Track d more performance less power - freescale final

May 4, 2011 1

Bringing More Performance in Less Power

Freescale Semiconductor Israel Ltd.May 4, 2011

TM

Sergey Sofer

Freescale and the Freescale logo are trademarks of Freescale Semiconductor, Inc., Reg. U.S. Pat. & Tm. Off. All other product or service names are the property of their respective owners. © 2011 Freescale Semiconductor, Inc.

May 4, 2011 2

Outline• IC Power consumption – background• MSC8157 brief overview• State of the art power saving techniques description and their applications to

MSC8157.– Architecture – enabled power saving techniques.– Design– enabled power saving techniques.– Manufacturing technology and assembly – enabled power saving techniques.

• Summary

May 4, 2011 3

Power consumption - general

• Integrated Circuit (IC) device power consumption is defined as amount of energy, supplied by an external power supply to the device in pre-defined time period.

• We distinguish peak, maximum, typical application average and stand-by power consumption.– Peak power consumption defines external power supply capability including

bulk/decoupling capacitors influence.– Maximum power consumption defines system thermal requirements and usually given for a

time period, equal to or bigger than the system thermal reactance.– Average power consumption stands to define system requirement for typical application.– Stand-by power consumption describes system power consumption in stand-by/idle mode.

• Power consumption reduction became important to Modern Ultra Large Scale Integration (ULSI) products. Design teams fight for lower power consumption during all design stages.

May 4, 2011 4

IC Power consumption partitioning• Power consumption comprises dynamic and static parts.• Static power consumption is defined by IC active devices leakage and short circuit

currents:– Leakage is mainly caused by MOSFET devices sub-threshold and gate current.– Short circuit current is characteristic to biased circuits, pull-up or pull-down devices and some

terminated circuits.

– Static power consumption greatly depends on the power supply voltage VDD and temperature T.

• Dynamic power consumption is mainly set by active devises switching.– Dynamic power consumption depends on the power supply voltage VDD and average

system/device clock frequency F.P = f(VDD2,F)

P = f(VDD3,T)

May 4, 2011 5

MSC8157 – Overview• Freescale’s MSC8157 multi core baseband

DSP is one of the most powerful devices available in the market.

• It includes six SC3850 DSP cores at 1GHz, powerful MAPLE2 baseband accelerator, 3MB of L2 cache, 3MB of the shared memory, Serial-RapidIO, CPRI, PCI-express interfaces, supporting various communication protocols and security engine.

• The device is manufactured in 45nm technology.


May 4, 2011 6

Power saving techniques• State of the art power saving techniques include the following:

– Low power modes;– Power supply separation;– Active Power Shut-off;– Active Clock gating;– Multi-Voltage approach;– Multi-Frequency approach;– Multi-threshold CMOS design;

• MSC8157 make use of various power saving techniques, including architectural, design and manufacturing technology enabled – almost all known state of the art solutions, mentioned above.

May 4, 2011 7

Low power modes (LPM)

• Low Power Mode (LPM) is a mean of purposeful and environmental operation of a functional unit to achieve its lowest possible power consumption.

• LPM should be aligned with the purposeful utilization of the functional unit (block).• Proper LPM definition allows best power consumption reduction.• Improper LPM definition may cause opposite result and affects IC functionality and

performance.

Full utilization

High utilization

Low utilization Not in

use

Block functioning profile P

ower

cons

umpt

ion

May 4, 2011 8

Low power modes (cont’d)

• Full utilization of the functional unit requires all its resources available, so no LPM can be applied.

• The functional unit design is usually targeted to this operating mode.

Full utilization

High utilization


use


ower

cons

umpt

ion

May 4, 2011 9


• High utilization allows some functional or environmental relaxation to save either dynamic or static power or both:– Some voltage reduction saves both dynamic and static power;– Frequency decrease or stopping some clocking signals saves dynamic power.

Full utilization

High utilization


use


ower

cons

umpt

ion

May 4, 2011 10


• Low utilization allows variety of low power modes to be applied to the block.– Voltage reduction saves both dynamic and static power;– Frequency decrease or alternatively disabling part of the clock signals toggling saves dynamic

power;– Power gating may be extensively used and greatly saves static power.

Full utilization

High utilization


use


ower

cons

umpt

ion

May 4, 2011 11


• When the block is not in use its physical disconnection from the power supply provides maximal leakage power saving.– Physically disconnected blocks are not available for potential use.

