Delivering Performance While Conserving Power (AMD) Brent Kerby Product Manager, Servers & Workstations Microprocessor Solutions Sector AMD www.amd.com Did You Know? The combined total of data centers in California are estimated to require 250MW – 375MW of energy. That’s equivalent to 3,495 – 5,242 barrels of oil a day! SOURCE: California Energy Commission http://www.energy.ca.gov/reports/2004-04-07_500-04-004.PDF IT Managers Have A Problem Power Consumption/Cooling Issues Tracked By Company: 71% Power Consumption Increased Amount Of Power Supplied To Data Center 12% 21% Cooling How Are Companies Addressing These Issues Both Equally Important 38% Stopped Buying More Servers And/Or Consolidated Existing Equipment 26% Implemented "Cool Aisle/Hot Aisle“ Layout 25% Increased Size Of Data Center Suspect That Both are Issues but do not Track at this Time Neither Presents An Issue at this Time 44% 23% 12% 17% Other 16% None Of The Above 15% Base: 1,177 IT Decision Makers November 2005 Strategy Group/Ziff-Davis Average of 18% of total rack space wasted due to power and cooling issues Microprocessor Innovations Reducing power and cooling Processor Power Management Virtualization Technology Memory Sub-System Dual-Core native design Efficient Architecture OS-Directed Power Management AMD PowerNow!™ technology with Optimized Power Management Multiple performance states for optimized power management Dynamically reduces processor power based on workload Lowers power consumption without compromising performance Leverages industry standards Up to 75% CPU power savings at idle P-State P0 HIGH 2600MHz 1.40V ~95watts P1 2400MHz 1.35V ~90watts P2 2200MHz 1.30V ~76watts P3 PROCESSOR UTILIZATION 2000MHz 1.25V ~65watts P4 1800MHz 1.20V ~55watts P5 1000MHz 1.10V ~32watts LOW EXAMPLE ONLY OS-Directed Power Management Avg CPU Core Power (SPECWebSSL) -7% AMD PowerNow!™ Technology with OPM - DISABLED MAX 30.000 AMD PowerNow!™ Technology with OPM - ENABLED POWER 25.000 20.000 -43% -55% -73% 15.000 -75% -69% 10.000 5.000 0.000 MIN 100% Utilization 80% Utilization 60% Utilization 40% Utilization 20% Utilization Idle Standby AMD PowerNow! technology can provide up to 75% power savings! Benchmarks show no measurable performance impact! Efficiency Through Instruction Sets ‘AMD64 Technology’ ‘Virtualization’ Virtual Machine Virtual Machine Virtual Machine Application Application OS OS Application OS 32-bit to 64-bit ~25% increase Getting more done without increased power and heat Can Reduce number of power consuming servers http://www.amd.com/us-en/Processors/ProductInformation/0,,30_118_8796_8800~100357,00.html Dual-Core Increasing performance without facilities upgrades AMD Opteron™ AMD Opteron™ (Single-Core) (Dual-Core) Power Drop of 9 KVA OR 29 VS. Room for more growth Less Power Less Heat Less Servers Less Space 41% 17 Dual-Core Processors Increasing throughput within the same power and cooling envelope as singlecore processors Performance based on SPECweb®99_SSL 82,244 connections http://www.amd.com/us-en/Processors/ProductInformation/0,,30_118_8796_8800~102056,00.html Architectures Effects On Power CORE 190 watts 8 GB/S 300 wattsCPU CORE CORE CORE CPU (150w MCP Native Dual-Core Native Dual-Core MCP Dual-Core MCP Dual-Core SRQ per CPU)MCP (95w per CPU) Crossbar Mem.Ctrlr SRQ Crossbar HT Mem.Ctrlr 8 GB/S I/O Hub I/O Hub PCI-E Bridge PCI-E Bridge PCI-E Bridge 22 watts Memory Controller Hub HT 8 GB/S PCI-E Bridge PCI-E Bridge 8 GB/S USB I/O Hub PCI Legacy x86 Architecture 300 watts for processors (150w each) 22 watts for external memory controller ~$451 per year (1 server) ~$225,698 per year (500 servers) ~69% More 322 watts AMD64 Technology with Direct Connect Architecture 190 watts for processors (only 95w each) Integrated memory controllers ~$266 per year (1 server) ~$133,000 per year (500 servers) Energy includes power input & cooling, Power Utility cost: $0.10/KW-hr, Publicly available processor & chipset specifications. The examples contained herein are intended for informational purposes only. Other factors will affect real-world power consumption and cost 190 watts Architectures Effects On Power CORE CORE CORE CORE MCP(173w MCP CORE CORE 692 watts perMCP CPU) CORE CORE MCP SRQ SRQ Crossbar Crossbar Mem.Ctrlr HT Mem.Ctrlr HT 380 watts (95w per CPU) 14 watts USB Memory Controller Hub I/O Hub Hub I/O PCI 8 GB/S PCI-E Bridge PCI-E Bridge PCI-E Bridge SRQ SRQ Crossbar Crossbar Mem.Ctrlr HT Mem.Ctrlr HT 8 GB/S 8.5 watts XMB 8.5 XMB watts 8.5 watts PCI-E Bridge 8.5 watts XMB 8 GB/S XMB PCI-E Bridge 8 GB/S USB I/O Hub PCI Legacy x86 Architecture 692 watts for processors (173w each) 48 watts for external memory controller ~$1037 per year (1 server) ~$518,592 per year (500 servers) ~95% More 740 watts AMD64 Technology with Direct Connect Architecture 380 watts for processors (only 95w each) Integrated memory controllers ~$533 per/year (1 server) ~$266,304 per/year (500 servers) Energy includes power input & cooling, Power Utility cost: $0.10/KW-hr, Publicly available processor & chipset specifications. The examples contained herein are intended for informational purposes only. Other factors will affect real-world power consumption and cost. 380 watts Challenges Efficient Power Supplies Typically only 72% efficient (at full load) That’s 280 watts in a 1000 watt load! Efficient Voltage Regulator Modules Processor, Memory Memory Technologies DDR2 DIMM = ~6 watts each FBDIMM = ~12 watts each Infrastructure Support of OS-Directed Power Management Driving Innovation And Awareness To Make A Difference… Delivering Performance While Conserving Power The mobile perspective Frank P. Helms Principal Member Technical Staff Lead Mobile System Architect Microprocessor Solutions Sector AMD www.amd.com Agenda AMD Mobile Power Management Philosophy AMD Turion™ 64 X2 Mobile Technology General Characteristics Power Management Mechanisms Dual Core Processor Benefits for Notebook PCs Windows Vista Power Management Demo Power Management Philosophy Both OS and hardware are important To optimize battery life and performance both the operating system (OS) directed and Hardware (HW) autonomous mechanisms must be involved in power management Power Management Philosophy Mechanisms controlled by the OS Mechanisms controlled by the OS have the following characteristics Involve user policy/preference Interactions with Applications and Drivers Knowledge of CPU usage by the applications which are running Examples of Mechanisms controlled by the OS include Transitions in and out of ACPI S-states Processor P-state control Halt (Processor C-state control) Power Management Philosophy Controlled Autonomously by HW Mechanisms that are controlled autonomously by Hardware have the following characteristics Transparent to the OS, Software, and the User Mechanisms that are on microsecond and finer time scales Examples Using pre-charge power down mode (memory) Fine grain clock gating (processor, GPU, etc.) HyperTransport™ Link power management AMD Turion™ 64 X2 Mobile Technology General characteristics Dual Core (512 KB L2 per core) Dual Channel DDR2 HyperTransport @ 800 MHz (6.4 GByte/s bandwidth total) Manufacturing process that is optimized for mobile (to minimize leakage current) 35x35mm lidless package and low profile socket AMD Turion™ 64 X2 Mobile Technology Core power management C0 P-states Dynamic fine grain clock gating of idle logic Pre-charge power down mode for system memory Cores can be halted (C1) individually Core by core power management C1E (Enhanced): When both cores are halted, additional power savings mechanisms can be applied HyperTransport™ link can be disconnected System Memory in self refresh On-die northbridge clock grid ramps down AMD Turion™ 64 X2 Mobile Technology Power management overview SODIMM Deeper Sleep: Core1 Core0 VDD plane C1 C0 C1 C0 voltage is reduced further Northbridge SODIMM HT Disconnected HT Connected LCD Chipset with GPU Both cores Core Both 1 cores Halted executing (C1) clock halted code (C0) grid enables ramped down hardware C1E Cores clock grids ramped down HT Disconnected System Memory in self refresh On-die Northbridge clock grid ramped down Chipset clock gating etc Deeper Sleep (AltVID) Dual Core CPUs Can Improve Performance And Battery Life Higher performance per watt Example TDP Breakdown for 2 GHz AMD mobile processors 2005 Single Core (2GHz) 2006 Core 0 (2GHz) + + Core 1 (2GHz) + On Die Northbridge = 32.8 W On Die Northbridge = 32 W I/O