Slide 1 © Carliss Y. Baldwin 2006 The Power of Modularity: The Financial Consequences of Computer and Code Architecture Carliss Y. Baldwin Harvard Business.

Slide 1 © Carliss Y. Baldwin 2006

The Power of Modularity: The Financial Consequences of Computer and Code Architecture

Carliss Y. BaldwinHarvard Business School

AOSD 06March 22, 2006


Unmanageable Designs—What They Are and their Financial Consequences

Carliss Y. BaldwinHarvard Business School

AOSD 06March 22, 2006


Three Points to begin Large, complex, evolving designs

– Are a fact of modern life– Need design architectures—

» “Description of the entities in a system and their relationships”» Way of assigning work (Parnas)

Designs create value– Value operates like a force in the economy– We fight to create it and to keep it—using strategy

Design Architecture and Strategy– How can you create and capture value in a large,

complex evolving set of designs?– Subject of this talk


In the economy, value acts like a force

Value = money or the promise of money

Consider the computer industry…


The changing structure of the computer industry

Andy Grove described a vertical-to-horizontal transition in the computer industry:

1995-“Modular Cluster”

1980-“Vertical Silos”


Andy’s Movie

Stack View in 1980Services S

PSystems Integration E

RR

Applications Layer Y D CVCMiddleware Layer U H E

Operating Systems IBM N P CS

Hardware Y XRCSAMP

ComponentsTI Intel

Top 10 Public Companies in US Computer Industry

Area reflects Market Value in Constant US $


Andy’s Movie

Stack View in 1995SPERRY D CVCU H E

IBM N P CSY XRCSAMP

TI Intel



ServicesFirst Data

Systems Integration EDSOracle

I CAApplications Layer B MSFTMiddleware Layer M

Operating Systems

Hardware: Printers HPHardware: Servers IBMHardware: Routers Cisco

Components IntelMicron


Andy’s Movie—the Sequel

Stack View in 2002SPERRY D CVCU H E

IBM N P CSY XRCSAMP

TI Intel



ServicesFirst Data



Operating Systems



Services First DataADP

Systems Integration

OracleApplications Layer IBM

Middleware Layer MSFT

Operating Systems

Hardware: Printers HPHardware: PCs Dell

Hardware: Servers IBMHardware: Routers Cisco

Components Intel TI


Turbulence in the Stack

Departures from Top 10: Xerox (~ bankrupt) DEC (bought) Sperry (bought) Unisys (marginal) AMP (bought) Computervision (LBO)

Arrivals to Top 10: Microsoft Cisco Oracle Dell ADP First Data

Sic Transit Gloria Mundi … Sic Transit


Contrast to the Auto Industry

Top 10 Public Companies in US Auto Industry


Other Chassis/Powertrain

Other Interior

Other MultipleElectrical/Electronics

Other Misc.

Toyota Nissan DaimlerChrysler

Honda GM Ford VW

Oth

ers

Johnson Controls

Magna

Eaton

20031982

AftermarketAUGAT (electronics)

General Motors

FORDC

hry

sler

Tenneco

Toyota

Ho

nd

a

Dana Eaton

TRWInterior

Other Powertrain

Value stayed in one layer!


Two patterns

“Manageable” designs = auto industry “Unmanageable” designs = computer

industry

What makes computer designs so unmanageable?

This was the question Kim Clark and I set out to answer in 1987.


After studying the history of computer designs and correlating their changes with value changes

We concluded that modularity was part of the answer…


Modularity is

The degree to which a set of designs (or tasks) is partitioned into components, called modules, that are

highly dependent within a module, nearly independent across modules

A property of architecture, somewhat under the architect’s control


Modularity in computers—IBM System/360

First modular computer design architecture (1962-1967)– Proof of concept in hardware and application

software– Proof of option value in market response and

product line evolution– System software was NOT modularizable

» Fred Brooks, “The Mythical Man Month”

» Limits of modularity


Strategically—IBM wanted to be the sole source of all of System/360’s Modules

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 34 35

1 SLT architecture and standard circuits2 Erich Bloch - August 19613 New Processor Line Architectural Ground Rules4 SPREAD Task Group - 12/28/615 New Processor Line control, product and programming standards6 Corporate Processor Control Group (CPC) - 4/1/62

