Computational Intelligence Techniques for Electronic Design … · 2015-02-26 · Southampton Seminar 10 Example: Analog IC Sizing (2) B. Liu, F. Fernández, G. Gielen, R. Lopez,

Computational Intelligence Techniques for Electronic Design Automation

Bo Liu

Department of Computing, Glyndwr University, UK

ESAT-MICAS, Katholieke Universiteit Leuven, Belgium

([email protected], [email protected])

mailto:[email protected]

mailto:[email protected]

Outline Electronic design and optimisation

Evolutionary algorithms in EDA: How it works and challenges Surrogate model assisted evolutionary algorithms (SAEA)

The SMAS framework, improvements and applications:

Medium-scale problem: automated design of complex antennas Constraint handling: automated design of mm-wave ICs Integer optimisation: NoC design optimisation Challenges and opportunities

Conclusions

Southampton Seminar 2

Microelectronic design


Analog

IC

Digital IC

Antenna

RF IC

MEMS

Design variables: W, L of each transistor, Cc

For the goal of (such as):

They are optimisation problems

Computational Intelligence (CI)-based EDA

Traditional Design Process and Challenges


Idea of CI-based EDA approaches:

TRANSFORM difficulties on “analysis, intuition and inference” TO difficulties on solving mathematical optimisation problems

Why CI is needed?


• Properties of EDA problems: • Multimodal and global optimisation is necessary

• They are simulation-based optimisation, while analytical equations are often unavailable

• The simulation may cost a long time

• CI-based EDA approaches aims at obtaining highly optimised designs (better than manual design) automatically in a practical time by: • Evolutionary computation (EC)

• Machine learning (ML)





Conclusions


Evolutionary Computation (EC)


• Evolutionary computation is a CI method for optimisation

• EC is based on natural selection, survival of the fittest (objective function) • EC has strengths on black-box and multimodal problems

• Different global optimization algorithms: GA, DE, PSO, IA, AC

Initialize population

Evaluation Fitnees

Select Survivors

Crossover

Mutation

Convergance?

Output Result

Yes

No

CompetitionRanking Based

SelectionRoulette wheelTournament

ReproductionOne point, Two point, Uniform, SBX, Linear, etcBinary, Uniform, Gaussian

How EA Works for EDA Problems?

Simulators: IC: Cadence Virtuoso, Synopsys HSPICE, … Electromagnetic device / antenna: Momentum, CST, SONNET, … MEMS: CoventorWare, MEMS+, COSMOL, … Energy system: AEPS, Energy+, …



Example: Analog IC Sizing (1)

ARCHITECTURE/

TOPOLOGY SELECTION

CIRCUIT SIZING

LAYOUT GENERATION

VERIFICATION

DRC+EXT+LVS

RE

DE

SIG

N L

OO

P

VERIFICATION

SPECS

M1 M2 vinvip

M13

vvcn

M2CM1C

M3C M4C

vvcp

M3 M4

VSS

M6 M12

M8 M10

cln

VDD

cc

M5 M11

M7M9

clp

cc

Mbp

ibb

Mbn

vonvop

min . . 80

250 60 4

4......

area

s b DC gain dB

GBW MHz

phase margin

output swing V

power mW

• Optimise Ws, Ls, Cc, Ibb

• No initial design


Example: Analog IC Sizing (2)

B. Liu, F. Fernández, G. Gielen, R. Lopez, E. Roca, "A Memetic Approach to the Automatic Design of High-Performance Analog Integrated Circuits", ACM Transactions on Design Automation of Electronic Systems, vol. 14, no. 3, pp. 1-24, 2009.

• Two-stage telescopic cascode amplifier sizing with tight constraints (high specifications)

Time: 20 minutes

CI-based EDA : Investigation & Application

EA-based design automation has been investigated and applied Academic:

IEEE: TCAD, TCAS, TVLSI, MTT, TAP, … ACM: TODAES, … Elsevier: Integration journal, Microelectronics journal, … …

Industry:

Cadence Neocircuit: Analog circuit sizing MunEDA WiCkeD: variation-aware analog circuit sizing SolidoDesign: various software for handling process variations …


Main Challenges


• Main challenges:

• Robust design

• System synthesis method: from block to system

• Long simulation time: impractical optimisation time • Analog IC with standard process parameters -> < 2 seconds /

simulation

• Yield estimation of analog IC -> a few minutes / simulation

• RF / mm-wave circuit -> 10-20 minutes / simulation

• Antenna / MEMS / NoC: various from 5 minutes to several hours / simulation

• Evolutionary algorithms often needs several hundreds to thousands of simulations to achieve highly optimised solutions.

