Future Power Technology in the Global Market...Power Continues to Increase - Reaching Limits of Air Cooling! Active and passive power (gate leakage and sub-threshold leakage) 10S Tox=11A

Integrated Supply Chain

© 2005 IBM Corporation1

Future Power Technology in the Global MarketFuture Power Technology in the Global Market

Dr. Scott D. Strand


September 20, 2005 © 2004 IBM Corporation

Program Director, Power/Cooling Procurement EngineeringIBM



The Technology Challenge/DriversMore Speed!Higher Density!More Data Capacity!More Bandwidth!Lower Cost!Smaller!More Portable!More Open!



Supercomputing Roadmap

1

10

100

1000

10000

100000

1995 2000 2005 2010 2015

Tera

flops

Source: ASCI Roadmap www.llnl.gov/asci, IBM, Supercomputer Conf., 2004Brain ops/sec: Kurzweil 1999, The Ace of Spiritual Machines

IBM Deep Blue®

US Dept. Of Energy ASCI

IBM BlueGene/P

IBM BlueGene/L®

2.8/5.6 GF/s4MB

5.6/11.2 GF/s1.0 GB

90/180 GF/s16 GB

2.8/5.6 TF/s512 GB

180/360 TF/s32 TB

System64 Racks, 64x32x32

Node Card32 Chips 4x4x2)

16 compute, 0-2 IO cards

Rack32 Node Cards

Compute Card

2 chips, 1x2x1

Chip2

processors

Moravec 1998, www.transhumanist.com/volume1/moravec.htm



Examples of MetadataMetadata provides information and adds meaning to data

! Text/Documents– Names, locations, phone

numbers, language, …

! Images– Name, date, time– Characteristics

! Rich, Streaming Media– Location, timing, scene

identification, participants, actions, ...

– Formats

! Relational data– Column attributes– Table values (domains)

! XML Documents– Tags, XML Schema

! Software Assets– Date, version, …– Interface definitions

! Web Services– Name, attributes, …– Interface definitions

! People Proxies– Name, location, serial no. …

Unstructured Information

Structured Information



Industry Implications For Metadata! The continued exponential growth of data will drive the need for,

and the explosion of, metadata

! The amount of metadata will significantly exceed the amount of data it describes

! Metadata will become a key enabler for information, application and business process integration

! Unexpected applications and businesses will emerge from the exploitation of integrated data and the creation and management of metadata

! Integrated search will become a key element of every productivity and enterprise application

! Metadata standards are emerging across every application domain and industry and will need to be reconciled with legacy corporate and industry schemata

! Enterprises need to re-think their information management strategy to exploit the emergence of new metadata standards



Storage Trends

1980 1990 2000 20100.0001

0.001

0.01

0.1

1

10

100

1000

Pric

e/M

Byt

e (D

olla

rs)

Range of Paper/Film

3.5 " HDD

2.5 " HDD

1 " Microdrive

Flash

DRAM

Storage Price

Store all customer dataTransactions, letters, phone calls

over the span of a lifetime

1980 1990 2000 20100.001

0.01

0.1

1

10

100

1000

10000

Are

al D

ensi

ty (G

b/in

2)

Rent Blockbuster™Not a video... the entire store

Storage Density



Semiconductor Technology

! “Traditional” CMOS scaling will slow, with innovation being the critical driver for future performance

! CMOS scaling coupled with innovation in device structures and materials is expected to provide an evolutionary path to the 25nm node

– Cycle time between technology nodes is very likely to continue to increase – Beyond 25nm, much uncertainty exists; much more radical devices may be needed

! Technology must now balance performance, power and variability– Power dissipation is limiting microprocessor performance, requiring new thermal

solutions– Process and statistical variability are emerging as dominant yield limiters

! Packaging will become an increasingly important component of performance

! Off-chip I/O bandwidth is becoming critical to continued system performance growth – Current technology approaching power and area limitations – Better electrical and optical solutions are needed

Semiconductor technology will continue to evolve and deliver performance but it will require innovation in device structures and materials, circuit design,

packaging, and a focus on process control



10000

Initial Ship Date

Freq

uenc

y (M

Hz)

1994 1996 1998 2000 2002 2004 2006 2008 2010100

1000

100000

10000

CMOS Microprocessor Frequency Growth

??! Microprocessor

operating frequencies will increase at half the historical rate

! Power dissipation is limiting performance

! Future differentiation based on function, not frequency



Future CMOS Roadmap: Less Predictive

! CMOS device performance will continue to improve rapidly, but in new ways" The concept of a scaled technology as we know it will cease to exist" Innovation will continue to drive performance improvements, but timing will be

harder to predict

19901990 19951995 20002000 20052005 20102010 20152015 20202020

Manufacturing Ramp StartManufacturing Ramp Start

NormalizedSemiconductor

Device Performance

Outlook

11

1010

100100

2000

2002

1995

2004

Molecular Scale Devices

High mobilityNew device structuresNew materials innovations



Power Continues to Increase - Reaching Limits of Air Cooling! Active and passive power (gate leakage and sub-threshold leakage)

10S Tox=11AGate Stack

0.010.110.001

0.01

0.1

1

10

100

1000

GateLength (microns)

Active Power

Passive Power

1994 2004Pow

er D

ensi

ty (W

/cm

2 )

Gate Length (microns)Gate dielectric approaching a fundamental limit (a few atomic layers)



Power is Limiting Microprocessor FrequenciesServer microprocessors cannot simultaneously utilize all their

transistors due to power limitations

Shippable PartsWith Leakage

Minimum Ship Frequency

Leff SlowFast

Nominal

Pow

er

Cooling/Power Limit

Max Freq(No Leakage)

Max Freq(With Leakage)

Power w/Leakage

Num

ber o

f Par

ts



Cross-section of NEC Earth Simulator showing cooling facilities

Supercomputers ……. Massive Cooling Needs



Power Management in Systems

P-III IdleP-III Loaded

P-IV System0

20

40

60

80

100

Syst

em P

ower

(wat

ts)

I/OI/O

CPUCPUMemoryMemory

I/OI/O

CPUCPU

MemoryMemoryIdleIdle

LoadedLoaded

PeakPeakPentium-III measuredPentium-IV estimated

Servers configured for expected peak loadAverage load can be much lessProcessor - over half of power when loaded

Hardware power management optionsSlow down CPU (mobile applications, etc.)

Power efficiency ~(processor voltage)2

Power down some nodes or slow down all nodesTrade-off power in memory & I/O vs. efficient CPU usage

Need power usage profile - measurements, models

Power efficiency gains from softwareBased on user models/expectations OS can further improve efficiency*More gains with other software approaches

Power consumption can be reduced at various levels, ranging from hardware controls to software methods, including power efficient OS and applications

Base Hardware OS-aided App-aided0123

4Relative Efficiency

* Source: J. R. Lorch & A. J. Smith, "Improving Dynamic Voltage Scaling Algorithms with PACE", SIGMETRICS 2001



Systems Will Be Designed With the Ability to Dynamically Manage and Optimize Power

! Power dissipation is increasingly limiting the performance of processors and systems driving new design techniques

– The power of high-end processor chips has exceeded 200 Watts

– Blades will push the power to 5 KW/sq-ft within the next 5 years

– Costs for cooling data centers will increase significantly

! Future systems will need to dynamically manage power at runtime

– Ability to quickly turn system elements on/off, control frequency including microprocessors, caches, memory and busses

I/O&Misc3% L3

11%

Fans23%

I/O Fans5%

DRAM System30%

Processors28%

Power distribution of a typical server

Power densities of all system-level elements (device to rack) are increasing exponentially



Implications! Systems performance leadership will be enabled by judiciously

balancing performance and power consumption, rather than maintaining historical frequency trends

! Power dissipation is limiting microprocessor performance, requiring aggressive power-management techniques and new thermal solutions

! Innovation in device structures and materials, not scaling, will be the primary driver of semiconductor performance improvements

! Integration and optimization over the entire systems stack will be critical to maintaining traditional server performance trends



2004 Platform Symposium Summary: Still valid today!

Technical Challenges and Solutions

Higher Reliability - MTBFHigher Power DensityHigher Transient Response - di/dt Higher EfficiencyLower Voltage - Higher CurrentVoltage/Current Distribution Increased Number Of Voltage DomainsAbility to Hot SwapError and Status ReportingIncreased Mobile Client Power NeedsLower CostShorter Development CyclesBest Industry Reliability

More IntegrationHigher Switching FrequenciesLower Switching and Conduction Losses Topology Influences

RES/ZVS/ZCSBetter EMI DesignInnovative Design Lower Output ImpedanceThermal ManagementComponent Improvements

IntegratedBattery TechnologyPower SemiconductorsCapacitorsInterconnect



Technical Challenges• More Processors Per Server• 1U & 2U Server Packages• Higher Power Density• Higher Transient Response - di/dt • Higher Efficiency• Lower Voltage - Higher Current• Voltage/Current Distribution Within Server• Ability to Hot Swap - Availability• Error and Status Reporting – Servicability• Best Industry Reliability• Lower Cost• Shorter Development Cycles



Digital Power Control: Is it the solution?

