What Computer What Computer Architects Architects Really DoReally Do
Bob ColwellBob ColwellE-M Talk ISCA ‘05E-M Talk ISCA ‘05
“Or should.”“Or should.”
2
2
But First, I’d Like to ThankBut First, I’d Like to Thank
Ron Hoelzeman (Pitt), Doug Jensen, Dan Ron Hoelzeman (Pitt), Doug Jensen, Dan Siewiorek (CMU), George Cox, Kevin Kahn Siewiorek (CMU), George Cox, Kevin Kahn
Paul Rodman, Dave Papworth, Rich Lethin, Paul Rodman, Dave Papworth, Rich Lethin, Josh Fisher & amazing Multiflow teamJosh Fisher & amazing Multiflow team
Incredible P6 team, esp. Randy Steck, Glenn Incredible P6 team, esp. Randy Steck, Glenn Hinton, Mike Fetterman, Andy Glew, Dave Hinton, Mike Fetterman, Andy Glew, Dave Papworth, Gurbir SinghPapworth, Gurbir Singh
My parents, Ellen, Kelly, Ken, KristenMy parents, Ellen, Kelly, Ken, KristenJoe MalingowskiJoe MalingowskiYale Patt, Wen-mei Hwu, Guri Sohi, Tom Yale Patt, Wen-mei Hwu, Guri Sohi, Tom
Conte, Computer Arch communityConte, Computer Arch community
4
4
A Misconception About What A Misconception About What Computer Architects DoComputer Architects DoEE Times May 23, 2005EE Times May 23, 2005
““Is the day of the architect over?”Is the day of the architect over?”““Microprocessor architects managed to re-Microprocessor architects managed to re-create almost the whole history of the create almost the whole history of the mainframe computer industry…they used all mainframe computer industry…they used all the tricks, from microprogramming and the tricks, from microprogramming and stripped-down pipelines with load-store stripped-down pipelines with load-store architectures to speculative execution architectures to speculative execution and branch prediction. and branch prediction. Best of all, Best of all, hardly anyone was unkind enough to hardly anyone was unkind enough to comment that all this ground had been comment that all this ground had been covered already, just at a lesser level covered already, just at a lesser level of integrationof integration.”.”
5
5
What would we have done…What would we have done…
Had we been born 300 years agoHad we been born 300 years agoSame IQ’s but no computers, no electronicsSame IQ’s but no computers, no electronics
Power source = waterwheels and oxenPower source = waterwheels and oxen
Same brains as today but different challengesSame brains as today but different challenges
Likewise with computer pioneersLikewise with computer pioneersIssue isn’t “why were they so innovative & why Issue isn’t “why were they so innovative & why
aren’t we”aren’t we”
They did what we do: whatever is necessaryThey did what we do: whatever is necessaryThose things change over timeThose things change over time
6
6
Intellectual Giant TheoryIntellectual Giant Theory
Intellectual giants did walk the earth in Intellectual giants did walk the earth in the ’60’sthe ’60’s– Eckert, Mauchly, von N, Conway, Cocke, Brooks, Eckert, Mauchly, von N, Conway, Cocke, Brooks,
Flynn, Tomasulo…Flynn, Tomasulo…– we should honor pioneering contributionswe should honor pioneering contributions– but today’s designers are not leeches living off but today’s designers are not leeches living off
that legacythat legacy
Intellectual giants did walk the earth in Intellectual giants did walk the earth in the ’60’s. the ’60’s. They still doThey still do..
7
7
Design today: more complexDesign today: more complex
1960’s complexity1960’s complexity– Poor tools, interaction of Poor tools, interaction of
electronics, packaging & ISAelectronics, packaging & ISA
Today’s complexityToday’s complexity– Today’s complexity from speed, Today’s complexity from speed,
hyper-aggressive uArch’s, power hyper-aggressive uArch’s, power limits, SW compatibility, number of limits, SW compatibility, number of usage models, usage models, CPU-MesmerizationCPU-Mesmerization
CPU
system
CPU-Mesmerization. Root cause? Profits.
