1 The Optimization of High- Performance Digital Circuits Andrew Conn (with Michael Henderson and Chandu Visweswariah) IBM Thomas J. Watson Research Center.

1

The Optimization of High-Performance Digital Circuits

Andrew Conn (with Michael Henderson and Chandu Visweswariah)

IBM Thomas J. Watson Research Center

Yorktown Heights, NY

2

Outline

3

Circuit optimization

4

Dynamic vs. static optimization

Transistor and wire sizes

SimulatorNonlinearoptimizer

Function and gradient values

Transistor and wire sizes

Static timinganalyzer

Nonlinearoptimizer

Function and gradient values

5

Dynamic vs. static optimization

Dynamic Static

Example: DELIGHT.SPICE,JiffyTune

Example: TILOS, Contrast,EinsTuner

Optimizes specified paths Optimizes all paths

Needs input vectors None required

Requires carefully thought-outproblem statement

Automatic

Hard to use Easy

Not susceptible to false paths,pessimism

Susceptible

Easier to add noise/powerconsiderations

Harder

6

Custom? High-performance?

7

EinsTuner: formal static optimizer

Embedded time-domain simulator

SPECS

Static transistor-level timerEinsTLT

Nonlinearoptimizer

LANCELOT

Transistor andwire sizes

Function andgradient values

8

Components of EinsTuner

Read netlist; create timing graph (EinsTLT)Formulate pruned optimization problemFeed problem to nonlinear optimizer (LANCELOT)

Snap-to-grid; back-annotate; re-time

Solve optimization problem, call simulatorfor delays/slews and gradients thereof

Obtain converged solution

Fast simulation and incrementalsensitivity computation (SPECS)

1

2

34

9

]464

,363

[max6

s.t.]

252,

151[max

5s.t.

]686

,585

[max8

s.t.]

676,

575[max

7s.t.

)]8

,7

([maxmin

dATdATATdATdATATdATdATATdATdATAT

ATAT

1AT

2AT

15d

25d

3AT

4AT

36d

46d

57d

68d

58d 67d

5AT

6AT

7AT

8AT

Static optimization formulation

10

1f2f

x0x

z

x

*z

Digression: minimax optimization

)( s.t.)( s.t.

min

])}(),(max{[min

2

1

,

21

xfzxfz

z

xfxfx

zx

11

)],min[max( 87 ATAT

],max[ 6865858 dATdATAT




Remapped problem

8

7

s.t.

s.t.

min

ATz

ATz

z

6868

5858

dATAT

dATAT

6767

5757

dATAT

dATAT

2525

1515

dATAT

dATAT

4646

3636

dATAT

dATAT

1AT

2AT

15d

25d

3AT

4AT

36d

46d

57d

68d

58d 67d

5AT

6AT

7AT

8AT

12

Springs and planks analogy

z7AT

8AT

5AT6AT

1AT 2AT 3AT 4AT15d 25d

36d 46d

57d67d 58d 68d

13

Springs and planks analogy

z7AT

8AT

5AT6AT


36d 46d

57d67d 58d 68d z

7AT8AT

5AT6AT


36d 46d

57d67d 58d 68d z

7AT8AT

5AT6AT


36d 46d

57d67d 58d 68d z

7AT8AT

5AT6AT


36d 46d

57d67d 58d 68d

14

Algorithm animation: inv3

Del

ay

Logic stages

PIs by criticality WireG

ate

Red=criticalGreen=non-criticalCurvature=sensitivityThickness=transistor size

• One such frame per iteration

15

Algorithm demonstration: 1b ALU

16

),,,(

),,,(

),,,(

),,,(

rinjpn

fij

fj

finjpn

rij

rj

rinjpn

fij

ri

fj

finjpn

rij

fi

rj

slewcoutwwsslew

slewcoutwwsslew

slewcoutwwdATAT

slewcoutwwdATAT

i j

pn ww ,

jcout

Constraint generation

17

Statement of the problem

arc geach timinfor ),( s.t.

arc timingeachfor),( s.t.POsallfor s.t.

min

SWijsjs

SWijdiATjATiRATiATz

z

POs allfor POmax s.t.

nets internal allfor internalmax s.t.

FETs allfor maxmin s.t.gates allfor max)(min s.t.PIs allfor itarget)(ipincap s.t.

targetarea s.t.

