From Theory to Practice - FBE - TitleFrame · The Intel ® Centrino. TM. Mobile technologyMobile technology. Announcing . Intel® Centrino™ mobile technology. Intel has expanded

The IntelThe Intel®® PentiumPentium®® M processorM processorPowerPower--Awareness Story Awareness Story

From Theory to PracticeFrom Theory to Practice

Ronny RonenRonny RonenSenior Principal EngineerSenior Principal EngineerDirector of Architecture ResearchDirector of Architecture ResearchIntel Labs Intel Labs -- HaifaHaifa

Intel CorporationIntel Corporation

Technion EE,Technion EE,Haifa,Haifa,June 2, 2003June 2, 2003

Page 2All dates, plans, and features are preliminary and subject to change without notice

Based on…Based on…

The IntelThe Intel®® PentiumPentium®® M Processor:M Processor:Microarchitecture and PerformanceMicroarchitecture and Performance

ByBySimcha Gochman, Ronny Ronen, Ittai Anati,Simcha Gochman, Ronny Ronen, Ittai Anati,Ariel Berkovits, Tsvika Kurts, Alon Naveh, Ariel Berkovits, Tsvika Kurts, Alon Naveh,

Ali Saeed, Zeev Sperber, Robert C. ValentineAli Saeed, Zeev Sperber, Robert C. Valentine

Intel Technology Journal Q2/2003Intel Technology Journal Q2/2003http://http://developer.intel.com/technology/itjdeveloper.intel.com/technology/itj//


IDC IDC –– Israel Development CenterIsrael Development CenterLocated on Israel's Mediterranean coast, Haifa is the home of Intel's Israel Development Center (IDC).

IDC was established in 1974, and is Intel's first development center outside the US. The center is a multi-disciplinary team, with more than 1000 employees.

Many of Intel's leading products were developed and originated at IDC.

IDC's employees are currently working on Intel's future microprocessors, CAD tools, advanced networking components and software technologies.

The Baha`i ShrineHaifa most known attraction


The IntelThe Intel®® CentrinoCentrinoTMTM Mobile technologyMobile technologyAnnouncing Intel® Centrino™ mobile technology. Intel has expanded its history of innovation with new notebook capabilities designed specifically for the mobile world. Now you can work, play and connect without wires. And choose from a whole new generation of thin, light notebooks designed to enable extended battery life.

This new innovative technology enables: Integrated wireless LAN capabilityBreakthrough mobile performance Extended battery lifeThinner, lighter designs

Intel® Intel® Pro/Wireless Pro/Wireless 2100 Network 2100 Network Connection Connection

ICH4-MICH4ICH4--MM

Intel® 855ChipsetFamily

IntelIntel®® 855855ChipsetChipsetFamilyFamily

IntelIntel®®

PentiumPentium®® M M ProcessorProcessor


AgendaAgenda

The TheoryThe Theory–– Power, EnergyPower, Energy–– Power AwarenessPower Awareness

The PracticeThe Practice–– The Intel Pentium M processor microThe Intel Pentium M processor micro--architecturearchitecture

The PerformanceThe Performance


Power and the digital world…Power and the digital world…

IN OUT

Power is consumed:Power is consumed:–– When capacitance is charged and discharged.When capacitance is charged and discharged.–– A charged cap is a logical ‘1’, a discharged cap is ‘0’.A charged cap is a logical ‘1’, a discharged cap is ‘0’.

The capacitance can be the gates of other transistors or wiresThe capacitance can be the gates of other transistors or wires(busses and long interconnects).(busses and long interconnects).

00 110011

E=E=11//22CVCV22


Power and the digital world (2)…Power and the digital world (2)…Secondary effects like leakage and shortSecondary effects like leakage and short--circuit current circuit current are increasing with advanced process technologies.are increasing with advanced process technologies.

Depends heavily on operating voltage and temperatureDepends heavily on operating voltage and temperatureLeakage is growing dramaticallyLeakage is growing dramatically–– Reaching 20% in current process technology, and growing…Reaching 20% in current process technology, and growing…

IN OUT

1/21/2

IN OUT

00 11

LeakageLeakage(sub(sub--threshold)threshold)

ShortShort--circuitcircuit


Power & EnergyPower & Energy

Total energyTotal energy–– Total of all switch energy and leakage wasteTotal of all switch energy and leakage waste–– Measured in either in joules or watt x hourMeasured in either in joules or watt x hour

“Energy per task”“Energy per task”Lower Energy per task meansLower Energy per task means–– Longer battery life.Longer battery life.–– Lower electric billsLower electric bills

Power = energy / time Power = energy / time = = ααCVCV22f f (+ leakage power)(+ leakage power)

((αα: activity, C: capacitance, V: voltage, f: frequency): activity, C: capacitance, V: voltage, f: frequency)–– Measured in wattsMeasured in watts


Power & EnergyPower & Energy

Average powerAverage power–– Total Total energy / energy / TotalTotal timetime–– Including lowIncluding low--activity and idleactivity and idle--timetime

Peak powerPeak power–– Higher power Higher power higher current.higher current.–– Higher power Higher power higher temperature.higher temperature.