Example: On-chip antenna design optimisation:

EM simulation for a candidate design: 20 minutes by ADS-Momentum

Convergence: 800 generations

Population size: 40

20min x 40 x 800 = 1.2 years!!!


Example

B. Liu, H. Aliakbarian, Z.Ma, G. Vandenbosch, G. Gielen, "An Efficient Method for Antenna Design Optimization Based on Evolutionary Computation and Machine Learning Techniques", IEEE Transactions on Antennas & Propagation, vol. 62, no. 1, pp. 7-18, 2014.

Solutions to the efficiency problem


Efficiency enhancement

Evolutionary algorithms

Speed up the simulation

Use fewer simulations

SAEA

Change the optimizer





Conclusions


Introduction to SAEA

Surrogate model assisted evolutionary algorithm (SAEA): Using surrogate models to replace exact function evaluations


Surrogate modelling, Prediction and Prescreening


Ordinary GP modeling

Given training data:

Correlation function:

Maximize likelihood function:

Note: solve in closed form, estimate the hyper-parameters

Best linear unbiased prediction and predictive distribution

variants: simple/blind/…


Prescreening

• With the uncertainty measurement, we can consider the quality of a candidate design in a global picture

• Even the predicted value is bad, promising solutions can still be discovered


D. Jones, 2001. “A Taxonomy of Global Optimization Methods Based on Response Surfaces”, Journal of Global Optimization, pp. 345-383.

Prescreening

• Prescreening methods utilize

the prediction uncertainty.

• Possible promising areas but with less training data can be effectively explored.

•The “guessed” promising points

may not be correct.

• In medium scale (20-30d), prescreening ≈ prediction


A Key Contradiction

Model quality Solution quality

Efficiency *x evalN

How to perform effective global optimisation without high quality surrogate model(s)?

Where are optimal locations of the samples?


Brute force, in-fill sampling, SAEA (1)

Brute force off-line surrogate modelling High quality model but too

time consuming

In-fill sampling techniques (EI, PI, LCB, etc) Design space understanding vs.

Optimisation? More than 15-20 dimensions? Do we need understanding of

the whole space or a single road is OK?


Most available SAEAs

Brute force, in-fill sampling, SAEA (2)

SAEA EA search -> the next sampling point Dozens of variables: EDA problems

EA search pattern Model quality? Most available SAEAs 20-50

dimensional problems still use many exact function evaluations (5000-50000)


Most available SAEAs





Conclusions


Southampton Seminar

Surrogate Model-Aware Evolutionary Search Framework

A new unified method of surrogate modelling and evolutionary search

•Not using more samples, but control the locations of samples, for the objectives of:

• Good prediction accuracy • Effective search mechanism

B. Liu, Q. Zhang, G. Gielen, "A Gaussian Process Surrogate Model Assisted Evolutionary Algorithm for Medium Scale Expensive Black Box Optimization Problems", IEEE Transactions on Evolutionary Computation (In Press).

25

SMAS vs. Present SAEA

SMAS Traditional SAEA


Advantages of SMAS (1)

Property 1: Using the current best candidates as the parent population.

Advantage 1: Optimal solutions are not far away from each other, so a good surrogate model with much fewer samples can be constructed.


Property 2: Only at most one candidate is different from two consecutive parent populations.

Advantage 2: The training data describing the current search region can be much denser.


from child population

Advantages of SMAS (2)

Convergence property of SMAS

For a standard EA, some diversity are useful, while some are not.

SMAS emphasizes exploitation.

The exploration ability can be maintained by selecting appropriate EA operators and parameters.

Simulation: 20-dimensional Ackley function (assuming absolutely

accurate model)


B. Liu, Q. Zhang, G. Gielen, “Behavioral Study of the Surrogate Model-aware Evolutionary Search Framework”, in Proceedings of IEEE World

Congress on Computational Intelligence, 2014 (In Press).

