intel speedstep technology of the pentium m processorskadron/tacs/rotem_slides.pdf · intel®...
TRANSCRIPT
Analysis of The Enhanced Analysis of The Enhanced IntelIntel®® SpeedstepSpeedstep®® Technology Technology of the Pentiumof the Pentium®® M Processor M Processor
Efi Rotem, Efi Rotem, AviAvi MendelsonMendelson, Alon , Alon NavehNaveh, , MichaMicha MoffieMoffie
June 2004June 2004
MOBILETECHNOLOGY
TACS June 2004 - 2 -
OverviewOverviewMobile computers challengesMobile computers challenges
Energy and Average powerEnergy and Average powerThermal Design PowerThermal Design Power
IntelIntel®® PentiumPentium®® M power management M power management The experimentsThe experimentsTest resultsTest resultsConclusionsConclusions
TACS June 2004 - 3 -
OverviewOverviewMobile computers challengesMobile computers challenges
Energy and Average powerEnergy and Average powerThermal Design PowerThermal Design Power
Intel Pentium M power management Intel Pentium M power management The experimentsThe experimentsTest resultsTest resultsConclusionsConclusions
TACS June 2004 - 4 -
Power & Density increasesPower & Density increasesThink Watts/CmThink Watts/Cm22
Complex architecture at faster and smaller technologyComplex architecture at faster and smaller technologyDenser power is harder to coolDenser power is harder to cool
Wat
ts/c
m2
1
10
100
1000
1.5µ 1µ 0.7µ 0.5µ 0.35µ 0.25µ 0.18µ 0.13µ 0.1µ 0.07µ
i386i386i486i486
PentiumPentium®®PentiumPentium®® ProPro
PentiumPentium®® IIIIPentiumPentium®® IIIIIIHot plateHot plate
Nuclear ReactorNuclear ReactorNuclear ReactorRocketNozzleRocketRocketNozzleNozzle
* * ““New Microarchitecture Challenges in the Coming Generations of CMNew Microarchitecture Challenges in the Coming Generations of CMOS Process TechnologiesOS Process Technologies”” ––Fred Pollack, Intel Corp. Micro32 conference key note Fred Pollack, Intel Corp. Micro32 conference key note -- 1999.1999.
PentiumPentium®® 44
CentrinoCentrino®®
TACS June 2004 - 5 -
The Mobile EnvironmentThe Mobile EnvironmentMaximize performance & features within given Maximize performance & features within given constraintsconstraints
Power / ThermalPower / ThermalSize Size –– form factorform factorNoiseNoiseEnergy / Battery lifeEnergy / Battery lifeetc. etc.
Mobile platforms offer Tradeoff preferencesMobile platforms offer Tradeoff preferencesUser defined or builtUser defined or built--in schemein scheme
Compromise performance for longer battery life, lower acoustic noise, cool box etc.
TACS June 2004 - 6 -
OverviewOverviewMobile computers challengesMobile computers challenges
Average powerAverage powerThermal Design PowerThermal Design Power
Intel Pentium M power management Intel Pentium M power management The experimentThe experimentTest resultsTest resultsConclusionsConclusions
TACS June 2004 - 7 -
Die Temperature measurementsDie Temperature measurementsTemperature based control mechanismsTemperature based control mechanismsA diode connected to an external A/D A diode connected to an external A/D -- reports temperaturereports temperatureFixed temperature sensorsFixed temperature sensors
Max spec junction temperatureMax spec junction temperatureCritical temperature detectorCritical temperature detector
Source currentSource currentMeasured Measured VVbebe
n: diode ideality factork: Boltzmann constantq: electron charge constantT: diode temperature (Kelvin)VBE: Voltage from base to emitterIc : Collector currentIs : Saturation currentN: Ratio of collector currents
Critical#Critical#
refref
refref
High#High#highhigh
CRCR
A/DA/D
TACS June 2004 - 8 -
Controlling TDP Controlling TDP -- LinearLinearP states and T statesP states and T states
ClockClock
PowerPower
Temp.Temp.
TriggerTrigger
P>TDP
TACS June 2004 - 9 -
Controlling TDP Controlling TDP -- LinearLinearSelf or S/W trigger Self or S/W trigger –– Stop clock asserted, Power and temperature Stop clock asserted, Power and temperature dropdrop
ClockClock
STOPCLKSTOPCLK
PowerPower
TriggerTrigger
Temp.Temp.
P>TDP
TACS June 2004 - 10 -
Controlling TDP Controlling TDP -- LinearLinearSTOPCLOCK continues toggling for a pre defined timeSTOPCLOCK continues toggling for a pre defined timeAverage power and temperature drop Average power and temperature drop –– performance dropsperformance drops
ClockClock
STOPCLKSTOPCLK
PowerPower
TriggerTrigger
Temp.Temp.