SiS

SiS

WiWWWi

WW

18

SPECS: fast simulation

• Two orders of magnitude faster than SPICE

• 5% typical stage delay and slew accuracy; 20% worst-case

• Event-driven algorithm

• Simplified device models

• Specialized integration methods

• Invoked via a programming interface

• Accurate gradients indispensable

19

LANCELOT

20

LANCELOT algorithms

•Uses augmented Lagrangian for nonlinear constraints (x,) = f(x) + [i ci (x) + ci (x)2 /2]

•Simple bounds handled explicitly

•Adds slacks to inequalities

•Trust region method

21

LANCELOT algorithms continued

Simple bounds

a

b

Trust-region

kx

u

f

wv

2x

1x

bx

ax

xxf

2

1

21

0

0

),(min

22

Customization of LANCELOT• Cannot just use as a black box

• Non-standard options may be preferable

eg Solve the BQP subproblem accurately

• Magic Steps

• Noise considerations

• Structured Secant Updates

• Adjoint computations

• Preprocessing (Pruning)

• Failure recovery in conjunction with SPECS

23

LANCELOT

• State-of-the-art large-scale nonlinear optimization package

• Group partial separability is heavily exploited in our formulation

• Two-step updates applied to linear variables

• Specialized criteria for initializations, updates, adjoint computations, stopping and dealing with numerical noise

24

Aids to convergence• Initialization of multipliers and variables

• Scaling, choice of units

• Choice of simple bounds on arrival times, z

• Reduction of numerical noise

• Reduction of dimensionality

• Treating fanout capacitances as“internal variables” of the optimization

• Tuning of LANCELOT to be aggressive

• Accurate solution of BQP

25

Demonstration of degeneracy

z

26

Demonstration of degeneracy

z

Deg

ener

acy!

27

Why do we need pruning?

28

Pruning of the timing graph

• The timing graph can be manipulated– to reduce the number of arrival time variables– to reduce the number of timing constraints– most of all, to reduce degeneracy

• No loss in generality or accuracy

• Bottom line: average 18.3xAT variables,33% variables, 43% timing constraints, 22% constraints, 1.7x to 4.1xin run time on large problems

29

Pruning strategy

• During pruning, number of fanins of any un-pruned node monotonically increases

• During pruning, number of fanouts of any un-pruned node monotonically increases

• Therefore, if a node is not pruned in the first pass, it will never be pruned

• Therefore, a one-pass algorithm can be used for a given pruning criterion

30

Pruning strategy

• The order of pruning provably produces different (possibly sub-optimal) results

• Greedy 3-pass pruning produces a “very good” (but perhaps non-optimal) result

• We have not been able to demonstrate a better result than greedy 3-pass pruning

• However, the quest for a provably optimal solution continues...

31

Pruning: an example

frf

rfr

frf

rfr

frf

rfr

f

r

dATATdATATdATATdATATdATATdATAT

ATzATzz

1212

1212

2323

2323

3434

3434

4

4

s.t. s.t. s.t. s.t. s.t. s.t.

s.t. s.t.min

rfrf

frfr

dddATzdddATz

z

3423121

3423121

s.t. s.t.

min

1 2 3 4

32

Block-based vs. path-based timing

463436

462426

461416

453435

452425

451415

ddATATddATATddATATddATATddATATddATAT

Blo

ck-b

ased P

ath-based

12

3

4

5

6

34342424

1414

4646

4545

dATATdATATdATAT

dATAT

dATAT

33

Block-based & path-based timing

• In timing graph, if node has n fanins, m fanouts, eliminating it causes 2mn constraints instead of 2 (m+n)

• Criterion: if 2mn 2(m+n)+2, prune!