Provides a huge advantageShorter Development Cycles

Advantage is in reuse and development costsLower Cost

Allows much more information to be saved and reportedError and Status Reporting

Support via softwareAbility to Hot Swap

Allows flexibility to define multiple voltage levelsIncreased Number Of Voltage Domains

Integration of function allows more space on boardsVoltage/Current Distribution

Integration of function allows more room for power components

Lower Voltage - Higher Current

More control of loops and frequencyHigher Efficiency

Control of compensation loops could allow greater transient response

Higher Transient Response - di/dt

More integration of functionHigher Power Density

Fewer componentsHigher Reliability - MTBF

Digital SolutionWhat We Need



Digital Power Control: Is it the solution?

Greater integration helps overall reliabilityComponent Improvements

Greater control of switching and loops based on loadThermal Management

Digital solution provides no advantageLower Output Impedance

Allows flexibility in the design/development phaseInnovative Design

Digital solution provides no advantageBetter EMI Design

More control of loops, frequency, and feedbackTopology Influences

Digital solution provides no advantageLower Switching and Conduction Losses

Digital control provides higher frequency capabilityHigher Switching Frequencies

Fewer componentsMore Integration

Digital SolutionHow We Get There



Comments Regarding Digital Power

! Digital Power Management is Not New!!!– Used in Telecomm and Computer Systems Since 1980’s (for example, SPCN for

AS/400, now iSeries and PCN for S/390, now zSeries)

– Also Has Been Used in UPS, Motor Drives, and Utility Industries

! Many Proprietary Solutions

! Standardization Efforts are under way, but need coordination from end users

! Little or No University Research to Satisfy industry needs

! Total Cost of Digital Power Solutions is Competitive



What is New in Digital Power?

1. Ability to Control Real-Time Cycle-by-Cycle Operation Of A Power Converter

2. Power Standard Command Protocol – PMBusTM

to manage power.

3. Cost Effective Solutions - Chip Set Manufacturers



Benefits of Digital Power Management

! Reduced Development, Qualification and manufacturing cost! Higher Quality and Reliability at competitive cost! Reduced Parts Count - Two chips at the most with unlimited functions due to

firmware. Same control card for different applications. Low Inventory

! Reduced Board Area due to Integrated modules. ! Flexibility Through Programmability

– Engineering Changes through firmware and not by Soldering Iron – Can be useful in Development, Qualification, Manufacturing, and Field.

! Calibration at Final Functional Test; application specific parameter! Less Sensitive to Environment and Component Drift! Better Noise Immunity ! Unlimited parameter monitoring – continual quality improvement



Barriers To Implementation of Digital Power! Legacy Systems and Risk Management! Perception …. It Will Cost More

– Piece Price Probably Will Increase– Total Cost of Ownership Is Difficult To Quantify, but this must be the measure…

! Not Many Digital Power Experts within the industry– Power Is an analog world – digital power is an uncomfortable area.

! Resources and Buy-In To Define Architecture To Best Implement A Digital Power Solution– NOTE: This is not a power system item only. This requires buy-in and innovation throughout the entire

system including the operating system, firmware, and service processor– Without this Buy-In, The Advantages of Digital Power Are Limited To Essentially What is Available

Today. Would An Optimized Digital Power Solution Justify The New Architecture?

! Each Chip Set Supplier Has Their Own Proprietary Tool for Optimizing the Loop And Setting Parameters. No standard Chip sets by semiconductor industry.

! Digital Power Offers Many Possibilities - Lack of usage of the available features. ! Digital Systems In A Noisy Environment Such As Power Demands Good Design

Practices for Common Mode Noise and Differential Mode Noise Rejection. – For Example: SF4 System Had An Issue Due to Noise on I2C Bus Causing Bits To Toggle Randomly



What Do We Need From The Industry?! Solutions, solutions, solutions!!!

– Provide drop-in solutions to legacy products– Solid designs to eliminate potential noise issues with higher

frequency digital signals

! Standardization without sacrificing function and capability– This needs to be driven by the end-users – we have to get what we

need

! Cooperation among component suppliers, power supply suppliers, and end users– Information and solutions must be shared while maintaining

competitive advantage and profit– Component suppliers and power suppliers should work together to

provide solutions to the industry – share expertise and knowledge



Future Power Technology in the Global MarketFuture Power Technology in the Global Market

Dr. Scott D. Strand


September 20, 2005 © 2004 IBM Corporation

Program Director, Power/Cooling Procurement EngineeringIBM