Thermal significant timeThermal significant timeAverage Average Temperature Temperature
Average Power Average Power ReducedReduced
P>TDP
TACS June 2004 - 11 -
Controlling TDP Controlling TDP -- LinearLinearSTOPCLOCK continues toggling for a pre defined timeSTOPCLOCK continues toggling for a pre defined timeAverage power and temperature drop Average power and temperature drop –– performance dropsperformance drops
ClockClock
STOPCLKSTOPCLK
PowerPower
TriggerTrigger
Temp.Temp.
Thermal significant timeThermal significant timeAverage Average Temperature Temperature
Average Power Average Power ReducedReduced
P>TDP
Linear dependency (<1:1)Linear dependency (<1:1)Power to PerformancePower to PerformanceEnergy = P*t Energy = P*t no gainno gainLeakage impacts energyLeakage impacts energy
Full PowerFull Power
LeakageLeakage
TACS June 2004 - 12 -
Dynamic Voltage Scaling (DVS)Dynamic Voltage Scaling (DVS)One One µ µ ARCH implementation ARCH implementation –– Power and energy controlPower and energy control
ClockClock
VccVcc
PowerPower
SwitchSwitchThermal orThermal orUtilization Utilization
TACS June 2004 - 13 -
Dynamic Voltage Scaling (DVS)Dynamic Voltage Scaling (DVS)PLL relock at lower frequency at same VccPLL relock at lower frequency at same VccFast change Fast change –– no user experience impactno user experience impact
ClockClock
VccVcc
PowerPower
Short timeShort timeO.K with S/WO.K with S/W
TACS June 2004 - 14 -
Dynamic Voltage Scaling (DVS)Dynamic Voltage Scaling (DVS)Vcc drops gradually while CPU activeVcc drops gradually while CPU activePower savings changes from linear to FPower savings changes from linear to F33
ClockClock
VccVcc
PowerPower CPU power drops byCPU power drops byP=C*VP=C*V22*F *F function of Ffunction of F33
Performance drops byPerformance drops byFunction of F (Frequency)Function of F (Frequency)
Power savings > time increasePower savings > time increaseEnergy net savingsEnergy net savings
TACS June 2004 - 15 -
Dynamic Voltage Scaling (DVS)Dynamic Voltage Scaling (DVS)Vcc is ramped up increasing powerVcc is ramped up increasing powerOnce stable Once stable –– PLL relock at high frequencyPLL relock at high frequency
ClockClock
VccVcc
PowerPower
Switch Switch BackBack
TACS June 2004 - 16 -
Adaptive Energy ControlAdaptive Energy Control
Applications have wide dynamic power rangeApplications have wide dynamic power rangeRequire high power high performance burstsRequire high power high performance bursts
Determine user experienceDetermine user experienceTrade power performance as neededTrade power performance as needed
Driven by Operating system ACPIDriven by Operating system ACPIAverage power control on the fly Average power control on the fly -- ADAPTIVEADAPTIVE
HighHigh
LowLow
HighHigh
LowLow
TACS June 2004 - 17 -
ACPI ControlACPI ControlOperating system feature Operating system feature –– Industry standardIndustry standard
Supported by MS and LinuxSupported by MS and LinuxPassive and active policies defined for each zonePassive and active policies defined for each zoneUser preference User preference –– Max performance or Max batteryMax performance or Max battery
Switches _PSV and _ACxDVS or clock throttling used for passive power controlDVS or clock throttling used for passive power control
Available Power states published by BIOSDVS and once reached min Vcc - linearImplements PD controller algorithm
Fan on/off and speed used for active coolingFan on/off and speed used for active cooling_TMP = 60
-70--65--60--55--50--45--40--35--30--25-
-95-
-85--90-
-80--75-
_CRT
_PSV
_AC0