• Alternatively on-die power gating also provides significant leakage power saving.– This can allow the block usage in the future.

Full utilization

High utilization


use


ower

cons

umpt

ion

May 4, 2011 12

MSC8157 Low Power ModesMSC8157 employs rich set of low power modes, allowing power consumption optimization according to the use-case.

• Doze mode – Stops internal clocks and can be recovered by software without resetting the block or

the device.• Deep sleep Mode

– Stops the source of the clocks and can be recovered by hardware reset (block dependent).

• Power Down Mode (Full or Partial)– Disables the power supply of some separated power domains (partial power down) or

all power domains (full power down) and can be recovered by power up sequence of the power domain(s) that was powered down.


May 4, 2011 13

Active Power Shut-off (PSO)

• On-die power gating (PG) or Power Shutoff (PSO) technique is used for leakage power reduction in VLSI devices – Characterized by disconnection of part(s) of a functional unit or the entire unit from the continuous

power supply network – Used when the unit’s operation is not required.

• The disconnection is provided by a switch, placed in power supply current path and controlled from another functional unit.

• PSO blocks when powered up, must start from the reset state.

VDDC

VDD gated

GND

Gated Power Domain

Switch

VDDC

GND

Continuous Power Domain

Control

Continuous supply

Gated supply

Continuous supply

May 4, 2011 14

Recovery from a PSO state

• Supply voltage recovery at the PSO exit takes time (dozens to hundreds of system clock periods). This is in order to keep low noise on the continuous or keep-alive power supply part (which provides the power supply voltage to devices that should remain powered on, e.g. data retaining devices, PSO controls etc).

• The consequence of that is the reduction in the PSO method efficiency, especially for short Low Power Mode (LPM) periods (like periods ## 2, 6).

tpower-up

7531

Time

Time

2 64L

PM

en

able

Vg

ated

su

pp

ly

uppoweroffleakonleaksaved PPPP

Short LPM period

May 4, 2011 15

MSC8157 Active Power Shut-off• Active power shut-off (PSO) technique is used in MSC8157 to reduce the static power

dissipation in “calm” periods for several functional units, which power consumption profile fits this feature.

– On/off switching is performed automatically by hardware – no application/user intervention required– Special (patent pending) fast power up/down sequence allows minimal impact on performance.


May 4, 2011 16

Clock disabling

• Clock disabling is an effective method of dynamic power saving.• Clock network toggling provides more than 50% of the dynamic power consumption of a

typical VLSI circuit.• Clock gating is a way of blocking the clock propagation, clock network toggling and

consequent clocking signal arrival to registers, when logic operation does not require that.

ENCK

GCK

CK

EN

GCK

P=f(VDD2,F)

Clock gate

May 4, 2011 17

Functional unit Clock disabling

EN_Z

CK

EN_AUnit A

Unit ZENClocksource

CKG_A

CKG_ZCKG

• Clock may be disabled either locally at the functional unit or globally for entire chip.• Local disabling allows fine resolution of the blocks power management:

• Recovery from the local disabling is usually immediate or very short.• Chip level clock distribution network and the clock source are still toggling even if all units are not in use.

Local clock gating

May 4, 2011 18

Global Clock disabling

EN_Z

CK

EN_AUnit A

Unit ZENClocksource

CKG_A

CKG_ZCKG

• Global clock disabling allows saving of the chip clock distribution network power.• All chip units are disabled.• Reset is usually required to recover from the global clock gating.

• Global clock source disabling (e.g. PLL stop) is the most efficient power saving way not employing power supply voltage manipulation.

• Reset is required to recover from the clock shut down.• Recovery time is usually long.

Global clock gating

Clock source disabling

May 4, 2011 19

MSC8157 - Active clock gating

• MSC8157 low power modes employ massive usage of active clock gating.– Clock source shut-down – deep sleep mode.– Global clock shut-down – doze mode.– Local clock shut-down – functional mode.

• Clock gating operation is implemented in the design and transparent to user.– Global clocks gating is applied by LPM activation.– Local clock gating is implemented during the design stage and is embedded in the

functional unit hardware


May 4, 2011 20

Power distribution network partitioning

• Power distribution network partitioning or power supply separation is a technique, used to provide different voltages to different parts of the IC.

• Static power saving for IC parts which are not in use for specific application.• Power saving for IC parts which operation is possible at lower voltage.