7 SLT Transistors8 SLT Modules9 SLT Cards

10 SLT Boards and Automatic Wiring11 Processor 1 - Endicott, New York12 Processor 2 - Hursley, England13 Processor 3 - Poughkeepsie, New York14 Processor 4 - Poughkeepsie, New York15 Processor 5 - Poughkeepsie, New York16 Main memories, Corporate Memory Group (1)17 Internal memories, CMG18 Read-only memories for control, CMG19 "Binary-addressed" Random Access Files20 Corporate File Group (2)21 Tape devices running at 5000+ char/sec22 Corporate Tape Group (3)23 Time-multiplex system for switching I/O devices24 DSD Technical Development Group

25 Techniques to measure processor performance, system26 throughput and software efficiency, Group Staff27 A unified Input/output Control Structure (IOCS)28 System Software for Configuration I (4)29 System Software for Configuration II (4)30 System Software for Configuration III (4)31 FORTRAN and COBOL compilers32 A unified programming language

33 Announcement and Marketing34 Production, Testing and Integration35 Shipment, Delivery and Installation


By 1980, 100s of firms made S/360 “plug-compatible” components

Code Category Definition 1960 1970 1980

3570 Computer and Office Equipment 5 2 93571 Electronic Computers 1 8 293572 Computer Storage Devices 1 6 36 *3575 Computer Terminals 2 5 23 *3576 Computer Communication Equipment 1 1 10 *3577 Computer Peripheral Devices, n.e.c. 3 5 12 *3670 Electronic Components and Accessories 11 7 11 *3672 Printed Circuit Boards 2 19 39 *3674 Semiconductors and Related Devices 8 4 10 *3678 Electronic Connectors 5 15 16 *7370 Computer Programming, Data Processing,

and Other Services 1 9 26 *7371 Computer Programming Services 0 2 12 *7372 Prepackaged Software 0 7 13 *7373 Computer Integrated Systems Design 1 3 167374 Computer Processing, Data Preparation

and Processing 0 5 29 *7377 Computer Leasing 0 10 7 *

41 108 298

* Firms in these subindustries make modules of larger computer systems. Firms making modules = 34 95 244Percent of total = 83% 88% 82%


Modularity and Option Value Interact

IBM did not understand the option value it had created

Did not increase its inhouse product R&D Result: Many engineers left

– to join “plug-compatible peripheral” companies San Jose labs —> Silicon Valley


Modularity in computers, cont. Bell and Newell, Computer Structures (1971)

– General principles of modular design for hardware– Basis of PDP-11 design—another ORMDA

Thompson and Ritchie, Unix and C (1971-1973)– Modular design of operating system software (contra

Brooks Law) Parnas (1972) abstract data structures, info hiding

– Object-oriented programming, C++, Java Mead and Conway, Intro to VLSI Systems (1980)

– Principles of modular design for large-scale chips


Modularity in computers (cont.) IBM PC (1983)

– DEC PDP-11 minimalist strategy (exclude and invite) – + Intel 8088 chip– + DOS system software – + IBM manufacturing – + Lotus 1-2-3– A modular design architecture with a mass market


Creating a modular architecture


. x x x x xx . x x x x x x x x x

Drive x x . x x xSystem x x x . x x x x x x x x

x x . xx x x x . x x x

x x x . x xx x x . x x x x

x x x . x x x x xMain x x x . x x xBoard x x x x x x x x . x x x x x

x x x x x . x xx x x x x x . x x x

x x x . x

x x x . x x xx x x x . x x x x

LCD x x x . x xScreen x x x x . x x x

x x x x x x x . x x xx x x . x

x x x x . x x x xx x x . x x x x

x x x x x . x x xPackaging x x x x . x x

x x x x x . x xx x x x . x x

x x x x x .x x x x x .

Graphics controller on Main Board or not?If yes, screen specifications change;If no, CPU must process more; adopt different interrupt protocols

Design Structure Matrix Map of a Laptop Computer


. x x x x xx . x x x x x x x x x

Drive x x . x x xSystem x x x . x x x x x x x x

x x . xx x x x . x x x

x x x . x xx x x . x x x x

x x x . x x x x xMain x x x . x x xBoard x x x x x x x x . x x x x x

x x x x x . x xx x x x x x . x x x

x x x . x

x x x . x x xx x x x . x x x x

LCD x x x . x xScreen x x x x . x x x

x x x x x x x . x x xx x x . x

x x x x . x x x xx x x . x x x x

x x x x x . x x xPackaging x x x x . x x

x x x x x . x xx x x x . x x

x x x x x .x x x x x .

Graphics controller on Main Board or not?If yes, screen specifications change;If no, CPU must process more; adopt different interrupt protocols

Every important cross-module interdependency must be addressed via a design rule.