–– Cannot exceed the thermal constrains.Cannot exceed the thermal constrains.

Typical figures (leading edge processors)Typical figures (leading edge processors)–– Average power: 1WAverage power: 1W--3W3W–– Peak power: 20WPeak power: 20W--100W100W


Power DensityPower Density

Think of watts/cmThink of watts/cm22..Denser power is harder to cool.Denser power is harder to cool.Complex algorithms lead to denser power:Complex algorithms lead to denser power:

–– Dense random logic.Dense random logic.–– Timing pressure leads to faster/bigger/powerTiming pressure leads to faster/bigger/power--hungrier hungrier

gates.gates.

Increased every process technology Increased every process technology generation (higher power @ smaller die size).generation (higher power @ smaller die size).


Power Density and ThermalPower Density and Thermal

Thermal Map (EDO System)2,3

(lowest) blue, green, yellow, orange, purple, white (highest)Power Density (Simulated)1

Color codes: (lowest) black, red, orange yellow, white (highest)

Pentium M Processor power density example1 Source : Intel® Pentium® M Processor Power Estimation,

Budgeting, Optimization, and ValidationDani Genossar, Nachum Shamir, ITJ Q2/2003

2 Source: Dani Genossar, Nachum Shamir, Intel 20033 The L2 left portion thermal map blank due to measurements limitations.


Voltage, Power, FrequencyVoltage, Power, Frequency

Transistor switches faster at higher voltageTransistor switches faster at higher voltagehigher voltage enables higher frequencyhigher voltage enables higher frequency

Maximum frequency grows about linearly with voltage.Maximum frequency grows about linearly with voltage.…Within a given voltage range Vmin…Within a given voltage range Vmin--Vmax.Vmax.–– V < VminV < Vmin

transistors won’t switch.transistors won’t switch.–– V > VmaxV > Vmax

the device may burn.the device may burn.

“The cube law”:“The cube law”:P P ≈≈ kVkV33(or ~1%V = 3%P)(or ~1%V = 3%P)ImplicationsImplications–– Can save energy/power whenCan save energy/power when

Performance is not a factor 0

100

200

300

400

500

600

700

800

900

1000

0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9

Fequency(Mhz)Power (mWatt)

XScale processor freq. & power vs. voltage *

Performance is not a factor* Source: Intel Corp. (http://developer.intel.com)* Source: Intel Corp. (http://developer.intel.com)


Bean Counting*Bean Counting*High VoltageHigh Voltage

1.1. Number of transistorsNumber of transistors 80 M80 M2.2. Die areaDie area 80 mm80 mm22

3.3. Operation VoltageOperation Voltage 1.48V1.48V4.4. Operation frequencyOperation frequency 1.6 GHz1.6 GHz5.5. Energy per switch per transistorEnergy per switch per transistor 0.85 fJ0.85 fJ6.6. Power per transistor (#4Power per transistor (#4xx#5)#5) 1.4 uW1.4 uW7.7. Activity factorActivity factor 20%20%8.8. Energy per cycle per chip (#1x#5x#7)Energy per cycle per chip (#1x#5x#7) 14 nJ14 nJ9.9. Power (#4x#8)Power (#4x#8) 22 W22 W10.10. Power Density (#9/#2)Power Density (#9/#2) 27 W/cm27 W/cm22

* These numbers are representative only and * These numbers are representative only and do not intend to reflect any existing devicedo not intend to reflect any existing deviceRepresentative NumbersHigh VoltageHigh Voltage Low VoltageLow Voltage