1

2

3

4 5

6

14d24d

34d

45d

46d

1

2

3

5

6

4514 dd

4524 dd

4634 dd

4614 dd

4624 dd

4534 dd

34

Detailed pruning example

123

456

7

8

9

10

11

12

13

14

15

16

35


1

2

3

7 9 11 14

4

5

6

8 10 13 16

Sink12 15Source

Edges = 26Nodes = 16 (+2)

36


7 9 11 141

2

345

6 8 10 13 16

Sink12 15Source

Edges = 26 20Nodes = 16 10

37


1

2

3

7 9 11

45

6 8 10 13

Sink12Source

14

16

15

14

Edges = 20 17Nodes = 10 7

38


9 11 14

45

6 8 10 1316

Sink12

15Source

14

1,72,7

3,7


39


9 11,14

45

6 8 10 1316

Sink12

15Source

14

1,72,7

3,7


40


9 11,14

45

6 8 1013,16

Sink12

15Source

14

1,72,7

3,7


41


9 11,14

45

6 8

10

10,13,16

Sink12

15Source

14

1,72,7

3,7


42


9 11,14

45

6 810,13,16

SinkSource

10,12,14

12,15

10,12,15

12,141,7

2,7

3,7


43

Pruning vs. no pruning

44

Adjoint Lagrangian modeAdjoint Lagrangian mode

– gradient computation is the bottleneck– if the problem has m measurements and n

tunable transistor/wire sizes:• traditional direct method: n sensitivity simulations• traditional adjoint method: m adjoint simulations

– adjoint Lagrangian method computes all gradients in a single adjoint simulation!

45

Adjoint Lagrangian modeAdjoint Lagrangian mode

– useful for large circuits– implication: additional timing/noise

constraints at no extra cost!– is predicated on close software integration

between the optimizer and the simulator– gradient computation is 8% of total run time

on average

46

Noise considerationsNoise considerations

– noise is important during tuning

– semi-infinite problem

],[in allfor ),( 21 tttNMtxv L

v

t

area = c(x)

NML

t1 t2

47


• Trick: remap infinite number of constraints to a single integral constraint c(x) = 0

• In adjoint Lagrangian mode, any number of noise constraints almost for free!

• General (constraints, objectives, minimax)

• Tradeoff analysis for dynamic library cells

v

t

area = c(x)

NML

t1 t2

48


49

Initialization of s

upstream downstream

jkij

1/2

1/2

1/6

1/4

1/4

1POs1/61/6

50

Some numerical results - Dynamic

Name # gatesStart delay

(ps)End delay

(ps)%

improvement#

iterationsCPU

time (s)

s390-1 22 269.4 255.5 5.2 51 100.2

s390-2 24 388.7 350.0 10.0 110 945.4

s390-3 41 1568.4 1548.6 11.4 49 462.7

ppc-1 155 584.2 526.4 9.9 35 2247.9

s390-4 175 306.2 231.8 24.3 55 17664.8

ppc-2 218 627.6 555.5 11.5 74 25342.2

ppc-3 408 589.8 584.4 0.9 30 2950.8

s390-5 430 943.5 623.4 33.9 228 27125.4

s390-6 559 884.1 742.6 16.0 107 52154.0

ppc-4 628 1276.7 1091.3 14.5 159 284445.0

51

Some numerical results - Static

Name # gatesStart slack

(ps)End slack

(ps)%

improvement#

iterationsCPU

time (s)

s390-1 22 645 668 3.5 34 29.3

s390-2 24 583 617 5.8 46 141.3

ppc-1 155 514 555 7.9 56 528.9

s390-4 175 609 684 12.3 59 355.5

ppc-2 218 454 525 15.6 64 1003

x1 138 -129.1 -41.36 68 49 1321

ppc-3 408 447 512 14.5 68 861.4

s390-5 430 5 424 8380 131 3600

x2 507 -12.53 166 1425 59 4096

x3 743 -265.8 -86.31 67.5 73 30850

52

– Motivation

Lagrangian relaxation

•Tuning community loves the approach

–reduces the size of the problem

–reduces redundancy and degeneracy

– BUT …Never get something for nothing

53

),,,(

),,,(

),,,(

),,,(

rinjpn

fij

fj

finjpn

rij

rj

rinjpn

fij

ri

fj

finjpn

rij

fi

rj

slewcoutwwsslew

slewcoutwwsslew

slewcoutwwdATAT

slewcoutwwdATAT

Lagrangian relaxation (continued)

ri

ri

fi

fi

RATATz

RATATz

•Complicating constraints

Relax into objective function

54

• Substituting the first-order optimality conditions on removes the dependence on z and AT’s --- because of the problem’s structure!