This is costly

Modularization may not pay


Design Structure Matrix of a Modular Laptop Computer

. x x x xx . x x

Design x . x x Design Rules Task GroupRules x x . x

x x x .x . x x x

x x . x x xDrive x x x x . xSystem x x x x x . x x Hidden Modules

x x x . x many Task groupsx x x x .

x . x xx x x x x . x x

x x . x x x xMain x x x x x . x xBoard x x x x x x x . x x

x x x x x . xx x x x x x . x

x x x x x .x x . x x x

x x x . x x xLCD x x x . xScreen x x x x x . x x

x x x x x . xx x x x x x .

x x . x x x xx x x . x x x x

x x x . x x xPack- x x x x x x . x xaging x x x . x x

x x x x x . x xx x x x x .

x x x x x x .x x x x x x . x x x x

System x x x x x x x x . x x System Testing x x x x x x x x x . x x x Integration& Integ- x x x x x x x x x x x x and Testingration x x x x x x x x . x Task Group

x x x x x x x x x x x .


Measuring modularity

Mozilla just after becoming open source Linux of similar size

Coord. Cost = 30,537,703Change Cost = 17.35%


Comparison of different software systems with DSM tools

© Alan MacCormack, Johh Rusnak and Carliss Baldwin, 2006

Mozilla just after becoming open source Linux of similar size



Conway’s Law: Different organizations deliver different architectures

One Firm, Tight-knit Team, RAD methods

Distributed Open Source Development


Mozilla Before Redesign Mozilla After Redesign

Refactoring for modularity


Change Cost = 17.35% Change Cost = 2.38%


What about Aspects?


Creating an Aspect Module

Imagine 100s of A’s and B’s!

APIsXPIs

A

B

C

Design Rules = APIs and XPIs

APIs

A

B

Design Rules = APIs only

Modular Operator “Inversion”

Cross-cutting concern


What is the Aspect worth?

If no change ever, Aspect is worthless Reasons to change the Aspect sub-blocks

– Achieve consistency– Better approaches– Changes in policy

Option value = the value of an unforeseen change


An Option is

The right but not the obligation to take an action– Action = Use a new design– If new is better than old, use new;– Otherwise, keep the old.


An Aspect changes the “Net Option Value” (NOV) of the codebase

Without Aspect:Net Option Value (NOV) =

Change Benefit – 100 Change Costs With Aspect:

Net Option Value (NOV) =Change Benefit – 1 Change Cost

If Change Benefit > 100 Change CostsNOV = 99 Change Costs

As Change Benefit decreases, NOV goes down:If Change Benefit= 0, NOV = 0


To justify an Aspect (to your CFO)

You need to argue:– Change Benefit is high

» consistency, better approaches, policy

– Difference in Change Costs is high» “99 changes vs. 1”

– Cost of refactoring is low» “You are going to have to refactor anyway, so do it right… .”

The NOV approach attempts to quantify the elements of this argument– Option value resides in “Change Benefit”– This is the bleeding edge of our science


What is this elusive property that gives rise to option value?

Where does it arise?

Can we measure it?


Global Design Rules v.1

Version 1.0Version 1.2

Version 1.5Version 1.8

Low Medium Zero High

Measuring Option Potential Successive, improving versions are evidence of option

potential being realized over time—after the fact Designers see option potential before the fact What do they see?


Major challenge in research and practice right now

Science may not be able to deliver tools to measure ex ante option potential reliably

But ex ante estimates are what’s needed


/Option potential is like dark matter in the universe

Scientists can measure its effects but we can’t measure “it”

“Wizards” can perceive /option potential– But wizards don’t talk to scientists!

Thus we lack ways to measure /option valuescientifically– It is a “research frontier”


/Option potential at work—Matlab programming contest


/Option potential at work—Transaction Processors and CPUs

1

10

100

1,000

10,000

100,000

1,000,000

10,000,000

Dec-91 Dec-93 Dec-95 Dec-97 Dec-99 Dec-01 Dec-03 Dec-05

System Availability Date

tpm

C a

nd

Sp

ec C

Int

Key Pfister TPC-C v.3 TPC-C v.5 Spec CPU 95 Spec CPU 2000

Architectural experimentation

Component experimentation


Sources of /option value Physics—

– Moore’s Law (dynamics of miniaturization) applies to MOSFET circuits and systems (Mead and Conway)

– Power and heat systems vs. logic systems (Dan Whitney)

User innovation– Users’ discovery of their own needs

– “Killer apps”

Architecture– Experimenting with different relationships among

components


But you might not want to tell your CFO that …

High option value makes designs unmanageable …

Call your venture capitalist, instead!