1.1. Number of transistorsNumber of transistors 80 M80 M 80 M80 M2.2. Die areaDie area 80 mm80 mm22 80 mm80 mm2 2

3.3. Operation VoltageOperation Voltage 1.48V1.48V 0.96V0.96V4.4. Operation frequencyOperation frequency 1.6 GHz1.6 GHz 0.6 GHz0.6 GHz5.5. Energy per switch per transistorEnergy per switch per transistor 0.85 fJ0.85 fJ 0.36 0.36 fJfJ6.6. Power per transistor (#4Power per transistor (#4xx#5)#5) 1.4 uW1.4 uW 0.21 0.21 uWuW7.7. Activity factorActivity factor 20%20% 10%10%8.8. Energy per cycle per chip (#1x#5x#7)Energy per cycle per chip (#1x#5x#7) 14 nJ14 nJ 2.9 2.9 nJnJ9.9. Power (#4x#8)Power (#4x#8) 22 W22 W 1.7 W1.7 W10.10. Power Density (#9/#2)Power Density (#9/#2) 27 W/cm27 W/cm22 2.1 W/cm2.1 W/cm22


Mobile Platform Goals & ChallengesMobile Platform Goals & Challenges

Goal: Higher performanceGoal: Higher performance–– Challenge:Challenge:

How much power one can afford to spend in How much power one can afford to spend in order to implement a performance feature?order to implement a performance feature?

Goal: Longer Battery lifeGoal: Longer Battery life–– Challenge:Challenge:

How to balance the design for maximum How to balance the design for maximum performance and extended battery life?performance and extended battery life?


Higher Performance vs. Longer Battery LifeHigher Performance vs. Longer Battery Life

Processor Processor average poweraverage power is <10% of platformis <10% of platformLCD and other components consume much moreLCD and other components consume much more

Even ideal processor canEven ideal processor canextend battery life by 11%extend battery life by 11%at most!at most!

Decision:Decision:–– Optimize for performance when ActiveOptimize for performance when Active–– Optimize for battery life when idleOptimize for battery life when idle

CaveatCaveat–– This observation is Pentium M specific.This observation is Pentium M specific.

May not hold as such in the future!

Display(panel + inverter)

33%

CPU10%

Power Supply10%

Intel® MCH9%

Misc.8%

GFX8%

HDD8%

CLK5%

Intel® ICH3%

DVD2%

LAN2%

Fan2%

May not hold as such in the future!Source: 2004 Extended Battery Life Technologies,Don J Nguyen, Intel Developer Forum, Spring 2003


Optimize for PerformanceOptimize for Performance““Maximize performance at given thermal constraints”Maximize performance at given thermal constraints”

Approximated by:Approximated by:Maximizing performance at given Power budgetMaximizing performance at given Power budgetThe test:The test:“A micro“A micro--architectural feature that gains performance architectural feature that gains performance or saves power should be better than simply using or saves power should be better than simply using voltage/frequency scaling”voltage/frequency scaling”That is:That is:

f f ≈≈ K*V K*V Power = Power = αα*C*V*C*V 22*f *f ≈≈ αα*C*f*C*f 33

∆∆power/Power = ((f+power/Power = ((f+∆∆f)f)3 3 –– ff 3 3 )/ f)/ f 3 3 ≈≈ 33∆∆f / ff / fPerf = IPCPerf = IPC*f*f

The right Performance/Power tradeoff:The right Performance/Power tradeoff:1%1% more performance in less than more performance in less than 33%% Power Power –– a gain!a gain!


Optimize for Battery LifeOptimize for Battery Life

““Minimize Energy per Task”Minimize Energy per Task”

Should address both active and idle energyShould address both active and idle energy

The active energy tradeoff:The active energy tradeoff:EnergyEnergyactiveactive = Power= Poweractiveactive * Time* Timeactiveactive

ororEnergyEnergyactiveactive ≅≅ PowerPoweractiveactive / Perf/ Perfactiveactive

The right Performance/Power tradeoff:The right Performance/Power tradeoff:1%1% more performance in less than more performance in less than 11%% Power Power –– a gain!a gain!


-60%

-40%

-20%

0%

20%

40%

60%

80%

100%

-30%

-27%

-24%

-21%

-18%

-15%

-12% -9% -6% -3% 0% 3% 6% 9% 12%

15%

18%

21%

24%

27%

30%

<= Performance loss | Performance gain =>

Energy LossConstrained Perf Loss

Wrong trade-off zone

Energy LossConstrained Perf

Gain

Energy GainConstrained Perf

Loss


Gain

Constrained-Performance

Breakeven line

Energy Breakeven

line

<= P

ower

Gai

n |

Po

wer

Los

s =>

-60%

-40%

-20%

0%

20%

40%

60%

80%

100%

-30%

-27%

-24%

-21%

-18%

-15%

-12% -9% -6% -3% 0% 3% 6% 9% 12%

15%

18%

21%

24%

27%

30%

<= Performance loss | Performance gain =>

Energy LossConstrained Perf Loss

Wrong trade-off zone

Energy LossConstrained Perf

Gain


Loss


Gain

Constrained-Performance

Breakeven line

Energy Breakeven

line

<= P

ower

Gai

n |

Po

wer

Los

s =>

Performance loss | Performance gain

Pow

er G

ain

| Pow

er L

oss

Putting it all together:Putting it all together:The Pentium M processor Approach The Pentium M processor Approach