Power with 64bit DDR interface = 2.2 W I/O Power with 128bit DDR2 interface = 3.0 W Fixed 35W TDP Windows Vista Power Management Processor power management AMD will be evaluating all Windows Vista configurable power management options AMD will publish recommendations for the use of Windows Vista power management options to optimize performance and battery life Dual Core Processor Frank P. Helms Lead Mobile System Architect AMD Japan Engineering Lab Call To Action Use AMD Turion™ 64 X2 Mobile Technology for the best mobile experience with Windows Vista Contact AMD regarding Recommendations for optimized Windows Vista Power Policy settings for notebooks later this year Additional information on AMD Turion™ 64 X2 Mobile Technology Delivering Performance While Conserving Power The desktop perspective Paul C. Stanley Desktop Strategic Platform Planning Microprocessor Solutions Sector AMD www.amd.com Agenda Desktop Power Consumption The Energy Efficient Desktop The Microprocessor Contribution Desktop Power Management Platform Strategies for Reducing Power Windows Vista Power Management on the Desktop Call to Action Desktop Power Consumption “Enthusiast” Class platform can easily exceed 800W - 1000W of total system power High-Performance CPU + Dual GPU can exceed 60A on 12V rails ATX 2.0 Specification = two rails at 20A per rail 8-phase VRM, 4 GB System Memory, GPU Physics Card, 4xSATA RAID, 2xGbE, USB, 1394, … Extreme Overclocking… High-End Home Theater can exceed 400W+ A/V tuner(s), external speakers, 7.1 Surround Sound, HD DVD, “big” LCD Monitor, etc. Let’s face it – higher performance = higher power The Energy Efficient Desktop Typical Mainstream PC, measured at the wall At Idle: 150W - 200W http://www.tomshardware.com/2005/06/03/dual_core_stress_test/ Under Load: 200 W (UMA) – 350 W (discrete) Energy Efficient Desktop: ~50W to ~100W http://www.silentpcreview.com/article313-page1.html Typical Commercial usage model 8 hours per day, 90%+ Idle Energy Efficient Desktop (estimated) electricity cost savings alone can easily exceed $100+ per year per platform IDC Prognostication On SFF http://www.eweek.com/article2/0,1759,1948894,00.asp?kc=EWRSS03119TX1K0000594 Platforms include Small Form Factor (SFF), Ultra-Small Form Factor (USFF), All-In-One (AIO), Living Room PC (LRPC), Blade PC... 2005 U.S. Markets: ~31M Minitowers (77% of desktop market), ~5M SFF (12%) 2009 U.S. Markets: ~11M Minitowers (31%), ~15M SFF (42%) 2005 to 2009 Global Markets: ~93M Minitower units (70%) drops to ~78M (< 50%) 2005 to 2009 Global Markets: SFF ~12M units (9%) rises to ~48M (30%) SFF produces less heat and noise Dual Core CPU provides more performance for less energy cost SFF in standard form factors, with standard components, keeps costs low eWeek On SFF http://www.eweek.com/article2/0,1895,1854790,00.asp http://www.eweek.com/article2/0,1895,1934114,00.asp Enterprise will drive the move to SFF Management costs, electricity, acoustics, all call for smaller, quieter desktop Expandability gives way to integration Shrinking the size (down to 6-9 liters) uses less power Standard desktop parts minimizes platform cost Data Center electricity usage is at the top of IT managers' list of worries Balance and tradeoff between performance and power "...the lowest-power system...pays me back over the life of the system" Desire for objective ways to measure power consumption, to be developed by third parties eWeek On SFF Trends Out To 2015 http://www.eweek.com/article2/0,1895,1949079,00.asp?kc=ewnws041306dtx1k0000599 Full-size tower is marginalized Minitower (dominates U.S. retail today) is diminished SFF (7-20 liters) becomes most popular, uses low-cost desktop components SFF acoustics become unobtrusive, form factor “fits anywhere” (LRPC, etc.) USFF (1-7 liters) uses higher priced notebook components (miniaturization) All-In-One (U)SFF a la Apple iMac Blade PC Client for the Enterprise The Microprocessor Contribution AMD Athlon™ 64 Processors AMD Athlon 64 X2 vs AMD Athlon 64 (V)EE CPU AMD Athlon 64 X2 4800+ @ 89 W standard TDP AMD Athlon 64 EE 4800+ @ 65 W, VEE 3800+ @ 35 W AMD Athlon 64 X2 drops to ~9 W at Idle (V)EE drops as low as ~1 W (matching AMD Turion™ 64) AMD Athlon 64 embeds TCaseMax and TDP