Technische Universität München We create tools for chips Institute for Electronic Design Automation U. Schlichtmann ISPD 2015 - March 31, 2015 Beyond GORDIAN.

Technische Universität MünchenTechnische Universität München

We create tools for chips

Institute forElectronic Design Automation

U. Schlichtmann

ISPD 2015 - March 31, 2015

Beyond GORDIAN and KRAFTWERK:EDA Research at TUM

Technische Universität München

Overview

The early years

Plantage: Analog Placement

PROTON: ONoC Physical Design

2


3

Topics covered by Prof. Antreich

• Since 1975: CAD for analog circuits• Early 1980s: Physical Design• Mid-1980s: ATPG• Late-1980s: Digital Simulation (Parallelization)• 1990s: Logic and RTL synthesis (FPGAs)• 1990s: Formal Techniques


4

Topics covered by Prof. Antreich

• Since 1975: CAD for analog circuits• Early 1980s: Physical Design• Mid-1980s: ATPG• Late-1980s: Digital Simulation (Parallelization)• 1990s: Logic and RTL synthesis (FPGAs)• 1990s: Formal Techniques

Technische Universität MünchenInstitute forElectronic Design Automation

Can you simulate what you want to design?

5

Circuit simulation

Simulation:

Performances f

Parameters x

Mathematical View

Physical View

Abstraction from the physical level


Statistical Parameter Distribution

• Normal distribution of with– mean vector– covariance matrix

6


Operating Range, Performance Specification

• Operating range– Bounds for environmental

parameters

– Simulator input

• Performance specification– Bounds for performances – Simulator output

7


Realistic Worst-Case Analysis

• Linear performance model with gradient for each performance

• Parameter tolerance ellipsoid

8


Relate Yield and Tolerance Region: Worst-Case Distance W

W -1 0 1 2 3 4

Y´ 15.9%

50% 84.1% 97.7%

99.9% 99.99%

sigma design

9


Example – Yield (Worst-Case) Analysis

NominalvaluePerformance

82.9%Overall yield

1.1(„“)/86.4%

2.6 ≤ 3.5DC power [mW]6.3(„“)/99.9%

67 ≥ 32Slew rate [V/s]1.8(„“)/96.9%

68 ≥ 60Phase margin[º]

7.7(„“)/99.9%

67 ≥ 30Transit frequency [MHz]

2.5(„“)/99.4%

76 ≥ 65Gain[dB]

Worst-casedistance/yield

Specifi-cation

10


Worst-case distance

Yieldpartition

50.00%

97.70%

99.99%

Example – Yield (Worst-Case) Analysis

11


Example – Optimized Yield

0

1

2

3

4

5

6

7

8

Initial Optimized

Gain Transit frequency Phase marginSlew rate DC power

50.00%

97.70%

99.99%

Worst-case distance

Yieldpartition

12

Empowering Innovation

© Copyright by MunEDA GmbH - All rights reserved - www.muneda.com

MunEDA Corporate Overview

Vision & Mission: Empowering Innovation

EDA Software Vendor – Design Tool Suite WiCkeDTM for IP Migration, Analysis, Modeling & Sizing of Nanometer IC designs

Founded in 2001 - Headquarters in Munich Germany

Worldwide Sales & Support Offices in USA, Korea, China, Taiwan, Japan, UK, Ireland, Scandinavia, South America

Worldwide Customer Base of Semiconductor IDMs, Fabless Design Houses & Foundries

13

Empowering Innovation

© Copyright by MunEDA GmbH - All rights reserved - www.muneda.com

MunEDA Worldwide Customers & Foundry Partners (Selection)

14

and more…for WiCkeDTM customer references see http://www.muneda.com/Customers


The Early Years


Overview


15


16

Analog Design Flow

Specification

Topology

Sizing

Placement

Routing

Layout

Sizing rules

Placement rules

Routing rules

Process variationsParasitics etc.

manual

mostlyautomated

mostlymanual


17

Modeling: inter-set relations

Placement rules

close proximity

A B

C D

E

symmetry

A B C E

alignment

BA

C

E

minimum distance

A1 B2 A3 B4

B1 A2 B3 A4

commoncentroid

variant constraints

R1

R2

R3

R4

R5

R6

Modeling: linear (in-)equalities


18

State-of-the-Art

• Two classes of algorithms:– Absolute coordinates

• [Jepsen/Gellat, ICCD‘83]• [Kohn/Garrod/Rutenbar, IEEE Journal SC‘91]

– Topological relations• Sequence Pair [Murata/Fujiyoshi, TCAD‘96],

[Tam/Young/Chu, ICCAD‘06]• BSG [Nakatake/Fujiyoshi, ICCAD‘96]• B*-Tree [Chang/Chang, DAC‘00], [Balasa, TCAD‘04],

