Brian w bramlett intel corporation for mit media lab : center of bits and atoms : emergent engineering : 2002 october 16 engineering emergence a view of

brian w bramlettintel corporationfor mit media lab : center of bits and atoms : emergent engineering : 2002 october 16

engineering emergence

a view of complex system engineeringfrom integrated circuit design evolution


disclaimer:

THIS DOCUMENT AND RELATED MATERIALS AND INFORMATION ARE PROVIDED "AS IS" WITH NO WARRANTIES, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY IMPLIED WARRANTY OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, NON-INFRINGEMENT OF INTELLECTUAL PROPERTY RIGHTS, OR ANY WARRANTY OTHERWISE ARISING OUT OF ANY PROPOSAL, SPECIFICATION, OR SAMPLE. INTEL ASSUMES NO RESPONSIBILITY FOR ANY ERRORS CONTAINED IN THIS DOCUMENT AND HAS NO LIABILITIES OR OBLIGATIONS FOR ANY DAMAGES ARISING FROM OR IN CONNECTION WITH THE USE OF THIS DOCUMENT.

Intel and the Intel logo are trademarks or registered trademarks of Intel Corporation or its subsidiaries in the United States and other countries.

* Other names and brands may be claimed as the property of others.

Copyright © 2002 Intel Corporation.


overview: four target research areas

language commonality across domains+disciplines

limits of atomic system complexity

formalization of complexity engineering

design system implications(methodologies, tools, frameworks, infrastructures)


on language: semiotics and semantics

since these fields are only beginning to interact broadly across "traditional" disciplines, there is a great deal of semantic disconnect [and things that fill that gap…]

more critically in systems science when exploring the relationships among and emergent behavior of connected components in dynamic systems near chaotic boundaries

1: common semantics for complex, multi-disciplinary systems is a unique, arguably critical challenge


context: setting the stage

emergeto come into view

evolveto unroll or unfold; yet grow in complexity

convergeto come together: close, meetto direct toward a common center: concentrate, focus

convergence is a core concept in modern cpu design


integration: from the top

• Central Processing Unitvery semantically accurate(see «atomic system complexity»)

• integrated (circuits)bound or tightly coupledoften implies «non-modular» within the system

• complex integrationcouplings among abstractions become so profound, they

challenge their meaningcomponent boundaries blur beyond practical use

to characterize: how well-defined are the interfaces, how much mapping is required across abstract domains


*lsi design space: notable parameters

• volume-margin-yieldcost per manufactured part before NREs; ASPscoupled to die size

• structural complexityunique patterns, equivalence classes, «compressibility»

• performancehighly visible within the functional domainothers prominent due to cost of quality+reliability

emerging: metrics like MIPS/Watt, peak+average


*lsi: types of products

semiconductor memory, displays, ASICs, graphic controllers, CPUs, chipsets, interfaces (USB, FireWire*, WiFi*...), SOC products, microcontrollers…

[imagine a three-axis plot]

• low-volume, high complexity, high performance designs are not economically justified; breaks the “choose two of three” phenomenon

• mass-market performance cpu design is highly constrained, which dramatically restricts the design space due to coupling effects

* Other names and brands may be claimed as the property of others.


cpu design: unique systems issues

other vastly complex engineered artifacts are currently produced, but:

• many rely on complex feedstock or bulk properties• most are composed of less complex,

replaceable/serviceable components• few are as complex, reliable, and optimized at physical

limits of performance, with commodity volumes

[…a twist: prerequisite to evolving engineering disciplines]

so, to focus: highly optimized, closed complex systems


about: atomic system complexity

• a hallmark of this degree of coupling; strong local coupling or large statistical contribution to intertwined first-order effects

• intrinsically coupled design, ecologically «closed» systems• coupling approaches the limit of what can be distinguished as a

system of connected components; inseparable, insufferable• design abstraction becomes very difficult to manage mentally

and in the data models

• a useful view during the design process is a series of phase transitions in convergence toward design requirements: an emergent property

…and since you will need to get used to me saying it: "coupling, coupling, coupling…”


why/how: exploring atomic system limits

2: is there a practical or theoretical limit to complexity and reliability in an «atomic system» component?