data AMD Athlon 64 TDP has steadily declined Examples (CPU samples are widely divergent, your mileage may vary) CPU A64 X2 3800+ (Toledo) A64 4800+ (Toledo) A64 3000+ (Venice) A64 4000+ (San Diego) TCaseMax 63°C 59°C 57°C 59°C TDP 65.6 W [standard 89 W] 85.0 W [standard 110 W] 44.1 W [66% of standard 67 W] 50.0 W [56% of standard 89 W] AMD Athlon™ 64 Energy Efficient CPUs Energy-efficient processors drive new class of platforms Thin Clients and PC Blades for Enterprise SFF Desktops for Home and for Business Smaller footprint reaches further into the Home Better Acoustics are Required Accurate Temperature Monitoring enables improved fan control Total-System Energy Efficiency is Required Regulatory Programs (EnergyStar, Blue Angel, Nordic Swan…) are driving challenging new requirements for Energy Efficiency Building Blocks Energy efficient components Improved power management leveraging mobile technology Robust implementation of Cool’n’Quiet™ Technology Accurate Temperature Monitoring infrastructure Desktop Power Management It’s not about power management, per se It’s really about delivering a better user experience, lowering costs, and enabling new usage scenarios Lower thermals = fewer/slower system fans Less noise = improved acoustics Desktops are consuming more power, electricity costs are rising – Do the math Home use scenarios demand quiet systems 7 of 10 IT Shops track power consumption http://enterprise.amd.com/Downloads/Ziff_Power_and_Cooling_IT_survey_en.pdf AMD Power Management AMD Cool‘n’Quiet™ on AMD Athlon™ 64 CPU Performance On Demand Up to 75% or more CPU power savings at Idle No impact on the performance of other platform subsystems (Memory Controller, HyperTransport...) Fully synchronized Dual Core power management Load Balancing across dual cores can lower total CPU power consumption by as much as 50% AMD Dual Core architecture optimizes chip-level interconnects for faster, lower-power operation AMD DirectConnect Architecture eliminates bus interface silicon and other overhead signals Saves “tens of watts per system” http://enterprise.amd.com/Downloads/34146A_PC_WP_en.pdf Platform Strategies For Lowering Power Undervolting and underclocking components Lower voltage chips running at slower clock speed consume less power Deliberately choosing low-end components Often lower performance and lower voltage, but also lower power Selecting more efficient components Higher performance-per-watt ratio Using notebook PC components and subsystems But Desktop CPU is preferred over Mobile CPU Ease of implementation, infrastructure cost model… Platform Impact Trends And Potentials AMD Athlon™ 64 CPU Roadmap now includes AMD Athlon 64 Energy Efficient (EE) CPU at 65 W and Very Energy Efficient (VEE) CPU at 35 W Enhanced Power Management Possibilities C1E Enhanced Sleep Clock gating can be beneficially implemented in virtually any silicon component GPU performance scaling is gaining traction Away Mode/Instant On for Media Center PC Enhanced Thermal and Acoustics Management Multiple thermal sensors to optimize fan control Instrumented PSU to enable PSU fan control Additional Hardware Options CPU, Motherboard, VRM, Chipset, PSU, System Memory, Storage (HDD, ODD…)… Select power efficient components Example: Notebook HDD consumes ~1 W Discrete Graphics adds 15 W+ at Idle and 30 W+ under load Many and varied platform power conversions waste as much as 50% of AC power input Improve PSU efficiency Today’s “better” PSU = ~70%, EnergyStar proposal = 80% Other global programs: Blue Angel, Nordic Swan... Improve VRM efficiency (today ~80%) Implement more highly power-optimized “sleep” components and algorithms Windows Vista Desktop power management Robust support for Sleep states encourages Windows Vista Power Management usage on Desktop PC Desktop Default is Hybrid Sleep: S3 Standby → S4 Hibernate Focus on Desktop power consumption Group Policy control for IT usage Optimized S3 Resume performance (< 2 seconds) Call To Action The entire PC ecosystem must cooperate and contribute to maximize the potential of the Energy Efficient Desktop Other Resources Download A64MaxTemp.zip (TCaseMax and TDP readout) freeware at http://www.thecoolest.zerobrains.com/forums/viewtopic.php?t=83