[Lin/Lin, DAC‘07], [Lin, DAC‘08]

• Optimization using Simulated Annealing


19

CM

DS

CM

DP

Hierarchically Bounded Enumeration

• Hierarchical structure of analog circuits

• Automatic structure recognition [Gräb, Zizala, Eckmüller, Antreich ICCAD‘01]

[Massier, Gräb, Schlichtmann DATE‘08]

M5 M6 M7

CM

M1 M2

DP

M3 M4

CMM1 M2

M5 M6 M7

M3 M4

M8

CDS

CORE

OPA

M8 C

CM: Current Mirror

DP: Differential Pair

DS: Differential Stage


20

Enumeration of placements of fundamental module sets

Hierarchically Bounded Enumeration

M5 M6 M7

CM

M1 M2

DP

M3 M4

CM

DS

CORE

OPA

M8 C

Fundamental

module sets

M5 M6 M7

M5

M6

M7M5

M6M7

…

M3 M4

M3

M4M3

M4

…

M8 C

M8

C

M8

C

…

M1a

M2a

M2b

M1b

…

M2

aM

1a

M1

bM

2b

M5 M6 M7

CM

M1 M2

DP

M3 M4

CM

DS

CORE

OPA

M8 C


21

Overview

M5 M6 M7

CM

M1 M2

DP

M3 M4

CM

DS

CORE

OPA

M8 C

M5 M6 M7

M5

M6

M7M5

M6M7

…

M3 M4

M3

M4M3

M4

…

M8 C

M8

C

M8

C

…

M1a

M2a

M2b

M1b

…

M2

aM

1a

M1

bM

2b

• Hierarchical approach• Enhanced shape functions• New algorithm for placement:

B*-Tree CG LP• New addition of shapes

by addition of B*-Trees

CM

DS

CM

DP M1 M2

M5 M6 M7

M3 M4

M8

C

Enhanced Shape Functions

Enhanced Shape Addition


22

Results

• OTA with folded cascode

#10

12 shapes in 44 sWSCM: Wide Swing Cascode current MirrorDP: Differential Pair

DP

WSCM #3optimum

area

#8

DP

WSCM


Industrial Example: wls_input_stage

23

(a)

(c)

(b) (a)

(b) (c)


The Early Years


Overview


24


The Architecture

25

TSV

TSV

TSV

TSV

M1

2

3

4

λ1

λ2

λ3

λ4

Array of off-chip CW lasers

Electronic layer

Photonic layer

Off-chip memory

Clusters of processors

M2

M3M

4H4

H3

H2

H11


The Photonic Layer

26

M1M1

Photonic layer

M2M2

M3M3

M4M4

H4

H3

H2

H1

M2M2 H3

• A path connects two Hubs or a Hub and a Memory Controller via Photonic Switching Elements (PSEs) and waveguides

Memory Controller PSE PSE Hub

λi λj


Example: Logic Scheme of Optical Layer

27

Example: 8x8 λ-Router [Scandurra NoCArc’08]

λ1

I1 O1

O2

Wavelength λ1 is redirectedWavelength λn crosses

λ4

I2


Logic Scheme vs. Physical Design

28

H1

H2

H3

H4

M1

M2 M4

M3

Logic Scheme Manually created layout

Creating layout manually is time consuming, error prone and suboptimal

PROTON: The first automatic place & route tool for 3D optical NoCs

[Ramini NOCS’12]


Placement and Routing Problem

• Minimize maximum insertion loss over all paths, e.g. minimize– Waveguide length– Number of crossings between

waveguides– Number of bends

• Constraints:– Place all PSEs and waveguides

inside chip area– No overlap

Netlist, chip area, positions of hubs and memory controllers

Valid and optimal layout

29


Placement and Routing Algorithm

RoutingRouting

• Approximate waveguide length of each path by using a quadratic net model

• Approximate number of crossings

• Solve non-linear optimization problem

• Route waveguides by allowing but penalizing crossings

• E.g. Maze Router

PlacementPlacement

30


Approximation of Crossings

31

Module r

Module s

Module i

Module j

• Approximated number of crossings is calculated with the weights of Simpson’s rule


Solving the Optimization Problem

32

• Find positions of all PSEs

• Constraints– Place all PSEs inside chip area– No overlap between optical devices (PSEs, hubs, memory

controllers)

• Nonlinear Optimization Problem solver: IPOPT (Interior point method)


Manual Layout vs. PROTON

33


Contributions & Acknowledgements

Anja von Beuningen Helmut Graeb Martin Strasser Michael Eick Frank Johannes

34


TUM. Technische Universität München

35

Technische Universität München We create tools for chips Institute for Electronic Design Automation U. Schlichtmann ISPD 2015 - March 31, 2015 Beyond GORDIAN.

Documents