10
10
Design today: scarierDesign today: scarier
No-recalls much harder than design-No-recalls much harder than design-for-minimal-field-servicefor-minimal-field-service– Pioneers designed for 1,000 usersPioneers designed for 1,000 users
Design errors? Charge ‘em for service callsDesign errors? Charge ‘em for service calls
– Today we design for 1,000,000,000 usersToday we design for 1,000,000,000 usersDesign errors? Pray…Design errors? Pray…
11
11
What architects really doWhat architects really do
Insidious error: thinking architects design Insidious error: thinking architects design instruction sets & uArch mechanismsinstruction sets & uArch mechanisms
We have, and do, but that misses the pointWe have, and do, but that misses the point– Architects start out as generals, moonlight as Special Architects start out as generals, moonlight as Special
ForcesForces
– Range freely, identify needs, apply appropriate forceRange freely, identify needs, apply appropriate force
– Ensure that biggest risks are attacked firstEnsure that biggest risks are attacked first
– Make sure project goals are clear & focussedMake sure project goals are clear & focussed
– Seek odd viewing angles to drive out problemsSeek odd viewing angles to drive out problems
– Supply judgment calls where data is lackingSupply judgment calls where data is lacking
And judgment as to when data should be And judgment as to when data should be collectedcollected
12
12
Recent arch historyRecent arch history
– For past two decades an architect’s point of For past two decades an architect’s point of highest leverage has been microarchitecturehighest leverage has been microarchitecture
Re-use what worksRe-use what works
– Pipelining, caches, shared buses, superscalar
Invent where necessaryInvent where necessary
– Microdataflow/OOO, trace caches, speculative branch predictors, cache coherency
With some major ISA work on RISCsWith some major ISA work on RISCs
– But this is changing. But this is changing. Right now.Right now.
CPU architects must evolve into system designersCPU architects must evolve into system designers
13
13
System Designers todo listSystem Designers todo list
““Whatever needs doing” has become…Whatever needs doing” has become…1.1. Products, not CPUsProducts, not CPUs
2.2. Power-constrained system design Power-constrained system design
3.3. Multicore (gotta pay the bills, too)Multicore (gotta pay the bills, too)
4.4. Reliable systems from unreliable Reliable systems from unreliable componentscomponents
No longer “what I’d like to sell you” but No longer “what I’d like to sell you” but designing what buyers wantdesigning what buyers want
14
14
Products Not CPUsProducts Not CPUs
“PC era” Ubiq. comp.
2004
Time
$$$
iPOD“killer apps” + PC’s
Cell phonesRay tracing
Portable computing
“Speed at any price” “What’s in it for me?” -buyer
“killer products”
15
15
Power-constrained designPower-constrained design
““fast as possible at max power” will yield to fast as possible at max power” will yield to “fast enough, no faster”“fast enough, no faster”
– Lesson from the embedded spaceLesson from the embedded space
Thermal variability vs. guaranteed real-timeThermal variability vs. guaranteed real-time– Throw in wireless links for good measureThrow in wireless links for good measure
Battery life, not just cooling costBattery life, not just cooling costGlobal warming, energy crisis loomsGlobal warming, energy crisis looms
– It ain’t just cars and oil pricesIt ain’t just cars and oil prices
– Be synchronized to public taste or loseBe synchronized to public taste or lose
16
16
MulticoreMulticore
““Here I come, ready or not…”Here I come, ready or not…”We can build ‘em. Can we…We can build ‘em. Can we…
– Compile to them?Compile to them?
– Feed them? (bandwidth)Feed them? (bandwidth)
– Cool them? (power)Cool them? (power)
– Write apps for them?Write apps for them?
Clear and present challengeClear and present challenge– There are pots of gold associated w/ thisThere are pots of gold associated w/ this
17
17
Reliable Systems from Reliable Systems from Unreliable ComponentsUnreliable ComponentsN-mod redundancy too expensiveN-mod redundancy too expensiveTransient errors, manufacturing Transient errors, manufacturing
defects, design errorsdefects, design errors– Must survive them allMust survive them all
– Solution can’t drive power upSolution can’t drive power up
– Still want guaranteed performance for real-timeStill want guaranteed performance for real-time
Intel’s Shekhar Borkar says we have at Intel’s Shekhar Borkar says we have at most 10 years to figure this outmost 10 years to figure this out
18
18
AdjurationsAdjurations
Computer revolution is only getting Computer revolution is only getting startedstarted
Role of architects is changingRole of architects is changing– If we don’t do it who will?If we don’t do it who will?
Your grandchildren will thank youYour grandchildren will thank you– And wonder if they’re as smart as you were And wonder if they’re as smart as you were
Top Related