Back to Basics: “Less is More” Back to Basics: “Less is More”

Less instructions per taskLess instructions per taskLess microLess micro--ops per instructionops per instructionLess transistor switches per microLess transistor switches per micro--opopLess energy per transistor switch.Less energy per transistor switch.


““Less is More” Less is More” in the Pentium M processorin the Pentium M processor

Less instructions per taskLess instructions per task–– Advanced branch prediction, Advanced branch prediction, SSE instructionsSSE instructions

Less microLess micro--ops per instructionops per instruction–– MicroMicro--ops fusionops fusion–– Dedicated stack engineDedicated stack engine

Less transistor switches per microLess transistor switches per micro--opop–– 1MB “power managed” L2 cache1MB “power managed” L2 cache–– the Pentium M processor busthe Pentium M processor bus–– Various lowerVarious lower--level optimizationslevel optimizations–– Advanced clock gatingAdvanced clock gating

Less energy per transistor switchLess energy per transistor switch–– Enhanced Intel® SpeedStep® technology Enhanced Intel® SpeedStep® technology

Power-awareness top to bottomPower-awareness top to bottom


Advanced Branch PredictionAdvanced Branch PredictionTypical today’s processors spend about Typical today’s processors spend about 11//44--11//33 of of the time in branch misprediction recoverythe time in branch misprediction recovery

–– Losing both performance and energy!Losing both performance and energy!

The Pentium M processor employs bestThe Pentium M processor employs best--inin--class class branch predictionbranch prediction

–– Uses Hybrid BiUses Hybrid Bi--Modal/Global branch predictorModal/Global branch predictor–– Loop detectorLoop detector–– Indirect branch predictorIndirect branch predictor

Captures all standard program behaviors and new Captures all standard program behaviors and new programming paradigms:programming paradigms:

–– Loops of all iteration sizeLoops of all iteration size–– JITed and object oriented codeJITed and object oriented code

20% less mispredcitions 7% performance gain*20% less mispredcitions 7% performance gain** Relative to the Pentium M processor w/ traditional branch pred* Relative to the Pentium M processor w/ traditional branch predictor (simulated)ictor (simulated)


Dedicated Stack ManagerDedicated Stack ManagerIA32 instruction set provides explicit management IA32 instruction set provides explicit management of S/W stackof S/W stack

–– E.g. PUSH, POP, RET, CALLE.g. PUSH, POP, RET, CALL

Stack management operations are an overheadStack management operations are an overhead–– E.g. Stack pointer incrementE.g. Stack pointer increment–– Normally done via machines main execution pathNormally done via machines main execution path

The Pentium M processor employs sophisticated The Pentium M processor employs sophisticated H/W control for stack managementH/W control for stack management

–– Maintains/updates the stack pointer value in the decoderMaintains/updates the stack pointer value in the decoder–– Includes synchronization mechanismIncludes synchronization mechanism

Replaces a power hungry microReplaces a power hungry micro--op controlop control

H/W management instead of power hungry Micro-opsH/W management instead of power hungry Micro-ops


Dedicated Stack ManagerDedicated Stack Manager

ESP+4

Scheduler

ALU

32 Bit adder

Wide uOp

Achieving >5% of Micro-op reductionAchieving >5% of Micro-op reduction

+4DeltaESP

PUSH/POP/CALL/RET

Decoder


MicroMicro--op Fusion op Fusion –– Best of all WorldsBest of all Worlds

IA Instructions are typically broken into IA Instructions are typically broken into micromicro--opsops

–– Normally handled individuallyNormally handled individually

Pentium M processor employs MicroPentium M processor employs Micro--op op fusionfusion

–– Instructions with memory operands are fusedInstructions with memory operands are fused–– Single MicroSingle Micro--op during most of the instruction lifetimeop during most of the instruction lifetime–– Enhanced performance and power characteristicsEnhanced performance and power characteristics

MicroMicro--ops are split just in time for executionops are split just in time for execution–– Allows outAllows out--ofof--order, superorder, super--scalar executionscalar execution


MicroMicro--op Fusion op Fusion –– Best of all WorldsBest of all Worldsadd eax, dword ptr data

