magestic: machine-learning for automatic generation of

26
1 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited ERI Design: IDEA and POSH IDEA and POSH Integration Exercise Detroit, MI July 17 July 19, 2019 Cadence Design Systems, NVIDIA and Carnegie Mellon University MAGESTIC: Machine-Learning for Automatic Generation of Electronic Systems Through Intelligent Collaboration This research was developed with funding from the Defense Advanced Research Projects Agency (DARPA). The views, opinions, and/or findings expressed are those of the author and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government.

Upload: others

Post on 24-Nov-2021

5 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.1 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

ERI Design: IDEA and POSH

IDEA and POSH Integration Exercise

Detroit, MI

July 17 – July 19, 2019

Cadence Design Systems, NVIDIA and Carnegie Mellon University

MAGESTIC: Machine-Learning for Automatic Generation

of Electronic Systems Through Intelligent Collaboration

This research was developed with funding from the Defense Advanced Research Projects Agency (DARPA). The views, opinions, and/or findings expressed are those of the author and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government.

Page 2: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.2 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

• Team: Cadence Design Systems, NVIDIA and Carnegie Mellon University

• Goals: – No Human in the Loop (NHIL) Design Flow, Silicon Compiler in < 24 hours TAT

– Research in Analog IC Design through Package and PCB Design

– Achieve Automated Design Across a Broad Spectrum of Designs in Year 2

• Agenda for Today:

– Software Demonstration of NHIL Package and PCB Design Progress

– Focus for NHIL Package and PCB Design Research in Year 2

– Software Demonstration of NHIL Analog IC Design Progress

– Focus for NHIL Analog IC Design Research in Year 2

• Visit MAGESTIC Team to see Live Demos of the Design Data Presented Today

MAGESTIC Program

Taylor Hogan

Elias Fallon

Page 3: MAGESTIC: Machine-Learning for Automatic Generation of

DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

PCB & Package SynthesisERI Conference 2019

Page 4: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.4 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

• The Dark Age–Interactive, NAIL

• The Industrial Age–NHIL for Some, Hope

• The Age of Enlightenment–NHIL for All, Explorations

EDA Ages

Page 5: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.5 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

The Dark Ages

Page 6: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.6 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

The Dark Ages Continue

Page 7: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.7 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

The Industrial Age

Page 8: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.8 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

Industrial Age Training Neural Networks to Avoid Intractable Computational Problems

GO Floorplan

TRAIN modeNetlist

FloorplansRUDY

congestion

images

CN

N

GR

routability

metric

GO Floorplan

INFERENCE

mode

Netlist

CNN

Design Features

𝐻𝑃𝑊𝐿𝐵𝑜𝑎𝑟𝑑 𝐷𝑖𝑚𝑠

⋮+

Page 9: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.9 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

The Industrial Age … Hope

Page 10: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.10 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

Industrial Age Continues

Page 11: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.11 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

What WorksWhat Does Not Work

55

88

100

60

90

100 100 100 10092

72

100

64

90

42

24

100

89

100

91 91

30

100 100 100 100 100

82 80

90

100

89 90

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33

Percent Complete

Open

Source

RISC PCB

Complex

Interposer Mobile

PCB

GPU

PCB

Small

Analog

PCB

Complex

Analog

Page 12: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.12 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

Achieving NHIL for All Year 2 Goals

Page 13: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.13 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

Age of EnlightenmentFeature Sensitivity Analysis

• Development of methods and tools that aid in the interpretability of neural networks by humans

• Explain what features matter

Feature Analysis

Input

Feature

Dataset

Model

Long-Range RUDY

Macros

Pin Density

RUDY Pins

. . .

. . .

Year 2 Goals

Page 14: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.14 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

The Age of EnlightenmentData Mining to eliminate typing in the known

Old Designs

Data SheetsDesign parser

Information extraction

database

Machine learning

Model training

Model Prediction

Python models

new

designs

Year 2 Goals

Page 15: MAGESTIC: Machine-Learning for Automatic Generation of

DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

Analog ICERI Conference 2019

Page 16: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.16 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

MAGESTIC Analog

Inte

gra

ted F

low

Analog

Circuit

Analog

Layout

Analog

Circuit Sizing

Floorplanning

Placement

Routing

Layout

Finishing

Verify &

Extract

Electrical

Performance

Exploration and Innovation

Genetic

Optimizing Router

Full Graph-

Embedded RL

Global Router

CMU RL Agent

Global Router

AlphaRoute

Research Dimensions

• Integrated flow− See complex interactions between steps

− Evaluate overall electrical performance

• Exploration and Innovation − Creative research into new approaches

Page 17: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.17 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

ML based yield

improvements

Within design

metrics, better yield

Page 18: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.18 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

Inte

gra

ted F

low

Page 19: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.19 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

Analog

Circuit

Analog

Layout

Analog

Circuit Sizing

Floorplanning

Placement

Routing

Layout

Finishing

Verify &

Extract

Electrical

Performance

Inte

gra

ted F

low

Page 20: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.20 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

Analog

Circuit

Analog

Layout

Analog

Circuit Sizing

Floorplanning

Placement

Routing

Layout

Finishing

Verify &

Extract

Electrical

Performance

Inte

gra

ted F

low

Page 21: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.21 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

Layout Group Prediction Problem

• It is common practice for layout designers to group devices that need good matching– Grouped devices are placed closely together,

and there may be interdigitation

• Grouping decisions are largely made from experience– Even simulation cannot fully predict effects of grouping

– Can we learn them from good hand-crafted layouts?

• To formulate a machine learning problem, we need to define grouping as a label and predict it from schematic features

M0.1 M1.1

M1.2 M0.2

Page 22: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.22 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

Gro

upin

g M

odel

Utiliz

ed

Analog

Circuit

Analog

Layout

Analog

Circuit Sizing

Floorplanning

Placement

Routing

Layout

Finishing

Verify &

Extract

Electrical

Performance

Inte

gra

ted F

low

Page 23: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.23 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

But Wait! There’s More!

Figure 3.13 Actual Capacitance v/s Error Percentage plots across all nets.

Parasitic Prediction Graph Neural Networks for

Routing Cost Prediction

AlphaRoute /

AlphaAnalogDesign Updates

Exploration and Innovation

Please stop by during the IDEA Integration Exercise to see a demo

or discuss in more detail many of our other exciting projects

Page 24: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.24 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

Grouping Cold Start Problem

0 1

2 3

4 5

6 7

8 9 10 11

Suppose we are given the following (sparse) labels:

- Transistor 0, 4, 6, and 8 will be in different groups

- Transistor 6 & 7 belong to the same group

Can we train a reasonable grouping model using the netlist information?

Exploration and Innovation

Graph Convolutional Network (GCN)

Representation of Circuit

Page 25: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.25 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

GCN Training for Grouping Cold Start Model

The learning curve of 10 runs, each with a

maximum of 50 epochs

Page 26: MAGESTIC: Machine-Learning for Automatic Generation of

© 2019 Cadence Design Systems, Inc. All rights reserved.26 DISTRIBUTION STATEMENT A: Approved for Public Release, Distribution Unlimited

Cadence Design Systems, NVIDIA and Carnegie Mellon University

This research was developed with funding from the Defense Advanced Research Projects Agency (DARPA). The views, opinions, and/or findings expressed are those of the author and should not be interpreted as representing the official views or policies of the Department of Defense or the U.S. Government.