• Power distribution network partitioning impacts global power distribution system quality. It should be avoided if not required.

• Exaggerated power distribution network partitioning requirement and improper use may bring to opposite results.

VDD1

Power Domain

1

GND

VDD2

Power Domain

2

GND

Die

P = f(VDD2,F)

May 4, 2011 21

Dynamic voltage scaling (DVS)

IchipSoftware control function G(Ichip)

VDD

• Dynamic Voltage Scaling stands for varying the power supply voltage based on the system needs and serves for dynamic and static power saving:

• Manufacturing technology and environmental conditions (like “Faster” manufacturing process corner or lower temperature).

• Functional conditions (like LPM: “power down”, “deep sleep”, “doze” etc.)• Functional constraints (application demand).

• The disadvantage of DVS is that transition between different voltages takes time:• Functional unit operation during the voltage transition may be limited or blocked.

P=f(VDD2,F)

GND

May 4, 2011 22

MSC8157 DVS: Drowsy memory concept• MSC8157 makes use of special DVS arrangement for L2 cache and Shared

memories, known as the Drowsy memory approach. • The essence of the approach is that the power supply voltage of the memory is

kept at low level, guaranteeing the stored data retention, and is elevated when specific part of the memory is accessed.

• It allows substantial reduction in bitcell array leakage, characteristic to the high-end manufacturing technologies, at very high access resolution level.

• The control over the memory power supply voltage is done as the function of the memory state (read/write/retention).

• Zero impact in memory latency or performance is kept.

22


May 4, 2011 23

Frequency scaling (FS)

IchipSoftware control function G(f,Ichip)

VDD

• Frequency Scaling stands for selection of the system clock frequency, based on the system needs:

• Running not extensive application that does not require full system bandwidth.

P=f(VDD2,F)

GND

May 4, 2011 24

MSC8157 Multi-frequency approach • MSC8157 device provides multiple combinations of the frequencies for cores,

MAPLE2, peripherals, interfaces etc.• It has 3 system PLLs and one DDR PLL for supporting the clock frequency

programming of all functional units. • Pre-defined clock modes with various options for different frequency

combinations, allow running every unit at exactly the frequency required by application.

24


May 4, 2011 25

Multi-Threshold CMOS (MTCMOS)

• With the continuous scaling of CMOS devices, the device leakage current is becoming a major contributor to the IC power consumption and it is further expected to increase with the technology scaling.

• The leakage current of MOSFET devices greatly influenced by its threshold voltage Vth.

P=f(1/Vth)

May 4, 2011 26

MTCMOS (cont’d)

• For a VLSI circuit, some MOSFETs in non-critical logic signals electrical paths may be manufactured having higher threshold voltage.

• the leakage current is reduced;• the performance is maintained since low threshold transistors are used in the critical signal path(s).

• Both high performance and low power are achieved simultaneously. • Multi-threshold technique is good for leakage power reduction during both standby and

active modes without performance and die area penalty.

May 4, 2011 27

MSC8157 MTCMOS approach • MSC8157 SoC use 45nm technology providing different levels of MOSFET threshold

voltage from the lowest to the highest to find proper balance between the performance requirement and the maximal leakage power saving.

• This set allows saving most of the leakage power.


May 4, 2011 28

Power Saving in VLSI - Summary• Defined IC Power consumption, including its parts and its influencing operational and

environmental factors.• Overviewed Architecture – enabled power saving techniques: Low power modes

definition, Active Power shut-off (PSO) and Active Clock disable function.• Outlined Design– enabled power saving techniques: Multi-Voltage and Multi-Frequency

operation.• Discussed Manufacturing technology – enabled power saving technique - Multi-

threshold CMOS approach.

May 4, 2011 29

MSC8157 low power features - Summary

• Power consumption reduction is one of the primary targets of DSP design along with high performance and feature set enrichment.

• MSC8157 make use of almost all state of the art power consumption reduction techniques, allowing to achieve dozens of percents of the power consumption reduction while not compromising with performance and features set.

• Most of the low power solutions are hardware and transparent to user.• Those which require user involvement are well documented and supported by the

development tools.


May 4, 2011 3030

TM

Thank you for attention!


Track d more performance less power - freescale final

Education

dynamic power consumption

static power consumption

system power consumption

average power consumption

maximum power consumption

peak power consumption

lower power consumption

power saving techniquesstate