_AC1
500
700
900
1100
1300
1500
1700
1900
8:56:59 8:57:42 8:58:25 8:59:08 8:59:51 9:00:3540
50
60
70
80
90
100
Frequency
Temperature
TACS June 2004 - 18 -
TACS June 2004 - 19 -
TACS June 2004 - 20 -
OverviewOverviewMobile computers challengesMobile computers challenges
Average powerAverage powerThermal Design PowerThermal Design Power
Intel Pentium M power managementIntel Pentium M power managementThe experimentThe experimentTest resultsTest resultsConclusionsConclusions
TACS June 2004 - 21 -
Test SetupTest SetupThermal controlThermal control
CPU temperature is a function of power and ambient temp.CPU temperature is a function of power and ambient temp.Long time to heat the ambient and heat sinkLong time to heat the ambient and heat sink
Case temperature control formed stable environmentCase temperature control formed stable environmentHeating plate and fan to keep temperature at fixed temperatureHeating plate and fan to keep temperature at fixed temperatureForce extreme conditionsForce extreme conditions
Test casesTest casesSPECSPEC--IntInt and SPECand SPEC--FP componentsFP componentsUsed the self trigger mechanismUsed the self trigger mechanism
For repeatable & consistent resultsFor repeatable & consistent resultsCollected power temperatureCollected power temperatureand performanceand performance
Tj
DieHeating Plate
Main Board - Mobile
Tc
Fan
Controller
Testing on real silicon the Testing on real silicon the theoretical expected behaviortheoretical expected behavior
TACS June 2004 - 22 -
OverviewOverviewMobile computers challengesMobile computers challenges
Average powerAverage powerThermal Design PowerThermal Design Power
Intel Pentium M power management Intel Pentium M power management The experimentThe experimentTest resultsTest resultsConclusionsConclusions
TACS June 2004 - 23 -
Lineal vs. DVS controlLineal vs. DVS control
Thermal Throttle impact
Cooling [% of max]
SPEC
2K P
erfo
rman
ce [
% o
f max
]
~3%
94%
95%
96%
97%
98%
99%
100%
55% 60% 65% 70% 75% 80%
Linear
DVS
StartThrottle
Higher temp. Less coolingHigher temp. Less cooling
TACS June 2004 - 24 -
Lineal vs. DVS controlLineal vs. DVS control
Thermal Throttle impact
Cooling [% of max]
SPEC
2K P
erfo
rman
ce [
% ]
~3%
94%
95%
96%
97%
98%
99%
100%
55% 60% 65% 70% 75% 80%
Linear
DVSHigher temp. Less coolingHigher temp. Less cooling
TACS June 2004 - 25 -
Lineal vs. DVS controlLineal vs. DVS control
Thermal Throttle impact
Cooling [% of max]
SPEC
2K P
erfo
rman
ce [
% ]
~3%
94%
95%
96%
97%
98%
99%
100%
55% 60% 65% 70% 75% 80%
Linear
DVS
~3%
~4%
~5%
~6%
Shallower Slope:Shallower Slope:More cooling at same More cooling at same performance performance
Higher temp. Less coolingHigher temp. Less cooling
TACS June 2004 - 26 -
DVS set pointDVS set pointThrottle impact on performance
50%
60%
70%
80%
90%
100%
Cooling [% of max]
Perf
orm
ance
[%]
50 60 70 80 90 100
600 Mhz800 Mhz1000 Mhz1200 Mhz
Optimal point (performance) is at the highest Optimal point (performance) is at the highest frequency that dose not accede frequency that dose not accede Tj_maxTj_max with no with no transitions up and downtransitions up and down
100% throttleOverheat
TACS June 2004 - 27 -
0
0.2.
0.4
Ave
rage
CPU
Pow
er [
W]
0
50
100
150
200
Mob
ile M
ark
02 [S
core
]
AveragePowerScore
0.6
0.8
1.0
1.2
Average Power managementAverage Power management
112
600 MHzPower saving (Battery)
1600 MHzPerformance (AC)
1.1
216
0.37
Lowest freq.Lowest freq.
66% PowerEfficiency 1:1.4 - Static
48% Performance
TACS June 2004 - 28 -
Average Power managementAverage Power management
0
0.2.