High option potential induces entry => industry structure change

Other Chassis/Powertrain

Other Interior

Other MultipleElectrical/Electronics

Other Misc.

Toyota Nissan DaimlerChrysler

Honda GM Ford VW

Oth

ers

Johnson Controls

Magna

Eaton

ServicesFirst Data



Operating Systems



Autos Computers


High ’s are what make designs unmanageable

“Unmanageable” = Cannot be confined in a single company or supply chain

There are modular design architectures that are very manageable– Example: Toyota Production System (TPS)

If System/360 had had low /option potential, IBM would be like Toyota

It would be a different world…


Recapping the argument Designs create value

– Value operates like a force in the economy– Changes the structure of industries

Designs have architectures– Modularity and /Option value are the key economic

properties of a design architecture– Modularity + High /Option value => Unmanageable

» Why computers are not like autos» Why IBM is not Toyota

We have NOT talked about—

How do you capture value in a modular, high-design?


How can you capture value?

Architectural/Technological question:– Where/when does modularization stop?

Business/Strategic question: – How do you make money in a modular, high-

world?


Where does modularization stop?

Strojwas (2005)

Semiconductor Industry

Top 10 Firms:

1994 and 2004


How do you make money in a high-world?

Old paradigm

– “Plunge in”

– “Get lucky”

– “Watch out for Microsoft”

– “Get bought by HP”


A new paradigm…– Expect a cluster– Use M&A to be the “lead firm” in some slice(s)

of the stack– Use design architecture to reduce your

“footprint” in each slice => high ROIC– Use open source methods to clone your

complementor’s products => price and innovation discipline

How do you make money in a high- world?


Our research strategy—Look for

Stable patterns of behavior involving several actors operating within a consistent framework of ex ante incentives and ex post rewards

==> Equilibria of linked games with self-confirming beliefs (Game theory)


How a “stable pattern” works

Anticipation of $$$ (visions of IPOs) Lots of investment Lots of design searches Best designs “win” Fast design evolution => innovation Lots of real $$$ (an actual IPO)

“Rational expectations equilibrium”


Three Stable Patterns

“Blind” Competition– PCs in the early 1980s

High ROIC on a Small Footprint – Sun vs. Apollo– Dell vs. Compaq (and HP and …)

Lead Firm Competition– Monopoly—MSFT– Mergers & Acquisitions—Cisco, Intel …


“Blind” Competition

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

All Zeros (1, 0, 0)

All Ventures (1, 0, 0) All Fighters (1, 0, 0)

ESS (1/8, 3/8, 4/8))

Ventures do best

Fighters do best

Zeros do best

x A

B x


“Blind” Competition


“Footprint” Competition—Apollo1 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 34 35 36 37 38 39 40 41 42

1 O Processor chip—CPU2 Outsourced—Motorola 680x0 Key:3 O Floating Point Accelerator x= transfer of material or information from column4 Outsourced task to row task;5 O Memory chips DRAMs, ROM T= transaction: sale of good by column owner to row6 Outsourced—Commodity owner;7 O Storage—Disk Drives O= outsourced task blocks;8 Outsourced D= downstream or complementary task blocks;9 O Storage—Tape Drive highly interdependent task blocks with many iterations

10 Outsourced and high within-block mundane transaction costs;

11 O Printed circuit boards Apollo's footprint (tasks performed inhouse).

12 Outsourced—Commodity13 O Display Monitor14 Outsourced15 O Keyboard, Cabinet, Fans16 Outsourced

17 x x x x x x x x x x x x x x x x Aegis proprietary18 Inhouse Operating System19 Design20 x x x x x x x x x x x x x x x x OS DOMAIN proprietary21 Network Network Architecture2223 x x x x x x x x x x x x x x x x Hardware Design24 DN series = 3-4 boards incl.

25 Hardware IO and Display controllers,26 Power supply27 T T T T T T T T T T T T T T T T x x x x Purchase Components28 Component Test x x x x x29 Kits x x x x x x Inhouse30 Board stuff and Solder x x x x x x Manu-31 Test Boards x x x x x x facturing32 Board Assembly x x x x x x33 System Assembly x x x x x x34 System Test x x x x x x35 Quality Assurance x x x x x x36 Consolidate and Ship x x x x x x

37 x x x x x x D38 x x x x x x Many Software Applications D39 x x x x x x D40 x x x x x x x x x x T D41 x x x x x x x x x x Many OEMs T D42 x x x x x x x x x x T D

Keeps Design Control


Then Sun came along…Apollo Computer

Aegis proprietaryInhouse Operating SystemDesignOS DOMAIN proprietary

Network Network Architecture

Hardware DesignDN series = 3-4 boards incl.