• how much of this is due to a "single-minded" entity defining design, rather than what might happen with a collective/emergent design intent?

• other examples of complex systems that are more economic to replace than repair? a biological tactic?

• some theory already, but do any comparable engineered systems exist? any coming?

• does this imply that this form of engineering a specialization? subsumable? bootstrap/replaceable?

• …essential? insufficiently developed?


cpu design: major subsystems

• economy / business ecology• business model / fiscal + market analysis• cross-platform design strategy• platform design (a highly distributed function)

» Instruction-level architecture» micro-architecture and logic design» circuit (dynamic) and physical (structural) design• physical fabrication / manufacturing process and system• physical test, systems integration, back-end q+r

although there are equally critical complex inter- and intra-system coupling (clear up to organizational change theory's “absorptive capacity”), many of these aren't exclusive to IC design generally, nor mass-market CPU design specifically.

manufacturing systems for CPUs is highly specialized in and of itself, and has large, often blunt effects on the others. the top few layers are not as well-structured (with implications to evolution).

although very strong and increasingly complex coupling exists across all these domains, those emphasized above are the among the tightest subset.


jargon file: engineering process 101

what you want, what you can't have, and how to reasonably get there on time and in budget

requirements: domains of convergence, design “qualities”:• function, performance, power, reliability, manufacturability, cost,

schedule, risk…

cost (resources + time):• human resources, compute systems, intellectual property (including

licenses), time

tools:• abstractions within domains, marked by level of granularity and

language; these levels usually cross sets of domains

rules: (process)• moving efficiently toward convergence over time using abstraction

granularities to define convergence phases, rates, and transitions


coupling: characteristics

Scope (across abstractions, information: in the design system)• Many levels of formal abstraction/granularity; often one or more

languages and models each• incremental mapping from specification to implementation; information

to physics• Hard to span with precision [hint: good emergence target]

Scale (numeric and physical: everywhere)What is often thought of as complexity, but often isn’t.• a billion transistors, each connected to at least three others, [graph

layout shares issues]• «computationally intractable», «incomprehensibly large»


coupling: characteristics

Strength/Degree (across domains, physics: in the artifact)toward the limits of information physics and space-time physics, and where

they intertwine (thermodynamics meets information theory)approaching the level of bits and quanta, and a very short distance to qbits

coupling is strong, complex, atomically non-linearcomponent isolation (well-defined interfaces) costs more (noise, parasitics)unlike most «complex systems» where simple rules imply complex behavior,

the rules are not simple, either.

Consider:

microarchitecture performance + device/wire placement + em noise + power + process variation + physical limits


an abbreviated evolutionary history

early on, coupling (for design) was primarily human-scale:practical to do by small groups, manual work

Moore-type scaling (you knew that was coming) in efficiency enabled by (among others):

increasing use of design abstractiondomain-specific engineering techniquesimproved process controlrefined and extended automation

sticking points generally defined by coupling issues

optimization is still a central, competitive engineering focus and driven to local maximum, even if global constraints change the landscape

proportion of design with this emphasis drops, but often with net growth


comparison to familiar software systems

Inversion of approach to parallelism• parallelism tends to be aligned to a physical basis• sequential/serial elements map to logic

massive default low-level parallelism with explicit serialization

versus

typically serial design, with explicit thread-based parallelism or high-symmetry distributed processing getting design attention, and optimizing compilers

defect tolerance, cost, and performancewithout the redundancy/intelligence for device-level fault-

detection at manufacture: • every transistor on every die must effectively be testable

and «perfect»• very low design defect tolerance; efficient robustness

clearly, there are software systems which exhibit similar qualities to cpu design


making it all fit and fit together

Why decomposition? (partitioning, hierarchy)optimal fit to:• computation, communication, capacity/cache• Mind or machine or tools[looks like a classical system model]