0.4
Ave
rage
CPU
Pow
er [
W]
0
50
100
150
200
Mob
ile M
ark
02 [S
core
]
AveragePowerScore
0.6
0.8
1.0
1.2
112
600 MHzPower saving (Battery)
1600 MHzPerformance (AC)
1.1195
216
0.37
600 – 1600ADAPTIVE
0.63
10% Performance43% Power
48% Performance66% PowerEfficiency 1:1.4 - Static
Efficiency 1:4
TACS June 2004 - 29 -
Energy efficiencyEnergy efficiencyEnergy consumed for the SPECEnergy consumed for the SPEC--IntInt and SPECand SPEC--FPFP
We measured energy as E=We measured energy as E=∫∫ttP(tP(t))
Energy consumed at 1600 MHz defined as 100%Energy consumed at 1600 MHz defined as 100%
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Rel
ativ
e En
ergy
eff
icie
ncy
1600 1200 1000 800 600
Frequency [MHz]
Energy efficiencyLowest energy at Lowest energy at lowest frequencylowest frequency
TACS June 2004 - 30 -
OverviewOverviewMobile computers challengesMobile computers challenges
Average powerAverage powerThermal Design PowerThermal Design Power
Intel Pentium M power management Intel Pentium M power management The experimentThe experimentTest resultsTest resultsConclusionsConclusions
TACS June 2004 - 31 -
Summary and ConclusionsSummary and Conclusions
Intel Pentium M Specifically designed for mobileIntel Pentium M Specifically designed for mobileTargeting energy and power efficiencyTargeting energy and power efficiency
Mobile systems trade performance for powerMobile systems trade performance for powerTo meet user preferenceTo meet user preference
Enhanced Intel SpeedStep Technology provides Enhanced Intel SpeedStep Technology provides significantly improved power to performance and significantly improved power to performance and energy control schemeenergy control scheme
Silicon measurement confirm the theoretical workSilicon measurement confirm the theoretical workOptimal point for performance to power is at the highest Optimal point for performance to power is at the highest frequency that dose not accede frequency that dose not accede Tj_maxTj_max
Policy implemented into ACPI algorithmOptimal point for energy and battery life is at the minimum Optimal point for energy and battery life is at the minimum frequency possible with DVSfrequency possible with DVS
TACS June 2004 - 32 -
The authors would like to thank Nachum shamir, IrinaIlatov and Riad Durr for the thermal clamping measurements, to Jessie Garcia, Ali Saeed, Marco Wirasinghe for the evaluation and data collection for this article and to Cohen Aviad and Lev Finkelstein for results verification using simulation.
ThanksThanks
TACS June 2004 - 33 -
Die Temperature measurementsDie Temperature measurementsA diode connected to an external A/D reports temperatureA diode connected to an external A/D reports temperatureFixed temperatureFixed temperature
Source currentSource currentMeasured Measured VVbebe
n: diode ideality factork: Boltzmann constantq: electron charge constantT: diode temperature (Kelvin)VBE: Voltage from base to emitterIc : Collector currentIs : Saturation currentN: Ratio of collector currents
Appling 2 CurrentsAppling 2 Currents
Critical#Critical#
refref
refref
High#High#highhigh
CRCR
A/DA/D
TACS June 2004 - 34 -
Controlling TDP Controlling TDP -- LinearLinearSTOPCLOCK continues toggling for a pre defined timeSTOPCLOCK continues toggling for a pre defined timeAverage power and temperature drop Average power and temperature drop –– performance dropsperformance drops
ClockClock
STOPCLKSTOPCLK
PowerPower
TriggerTrigger
Temp.Temp.
Thermal significant timeThermal significant time
P>TDP
TACS June 2004 - 35 -
Dynamic Voltage Scaling (DVS)Dynamic Voltage Scaling (DVS)Vcc drops gradually while CPU activeVcc drops gradually while CPU activePower savings changes from linear to FPower savings changes from linear to F33
ClockClock
VccVcc
PowerPower CPU power drops byCPU power drops byP=C*VP=C*V22*F *F function of Ffunction of F33
Performance drops byPerformance drops byFunction of F (Frequency)Function of F (Frequency)
TACS June 2004 - 36 -
The Pentium M Power Control SchemesThe Pentium M Power Control Schemes
Linear power scalingLinear power scalingChange CPU frequency onlyChange CPU frequency only
reduces both power and performance linearly with frequencyreduces both power and performance linearly with frequencySaves power, No energy savingsSaves power, No energy savings
New Dynamic Voltage Scaling (DVS)New Dynamic Voltage Scaling (DVS)Reduces Voltage and frequency on the flyReduces Voltage and frequency on the fly
Frequency is dependent ~ linearly on VoltageFrequency is dependent ~ linearly on VoltagePower is a function of C*F*VPower is a function of C*F*V22
Reduction of both voltage and frequency provides FReduction of both voltage and frequency provides F33
power reduction for linear performance reduction power reduction for linear performance reduction Also results with energy savingsAlso results with energy savings
TACS June 2004 - 37 -
ACPI InterfaceACPI Interface
TACS June 2004 - 38 -
The Pentium M Power Control SchemesThe Pentium M Power Control Schemes
Linear power scalingLinear power scalingChange CPU frequency onlyChange CPU frequency only
reduces both power and performance linearly with frequencyreduces both power and performance linearly with frequency
New Dynamic Voltage Scaling (DVS)New Dynamic Voltage Scaling (DVS)Reduces Voltage and frequency on the flyReduces Voltage and frequency on the flyReduction of both voltage and frequency provides FReduction of both voltage and frequency provides F33
power reduction for linear performance reduction power reduction for linear performance reduction
See detailsSee details……