Optimisation Kernel: Differential Evolution

,1 ,2 ,ˆ ( ) [ , , , ]

i i i MX t x x x 1, 2, , NPi

,,

,

( 1), ( ( ) ) ( ),( 1)

( ), , 1, 2, , i j

i j

i j

v t if rand j CR or j randn iu t

x t otherwise j M


Experimental Verifications (1)

Problems: 20-,30-dimensional Ellipsoid (F1-F3, opt:0), Rosenbrock (F4-F6, opt:0), Ackley(F7-F9, opt:0), Griewank(F10-F12, opt:0), 30-dimensional RS-Rastrigin(opt: -330), 30-dimensional RH composition function(opt:10). 1000 evaluations, 20 runs.

SMAS vs. GS-SOMA [Lim IEEE TEVC 2010]

Southampton Seminar

31

Experimental Verifications (2)

SMAS vs. SAGA-GLS [Zhou IEEE TSMC 2007]

SMAS vs. MAES [Emmerich IEEE TEVC 2006]



Automated Design of Complex Antennas (1)

• EAs have been widely used for antenna synthesis, but the long optimisation time largely limits their applications

Example: Four-element antenna array (3.4GHz – 3.8GHz, FR4 substrate)


Automated Design of Complex Antennas (2)

• Maximise realised gain (each sampling point at least 13dB) with S11 below -10dB

Synthesis finished in only one night, with 71.05dB (5 sampling points total) realised gain

Comparable solutions, 3-7 times speed enhancement compared to DE, PSO.

RF IC / mm-wave IC Design Difficulties for manual design

Passive component design is difficult, although circuit configurations are

simpler than analog IC Long simulation time: EM simulation, HB simulation Design experience intensive for the available step-by-step manual design

method

Difficulties for EA-based automated design

Accurate lumped models (computationally cheap) over a wide bandwidth for passive components are difficult to find at high-frequencies

Long simulation time: EM simulation, HB simulation Medium scale (15-40 variables) Complex and tight constraints


Review of RF IC Design Automation

Existing method: equivalent circuit model for passive components, only

applicable to less than 10GHz RF design automation. [Allstot2003Springer] [Ramos2005TCAS ]

mm-wave frequency / general: No design automation method before our work [Liu IEEE TCAD 2012/2014]


The GASPAD Method Focuses on 60GHz and above RF IC

Three Main Goals of the Synthesiser Provide Highly Optimised Results General Enough to Any Circuit Configuration, Any Technology and Any

Frequency Efficient Enough for Practical Use

Update SMAS on constraint handling


B. Liu, D. Zhao, G. Gielen, "GASPAD: A General and Efficient mm-wave Integrated Circuit Synthesis Method Based on Surrogate Model Assisted Evolutionary Algorithm", IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, vol. 33, no. 2, pp. 169-182, 2014.

Updating SMAS for Constraint Handling

New methods to model the

focused search region

New ranking methods

considering constraint

satisfaction

Separate modelling

objective and constraints

Prescreening + predicted

value


mm-wave IC Design Automation (1)

Synthesis of a 60GHz power amplifier in a 65nm CMOS technology (18 parameters)


15 min / simulation

mm-wave IC Design Automation (2)

Design parameters and their ranges


Synthesised Result

Synthesised results: Power added efficiency (@P1dB): 9.85% 1dB compression point: 14.87dBm Power gain: 10.73dB K factor: 10.68 (stable)

About 2 days synthesis time. Much better performance than manual design [He 2010 RFIC]

The first method of general mm-wave IC design automation


Expensive EDA Problems with Integer Variables

• Integer design variables lead to discontinuous landscapes, which is a challenge for:

– Surrogate modelling

– Evolutionary search

• NoC design optimisation


Updating SMAS for Handling Integer Variables Parameter setting and search strategy selection rules:

Scaling factor Self-adaptive crossover rate DE/ctb/1 strategy

Iterative surrogate-assisted neighbourhood exploration method Aims:

Jump out of local optima Direct improvement Assist the SMAS flow

Methods: A self-adaptive perturbation method Local surrogate modelling Greedy search + Opposite DE search


NoC Design Parameter Optimisation

NoC performance

Design parameter opt

A single simulation costs 15 min to 1 hour for 15x15 to 30 x 30 NoC


Architecture (Baseline, VCT, Hybrid, etc.)