Hardware IO and Display controllers,Power supply

x x x x Purchase ComponentsComponent Test x x x x x

Kits x x x x x x InhouseBoard stuff and Solder x x x x x x Manu-

Test Boards x x x x x x facturingBoard Assembly x x x x x x

System Assembly x x x x x xSystem Test x x x x x x

Quality Assurance x x x x x xConsolidate and Ship x x x x x x

And did even less!

How?

x x x x x Customize Unixx x x x x Inhouse Proprietary MMU

x x x x x Design Internal busx x Single Board Layout

T T T T x x x x Purchase ComponentsComponent Test x x x x x O

Kits x x x x T Manu-Board stuff and Solder x x x x x O facturing

Test Boards x x x x TBoard Assembly x x x x x

System Assembly x x x x xSystem Test x x x x x

Quality Assurance x x x x xConsolidate and Ship x x x x x


Then Sun came along…Apollo Computer

Aegis proprietaryInhouse Operating SystemDesignOS DOMAIN proprietary

Network Network Architecture

Hardware DesignDN series = 3-4 boards incl.

Hardware IO and Display controllers,Power supply

x x x x Purchase ComponentsComponent Test x x x x x

Kits x x x x x x InhouseBoard stuff and Solder x x x x x x Manu-

Test Boards x x x x x x facturingBoard Assembly x x x x x x

System Assembly x x x x x xSystem Test x x x x x x

Quality Assurance x x x x x xConsolidate and Ship x x x x x x

x x x x x Customize Unixx x x x x Inhouse Proprietary MMU

x x x x x Design Internal busx x Single Board Layout

T T T T x x x x Purchase ComponentsComponent Test x x x x x O

Kits x x x x T Manu-Board stuff and Solder x x x x x O facturing

Test Boards x x x x TBoard Assembly x x x x x

System Assembly x x x x xSystem Test x x x x x

Quality Assurance x x x x xConsolidate and Ship x x x x x

Design Architecture for high performance with a small footprint

Public Standards for outsourcing

And did even less!

How?


Result: ROIC advantage to SunAverage over 16 Quarters: Apollo Sun

Computer MicrosystemsInvested Capital Ratios (Annualized)Net Working Capital/ Sales (%) 29% 15% Low is goodEnding Net PPE / Sales (%) 24% 13% Low is goodInvested Capital/Sales (%) 57% 31% Low is good

ProfitabilityNet Income/Sales 0% 6% High is good

ROICROIC (excl Cash, Annualized) 2% 20% High is good

Sun used its ROIC advantage to drive Apollo out of the market

Apollo was acquired by HP in 1989


Compaq vs. Dell Dell did to Compaq what Sun did to Apollo …

Dell created an equally good machine, and Used design architecture to reduce its footprint in

production, logistics and distribution costs– Negative Net Working Capital– Direct sales, no dealers

Result = Higher ROIC


Higher ROIC always wins!1997 Compaq Dell

Computer ComputerInvested Capital Ratios (Annualized)Net Working Capital/ Sales (%) -2% -5% Low is goodEnding Net PPE / Sales (%) 8% 3% Low is goodInvested Capital/Sales (%) 8% -2% Low is good

ProfitabilityNet Income/Sales 8% 7% High is good

ROICROIC (excl Cash, Annualized) 101% -287% !!!

Dell started cutting prices; Compaq struggled, but in the end had to exit.

Compaq was acquired by HP in 2002


Competition and Beliefs

“Blind” competitors – don’t know others exist

“Footprint” competitors – Don’t expect to influence others—just compete

“Lead firms”– Must influence the beliefs of their competitors– FUD — “Fear, uncertainty and doubt” – Others cannot be blind!


“Lead Firm” Competition

Monopolist needs to deter all potential entrants with threats of price war– Very fragile equilibrium

– Potentially expensive to create “enough” FUD

M&A Lead Firm does not try to deter all entry in the design space– Expects to buy most successful entrants ex post

– More robust equilibrium

– Maybe more advantageous, when you count the cost of FUD


Thank you!

Slide 1 © Carliss Y. Baldwin 2006 The Power of Modularity: The Financial Consequences of Computer and Code Architecture Carliss Y. Baldwin Harvard Business.

Documents

transit slide

answer slide

talk slide

computer industry slide

modular architecture

measuring modularity

mass market slide

largescale chips slide