areas

" Designer can modify layout to eliminate problem

" 1 AutoIC license replaces 10-50 QuickCap™ licenses

ICCAD99 29

Critical Blocks� PCB-style layout (Cypress)

" Non-manhattan Geometries

" Varying Linewidths

" AutoIC 500x faster than QuickCap™!

" Aggressor net, 3D tunnel & 3D mesh shown

ICCAD99 30

Non-Ideal Fabrication� AutoIC simulates Process Variations

" Any geometry

" Any material

� AutoIC correctly solves floating filler metal" Note that filler adds 20% to self-cap!

" Non-zero potential on floating filler shown

ICCAD99 31

AutoIC-SMP Scaling� Symmetric multiprocessing support

" Excellent scaling with every added CPU

" Accelerates both large and small nets

SMP Scaling for Critical Net Extraction

1

2

3

4

1 2 3 4

Number CPUs (Sun Ultra Enterprise 450)

Rel

ativ

e S

calin

g

Large Monterey benchmark

global net (control90 !RESET)

small net (control90 !n6969)

ICCAD99 32

"The Layout is the Design"

� Timing Delay" Self-capacitances dominate

� Signal Integrity" “Noise” or “Crosstalk”

" Cross-capacitances dominate

� AutoIC generates lumped RC models" Extracts high-accuracy 3D self-caps and cross-caps from the

interconnect layout

" 3D visualization identifies areas for correction

� AutoNet generates distributed RC models" Important differences between lumped and distributed

simulations

ICCAD99 33

AutoNet� Go from Layout to Distributed Spice

� Variable model size" Lumped net =

1 segment

" Distributed net = multiple segments

� Segments" Spice star-model for each

segment

" Created from layout

" Cross-caps between segments

" Compact Spice model size

" No model reduction needed

A B

C

D

a b

T1

T2

V1

V2

V3

T3

T4T5

Segment1

A B

C D

a

b

V3

T1

T2

V2

V1 T3

T5

T4

Segment1Segment3

Segment2

Segment4

Segment5

ICCAD99 34

AutoNet� Automatic segmentation

" 5 segments on Net1

" 4 segments on Net2

� Distributed RC spice model" Each segment has self-capacitance

" Each segment has cross-capacitances to segments on other nets

Net2

Net1

a b c d

A

B

C D E

A B C D

E

ab

c

d

ICCAD99 35

AutoNet� Parallel bus lines

" Generated distributed model(8 segments per net)

" Generated lumped T-model(1 segment per net)

" Step Response

- Distributed and T model responses similar

" Coupling

- Distributed and lumped responses similar

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.2

0.4

0.6

0.8

1

time [ps]

volts

Crosstalk

Vin8 segmentsT-model

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.2

0.4

0.6

0.8

1

time [ps]

volts

Unit step response

Vin8 segmentsT-model

ICCAD99 36

AutoNet� Non-uniform bus lines

" Generated distributed model(4-5 segments per net)

" Generated lumped T-model(1 segment per net)

" Step Response

- Distributed and T model responses similar

" Coupling

- 2x more noise on distributed model

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.2

0.4

0.6

0.8

1

time [ps]

volts

Unit step responses

Vin4/5 segmentsT-model

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.2

0.4

0.6

0.8

1

time [ps]

volts

Crosstalk

VinVictim A, 4/5 segmentsVictim B, 4/5 segmentsVictim A, T-modelVictim B, T-model

Victim A

Victim B

Aggressor

ICCAD99 37

AutoNet� Parallel lines with varying cross-section

" Net1 switches

" Step Response

- Distributed and T models similar

- Commonly used self-cap model is inaccurate

" Coupling

- 30% larger noise with distributed model

0 10 20 30 40 50 60 70 80 90 1000

0.2

0.4

0.6

0.8

1

1.2

time [ps]

volts

[V]

Unit step response and coupling effects

Vinresponse, 4/5 segmentsresponse, T-modelresponse, self capacitancecoupling, 4/5 segmentscoupling, T-model

Net2

Net1

ICCAD99 38

AutoNet� Non-uniform parallel lines

" Nets switch in opposite directions

" Step Response

- Distributed and T models similar

" Coupling

- Distributed model shows non-monotonic behavior

0 10 20 30 40 50 60 70 80 90 1000

0.2

0.4

0.6

0.8

1

1.2

time [ps]

volts

[V]

High/low switch response

Vin1

Vin2

Vout1, 4/5 segments

Vout1, T-model

Vout2, 4/5 segments

Vout2, T-model

Net2

Net1

Non-monotonic

ICCAD99 39

AutoNet� Clock Tree Skew

" Hierarchical H-shaped layout

" 256 terminals

" RC model with 70 segments

" Non-constant clock skew due to cross capacitances

" 20% clock skew at 3 random clock terminals!

0 20 40 60 80 100 120 140 160 180 2000

0.2

0.4

0.6

0.8

1

time [ps]

volts

[V]

Clock skew

VinVout

1Vout

2Vout

3

sample segments

ICCAD99 40

AutoNet� Compare Floating Filler & Grounded Filler

" 3x3 crossing bus

" Net1 switches

� Significant filler problem" 25% difference in risetime

" 9x increase in noise level

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.2

0.4

0.6

0.8

1

time [ps]

volts

[V]

Unit step response, filler floating / grounded

VinVout

1, floating filler

Vout1, grounded filler

Vout2, floating filler

Vout2, grounded filler

Vout3, floating filler

Vout3, grounded filler

Net1Net2

Net3

ICCAD99 41

AutoIRdrop� AutoIC solves 3D field

outside interconnects

� AutoIRdrop solves 2D/3D field inside interconnects" Varying angles, width,

cross-section

" Vias

" Aluminum, copper

" Current sink/source

� 2D/3D Field Solver" Solves potential drop

- Resistance

" Solves current density through wire

- Enables electromigration analysis

1

2

3

4

ICCAD99 42

AutoIRdrop� "Naive" Resistance Network

" N-ports require N simulations, resulting in N2 resistances

� "Smart" Resistance Network" N-ports require N simulations,

resulting in N resistances

� Field Simulation" 2D resistance is very accurate,

much faster than 3D

" 2D suitable for resistance,3D suitable for capacitance

1 2

3 4

1

2 3

4

ICCAD99 43

AutoSubstrate� Solves 3D field

inside substrates" Contacts

" Multiple materials

� 3D Field Solver" Solves potential

distribution

" Solves electrostatic flux distribution

" Solves current density through contacts

12

34

56

7

89

10

ICCAD99 44

AutoSubstrate� Resistance Network

" Resistance calculated from calculated current densities

" Resistor for every port-port interaction

" Automatic spice model

12

34

56

78

910

12

34

56

78

910

0

2

4

6

8

10

12

x 105

PortPort

Ohm

ICCAD99 45

Shipping Now� Fastest, Most accurate, Easiest to use

" Replace multiple tools

" Large engineering time savings

� Licenses" End-user licenses

" OEM, Embedded licenses

� Available Now" Sun, Linux, HP platforms

" Uniprocessor & SMP configurations

ICCAD99 47

For More Information

� http://www.coyotesystems.com

� Contact" Dr. Per Ljung

2740 Van Ness Ave. #210San Francisco, CA 94109(415) 346 4223 x12pbljung@coyotesystems.com