Design parameters (Number of virtual channel, buffer depth of the router, etc.)

Application specific (a single loading environment) More general purpose (various loading environments)

Application Specific NoC optimisation


Optimisation of a 15 x 15 NoC with hybrid architecture with the environment of PIR=0.018, PS1=2, PS2=12, BR=0 and hotspot nodes at 35, 57, 108, 155, 175 with 3%, 3%, 2%, 2%, 3% (random traffic)

1 1 2 2 5 5

1 1 2 2 5 5

2

minimize ( , , , , , ,..., , )

s.t. ( , , , , , ,..., , ) 0.16

area 850mm

c p

c p

D N S X Y X Y X Y

E N S X Y X Y X Y J

2

37 hours optimisation55.780.153

805.14

D cycles

E J

area mm

[1,10], [1,12], [1,8], , [1,15] (all integers)c pN S B X Y

NoC Design for General Purpose Dozens of typical loading environments need to be

considered: Example: CMP with cache coherence protocols: 80 combinations of typical broadcast ratio, packet size and

packet injection rate

8 x 8 NoC optimisation for CMP with cache coherence protocols has been addressed, but 15 x 15 NoC, 30 x 30 NoC?


Prohibitively expensive simulation

A Few Challenges and Opportunities (1) Computationally prohibitively expensive

optimisation More than 1 hour / simulation Dozens of design variables Complex landscapes

EDA problems General purpose large dimensional NoC design optimisation Some problems in MEMS design optimisation 3D complex antenna design Some problems in optical devices design optimisation …


A Few Challenges and Opportunities (2) Computationally expensive multiobjective

optimisation

EDA problems: almost every problem


B. Liu, Q. Zhang, F. Fernández, G. Gielen, "An Efficient Evolutionary Algorithm for Chance-constrained Bi-objective Stochastic Optimization and Its Application to Manufacturing Engineering", IEEE Transactions on Evolutionary Computation, vol. 17, no. 6, pp. 786-796, 2013.

A Few Challenges and Opportunities (3) Possible starting points: A system utilising various electronic elements (IC,

MEMS, antenna, optical device, etc) Example: bio-medical system, energy harvesting system

Starting from computational expensive blocks Example: MEMS

From block synthesis to system synthesis

Co-design of hardware and algorithms






Conclusions


Conclusions (1) Handling expensiveness is unavoidable for future (intelligent) EDA

tools in many applications.

In most present SAEAs, the surrogate modelling and the evolutionary search are loosely cooperated.

New SAEA frameworks (unconstrained / constrained continuous / discrete optimisation) are presented, showing significant improvements in terms of efficiency while keeping the high solution quality. Some computationally intractible EDA problems can therefore be solved.


Conclusions (2) CI methods based EDA tools

is possible to have a large impact for electronic design (IC, MEMS, antenna, optical devices, energy systems, etc) in the coming future.

Engineers with a new knowledge structure are needed.

Promote this emerging multidisciplinary research and education.


M. Fakhfakh, E. Cuautle,

P. Siarry

Computational Intelligence in Electronic Design Springer, 2015

Acknowledgments

Sincere thanks to Prof. Georges Gielen, Prof. Francisco V. Fernandez, Prof. Alex Yakovlev, Dr. Patrick Degenaar, Prof. Qingfu Zhang, Prof. Guy Vandenbosch, Prof. Tom Dhaene, Prof. Vic Grout, Dr. Hao-ming Chao, Mr. Dixian Zhao, Mr. Ammar Karkar Mr. Noel Deferm, Dr. Brecht Machiels, Miss Ying He, Mr. Mengyuan Wu, Mr. Bohan Yang, Miss Borong Su, Miss Wan-ting Lo for their valuable help!


Thank you

Thank you!


Documents

Computational Intelligence Techniques for Electronic Design … · 2015-02-26 · Southampton Seminar 10 Example: Analog IC Sizing (2) B. Liu, F. Fernández, G. Gielen, R. Lopez,