Lagrangian relaxation (continued)

55

Lagrangian Relaxation

0

2000

4000

6000

8000

10000

0 2000 4000Transistors

Var

iabl

es

Lagrangian Relaxation

Pruning

Neither

56

Lagrangian relaxation (continued)Name FETs Iterations CPU(s)

w/o w w/ w/o w

inv3 8 61 2 36 5.1 5.3

a3.2 34 41 17 215 21.5 138.2

Agrph 34 30 28 303 38.1 434.7

ldder 46 41 16 182 21.5 121.1

s3901 72 40 34 109 87.1 479.7

s3902 102 67 61 471 990.5 842.6

s3903 154 26 29 131 236.8 1502

ppc-1 824 37 117 1181 1818 57430

s3904 882 20 156 386 532.8 14553

c8 584 53 194 1372 1004 37068

57

Future work and conclusions –Elaine/IPOPT (Andreas Waechter)

•Handle linear constraints as before/directly

•Nonlinear constraints as before/filter method

•Simple bounds via primal-dual interior point method

• Spherical trust region scaled appropriately/line search method

58

References

C. Visweswariah, A. R. Conn and L. Silva

Exploiting Optimality Conditions in Accurate Static Circuit Tuning

in High Performance Algorithms and Software for Nonlinear Optimization,

G. DiPillo and A. Murli, Eds, pages 1-19, Kluwer, 2002 (to appear)

A. R. Conn, P. K. Coulman, R. A. Haring, G. L. Morrill, and C. Visweswariah.

Optimization of custom MOS Circuits by transistor sizing.

To appear (2002) in the book "The Best of ICCAD - 20Years of Excellence in

Computer Aided Design". Originally appeared as IEEE International Conference

on Computer-Aided Design, pages 174--180, Nov 1996

A.R. Conn and C. Visweswariah

Overview of continuous optimization advances and applications to circuit tuning.

Proceedings International Symposium on Physical Design

(2001), pp. 74-81, ACM Press, New York.

•

59

References

C. Visweswariah, R. A. Haring, and A. R. Conn.

Noise considerations in circuit optimization.

IEEE Transactions on Computer-Aided Design of ICs and Systems,

Vol. 19, pages 679-690, June 2000.

C. Visweswariah and A. R. Conn

Formulation of static circuit optimization with reduced size, degeneracy and

redundancy by timing graph manipulation.

IEEE International Conference on Computer-Aided Design, pages 244--251, Nov. 1999.

A. R. Conn, L. N. Vicente, and C. Visweswariah.

Two-step algorithms for nonlinear optimization with structured applications.

SIAM Journal on Optimization, volume 9, number 4, pages 924--947, September 1999.

60

References

A. R. Conn, I. M. Elfadel, W. W. Molzen, Jr., P. R. O'Brien, P. N. Strenski,

C. Visweswariah, and C. B. Whan

Gradient-based optimization of custom circuits using a static-timing formulation.

Proc. Design Automation Conference, pages 452--459, June 1999.

A. R. Conn, R. A. Haring, and C. Visweswariah

Noise considerations in circuit optimization.

IEEE International Conference on Computer-Aided Design, pages 220--227, Nov 1998.

A. R. Conn, P. K. Coulman, R. A. Haring, G. L. Morrill, C. Visweswariah, and

C. W. Wu.

JiffyTune: circuit optimization using time-domain sensitivities.

IEEE Transactions on Computer-Aided Design of ICs and Systems, number 12,

volume 17, pages 1292--1309, December 1998.

61

References

A. R. Conn, R. A. Haring, C. Visweswariah, and C. W. Wu.

Circuit optimization via adjoint Lagrangians.

IEEE International Conference on Computer-Aided Design, pages 281--288, Nov 1997.

A. R. Conn, R. A. Haring, and C. Visweswariah.

Efficient time-domain simulation and optimization of digital FET circuits.

Mathematical Theory of Networks and Systems, May 1996.

A popular article on our work by Stewart Wolpin

http://www.research.ibm.com/thinkresearch/pages/2002/20020625_einstuner.shtml

1 The Optimization of High- Performance Digital Circuits Andrew Conn (with Michael Henderson and Chandu Visweswariah) IBM Thomas J. Watson Research Center.

Documents

number of timing

increasesduring pruning

timing constraints

nonlinear constraints

unpruned node

number of fanins

number of fanouts

timing graphthe timing