Power and Temperature Control on a 90nm Itanium … and Temperature Control on a 90nm Itanium®...

21
Power and Temperature Control on a 90nm Itanium ® Family Processor Jim Ignowski, Richard McGowen, Chris Poirier, Chris Bostak, Sam Naffziger Intel Corporation, Fort Collins, CO

Transcript of Power and Temperature Control on a 90nm Itanium … and Temperature Control on a 90nm Itanium®...

Power and Temperature Control on a 90nm Itanium®

Family Processor

Jim Ignowski, Richard McGowen, Chris Poirier, Chris Bostak,

Sam NaffzigerIntel Corporation, Fort Collins, CO

2

Motivation

• Maximize performance per Watt • Latest Itanium II (130nm) consumes 130 Watts

– Single core, 9MB L3 Cache

• Montecito (90nm) consumes 100 Watts– Dual core, 24MB L3 Cache

• Simple port would be well over 200 Watts– Need something more than low power circuit

techniques to address power problem

• Take advantage of P ∝ V2F

3

Take Advantage of P ∝ V2F

0

20

40

60

80

100

Pow

er (W

)

50 55 60 65 70 80 85 90 95

100

Frequency (% of Fmax)

75

Frequency

Frequency &Voltage

Core SwitchingCore LeakageOther

4

High Level View of System10s of µs

100s of ps

Voltage ControlVoltage Control

Frequency Frequency ControlControl

5

High Level View of System

MicroMicro--ControllerController

PowerPowerSensorSensor SupplySupply

VRMVRM

10s of µs

100s of ps

Frequency Frequency ControlControl

6

High Level View of System

MicroMicro--ControllerController

PowerPowerSensorSensor SupplySupply

VRMVRM

10s of µs

100s of ps

Frequency Frequency ControlControl

ThermalThermalSensorSensor

7

High Level View of System

MicroMicro--ControllerController

PowerPowerSensorSensor SupplySupply

VRMVRM

10s of µs

100s of ps

ThermalThermalSensorSensor

VoltageVoltageSensorSensor

Voltage to Freq.Voltage to Freq.Converter Converter ClockClock

8

Measuring Power

• Use package resistance to measure power• Avoids burning extra power in measurement• Portable, self-contained solution

– No dependence on external power supply

VConnectorVDie

RPackage

9

VID

6

Power Control System

+ -DAC+ -

PLimit

VDie

IIR

PCalc Calc

VConnectorA/D

A/D

Power SupplyMicro-Controller

RPackage

Package/Die

10

System Architecture –Power Measurement

( )Package

DieConnectorDieDieDie R

VVVIVPower

−==

• Measure voltages at both sides of the package to determine current and compute power (8µs sample period)

PowerSupply

RPkg

A/DConverters

MicroController

VDie

VConnector

6VID

LoadDie

11

Voltage Measurement

• By using an analog mux, we can reuse the same VCO to measure voltages from multiple sources– On die voltmeter

• High speed counter, >>10GHz for ~100µV measurement granularity

• 8µs counting interval for filtering and resolution

VCOVCO CounterCounter

VAnalogVDigital

12

VCO Detail

• Short loop for high speed• Analog “inverter”: bias circuit for complementary

control signal, also starts/stops oscillation• Performs averaging function by running at over

10GHz for 8µS

counthalt

VAnalog

VAnalog_n

13

VCO Detail– Delay Element

• Modulate output current via series output resistance combined with controlled drive fight– Large input voltage range with very linear response

GND

GNDVAnalog_n

VDDVDD

in out

VAnalog

14

• Two thermal sensors per core– Mux thermal diodes into VCOs to measure

temperature

Measuring Temperature

15

System Architecture –Temperature Measurement

• Calibrate the voltage drop at test to TJ target (90C)• Use the known -1.7mV per degree C temperature

coefficient to calculate die temperature• Measure the voltage drop across the diodes every

20ms

PowerSupply

A/DConverters

�����

�������

Die

��� �

������6

VID

16

Sources of Error and Inaccuracy

• RPackage variation with manufacturing process• RPackage variation with temperature• <25mV DC voltage drop across package

– 1% accuracy target requires 250µV• Noise on the measurement system supply

– VCO designed for speed, not power supply rejection• VCO gain and frequency variation over PVT• Inherent inaccuracies in the reference

components

Requires frequent calibration

17

Package Resistance Calibration

• Package resistance can be computed with two voltage measurements with processor stalled– Pulling quiescent current I0– Pulling I0 + a precision, on-die generated current IDelta

• On-package precision R for consistent IDelta

• Recalibrate every 66ms (<0.2% perf. hit)

I0 I0+IDelta

Vc2-Vd2

Vc1-Vd1

R Package

( ) ( )Delta

dcdcPackage I

VVVVR 1122 −−−=

18

System Overhead

• Burns less than 0.5W, less than <0.5% area

Foxton Controller

19

Frequency vs. Power LimitCore 0, Core 1, Avg Frequency vs. PLimit

84%

100%

100W 95W 90W 85W 80W 75W 70W 65W 60WPLimit

Freq

uenc

y

Core 0 Frequency Core 1 Frequency Avg Frequency

88%

92%

96%

20

Measured Data

Voltage

Power

Current

21

Conclusions• Power and temperature control system using

voltage and frequency modulation – Maximizes perf./Watt for at least a 25% efficiency gain– 90oC on-die temperature control with graceful throttle for

improved reliability and performance– Programmable power limit

• Voltmeter and package resistance calibration enables– 3% average power measurement accuracy– 1oC temperature control accuracy

• Embedded on die and package, consuming <0.5W and occupying <0.5% of the die