SchedulerLD

Cache ALU

OP

OP

Decoder

LD


MicroMicro--op Fusion op Fusion –– Best of all WorldsBest of all Worldsadd eax, dword ptr data

Decoder

Scheduler

Cache

LD + OP

LD + OP

Micro-op fusion enables effective

machine utilization

LD

Independent uOp OOO/Super-

scalar executionALUOP

Achieving >10% of Micro-op reductionAchieving >10% of Micro-op reduction


Enhanced SpeedStep™ TechnologyEnhanced SpeedStep™ TechnologyThe “Basic” SpeedStep™ Technology hadThe “Basic” SpeedStep™ Technology had–– 2 operating points 2 operating points –– NonNon--transparent switchtransparent switch

The “Enhanced” version providesThe “Enhanced” version provides–– Multi voltage/frequency operating points. The Pentium M processoMulti voltage/frequency operating points. The Pentium M processor r

1.6GHz operation ranges:1.6GHz operation ranges:–– From 600MHz @ 0.956VFrom 600MHz @ 0.956V–– To 1.6GHz @ 1.484V To 1.6GHz @ 1.484V

–– Transparent switchTransparent switch–– Frequent switchesFrequent switches

BenefitsBenefits–– Higher power efficiencyHigher power efficiency

2.7X lower frequency 2.7X lower frequency 2X performance loss 2X performance loss >2X energy gain>2X energy gain

–– Outstanding battery lifeOutstanding battery life–– Excellent thermal mgmt.Excellent thermal mgmt.

Voltage, Frequency, Power

0.0

0.4

0.8

1.2

1.6

2.0

2.4

2.8

3.2

3.6

4.0

0.8 1.0 1.2 1.4 1.6Voltage (Volt)

Freq

uenc

y

0

2

4

6

8

10

12

14

16

18

20

Typi

cal P

ower

Freq (GHz)Power (Watts)

2.7X2.7X

6.1X6.1X

Efficiency ratio = 2.3


Performance ResultsPerformance Results3 system3 system

–– Intel® Pentium® M Processor (1.6 GHz/600 MHz)Intel® Pentium® M Processor (1.6 GHz/600 MHz)–– Mobile Intel® Pentium® 4 ProcessorMobile Intel® Pentium® 4 Processor--M (2.4/1.2 GHz)M (2.4/1.2 GHz)–– Mobile Intel® Pentium® III ProcessorMobile Intel® Pentium® III Processor--M (1.2 GHz/800 MHz)M (1.2 GHz/800 MHz)

3 operation modes3 operation modes–– Always On (Max Frequency)Always On (Max Frequency)–– Portable/Laptop (Adaptive Frequency)Portable/Laptop (Adaptive Frequency)–– Maximum Battery (Min Frequency)Maximum Battery (Min Frequency)

3 benchmarks3 benchmarks–– Mobile Representative Office Productivity WorkloadMobile Representative Office Productivity Workload–– Internet Experience workload Internet Experience workload –– SPEC CPU 2000 V1.2SPEC CPU 2000 V1.2

Measured relative Performance and EfficiencyMeasured relative Performance and Efficiency–– Efficiency = “energy per task” = Performance/averageEfficiency = “energy per task” = Performance/average--power power


Performance ResultsPerformance Results

Always On (Max Frequency)Always On (Max Frequency) Portable/Laptop (Adaptive Frequency)Portable/Laptop (Adaptive Frequency)

Always onAlways on–– 2%2%--30% higher performance30% higher performance–– 2X2X--3X efficiency3X efficiency

Maximum BatteryMaximum Battery–– Lower performanceLower performance (15% or so)(15% or so)–– 4X4X--5X efficiency on all benchmarks

PortablePortable–– Enhanced SpeedStep Technology Enhanced SpeedStep Technology –– 3X 3X >4X >4X efficiency in officeefficiency in office

5X efficiency on all benchmarks Maximum Battery (Min Frequency)Maximum Battery (Min Frequency)


SummarySummary

The Intel Pentium M Processor The Intel Pentium M Processor

A key component of theA key component of theIntel CentrinoIntel CentrinoTMTM Mobile TechnologyMobile Technology

Exhibits Power awareness at all levelsExhibits Power awareness at all levels

Provides breakthrough mobile performanceProvides breakthrough mobile performanceand extended battery lifeand extended battery life


The EndThe End

From Theory to Practice - FBE - TitleFrame · The Intel ® Centrino. TM. Mobile technologyMobile technology. Announcing . Intel® Centrino™ mobile technology. Intel has expanded

Documents