Mapping [complex coupling makes this hard!]• Extraction, synthesis, and verification• Based on encapsulation of partition contents by interface/boundary• Across abstraction granularity, convergence domains

Common design system abstractions [architectural hierarchy]• design representation, query, transformation, measurement• state management• workflow management• role management


«…essential? insufficiently developed?»

for the purposes of this discussion, an assertion: yes and yes

the design system itself becomes (is) far more complex than the artifacts it produces

complex system design infrastructure requires high fiscal motivation not currently present

each wave of evolutionary improvement moves toward a local maximum around existing engineering methodologies

interactive design is still often single-cpu (even single-thread), and license bound; even graphite and pressed wood pulp

design-expertise IP is still costly, but little market justificationEDA is nowhere near the maturity/commodity of SW compilation;

plenty of room to grow


what's the emerging issue?(does this look familiar yet?)

• early attempts within EDA alone have fallen short; lack of incentive in existing business models

• current «open» design systems do not scale for required capacity, and lack modular integration required for evolution

• as other engineering disciplines evolve, this looks to become common

• de novo invention of specialized design systems for every new combination of disciplines introduces diversity without common abstraction

4: open architectures designed for emergent and cross-disciplinary engineering are needed; the design system itself seems a candidate for emergent engineering


complexity engineering: (re)defined …or “bramlett’s semantic gripe”

a frequent misnomer for emergent engineering

A discipline including partitioning a design (even dynamically and by evolutionary means) into various design approaches, such as traditional or emergent, by considerations such as requirement domains, effort/cost, and risk.

Another view:engineering how complexity is managed and manifested in a design instance and its design process as a method of optimization

Or, to appease the emergilentia:a formal specialization within system engineering


complexity: my own taxonomy

as if there weren’t enough:

• Intrinsicphysically or logically inherent

• ArtificialDue to design or construction of abstract model. This is

becoming more fluid, and can be subject to engineering

• PseudoAttributes that are misunderstood as complexity, like scaleA form of noise, basically.


simplistic design system evo scenario

how might emergent engineering be incorporated into existing disciplines and infrastructures (in vivo)?

1. traditional; extended + conventional reuse of standardized components

2. complexity engineering applied within the design model and automation space

3. complexity engineering with awareness of humans as part of design system

4. inclusion of emergent methods


directions in design system engineering

pattern languages and automation for query, transformation, mapping[often from graph/string domains; linguistic analog of…]

standardized, modular interfaces and visualization engines[aligned to human/machine relative efficiencies;biological pattern recognition/learning, machine-precise manipulation]

information design including information physics[self-assembling data visualization]

domain-agnostic architectures and frameworks[analogous to movement toward semantic web]

resource + model partitioning and allocation; complexity engineering[humans, computation/machines, design]

fine-grained workflow / state / design management[dynamic precision decoupling]

merging and convergence of engineering methods across disciplines

bootstrapping via formal systems and workflow analysis[reusing design tools on the design system itself; autocatalysis?]


questions: a step out from the formal scope

project planning:• how to staff and schedule for a set of goals or design intent• including direction of complex infrastructure development/investment

organizational composition, structure, and dynamics:• what roles and disciplines are needed? how do they relate/couple?• engineering for/to absorptive capacityeven within electrical engineering, augmented by computer science, the scope of

domain knowledge required is vast

on a larger scale, process / infrastructure to support design intent:• facilitating emergent design (expression of collective intent) evolved

from non-globally aware system components• what does design mean here as a manifestation of desire or intent?

and are these traditional, emergent, complex…?


postscript: complex design engineering ecology

design cost relative to manufactured margin is still small

[capital, nre, r+d cost/effort weighted in manufacturing]

design productivity driven by market opportunity cost/gain and capital investment/amortization

interesting to see what happens when cost of manufacturing and barriers to entry drop…

[hint: design engineering is coming up for commodification]

Documents

Brian w bramlett intel corporation for mit media lab : center of bits and atoms : emergent engineering : 2002 october 16 engineering emergence a view of