* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download ppt - ISLPED
Variable-frequency drive wikipedia , lookup
Power factor wikipedia , lookup
Power inverter wikipedia , lookup
Buck converter wikipedia , lookup
Immunity-aware programming wikipedia , lookup
Pulse-width modulation wikipedia , lookup
Utility frequency wikipedia , lookup
Standby power wikipedia , lookup
Wireless power transfer wikipedia , lookup
Rectiverter wikipedia , lookup
History of electric power transmission wikipedia , lookup
Voltage optimisation wikipedia , lookup
Power over Ethernet wikipedia , lookup
Electrification wikipedia , lookup
Electric power system wikipedia , lookup
Power electronics wikipedia , lookup
Distribution management system wikipedia , lookup
Audio power wikipedia , lookup
Amtrak's 25 Hz traction power system wikipedia , lookup
Power engineering wikipedia , lookup
Switched-mode power supply wikipedia , lookup
Alternating current wikipedia , lookup
Low Power Design From Technology Challenges to Great Products Barry Dennington Snr VP CTO/SoC Design Engineering October 5, 2006 Agenda Is power really a problem? Are there viable solutions? What are the challenges to use them? Designing low-power products Conclusions Is power really a problem? Scaling increases power more than expected CMOS 65nm technology represents a real challenge for any sort of voltage and frequency scaling – Supply voltages stable at 1.2v Starting from 120nm, each new process has inherently higher dynamic and leakage current density with minimal speed advantage – 90nm to 65nm: same dynamic power and ~5% higher leakage/mm2 Low cost continues to drive higher levels of integration Low cost technological breakthroughs to keep power under control are getting very scarce – Examples: changing device or tuning the process to the application 4 ISLPED Keynote, B. Dennington, October 2006 Modern SoC’s demand more power Logic: – Static power is growing really fast – Dynamic power kind of grows Memory – Static power is growing really fast – Dynamic power kind of grows Overall power is dramatically increasing “Power-Efficient System-on-Chip power Trends, System Drivers”, International Technology Roadmap for Semiconductors (ITRS) 2005 5 ISLPED Keynote, B. Dennington, October 2006 But, do we need to bother with power? The mobile device consumer demands more features and extended battery life at a lower cost – About 70% of users rate longer talk and stand-by time as primary mobile phone feature – Top 3G requirement for operators is power efficiency Customers want smaller, sleeker mobile devices – Requires this high levels of Silicon integration in advanced processes, but … – Advanced processes have inherently higher leakage current Therefore, we do need to bother with reducing power! 6 ISLPED Keynote, B. Dennington, October 2006 Increasing the Challenge; conflicting requirements Low cost is always critical in the consumer market – Cannot afford exotic packaging to solve power consumption issues – Products must consume less power Home consumers want products that enhance the user experience – Reduced noise (no fans) – Environmental issues When docking mobile devices for in-home use, consumers expect the same performance as tethered products – Relief from device battery life constraints – Products must be able to deliver high performance when docked 7 ISLPED Keynote, B. Dennington, October 2006 Thus, is power really a problem? Yes Power is a problem & the user needs increase the challenge !!! 8 ISLPED Keynote, B. Dennington, October 2006 What can we do? An holistic approach for a pervasive problem Low Power requires an holistic approach across many areas – System solutions: Software power management control, OS and Firmware, instruction set extensions, power management devices – SoC design technologies: Optimized processors, voltage and frequency scaling, design architectures, tools and flows, quality of service – Low-power building blocks: Ultra low power processes, low power IP, advanced packaging strategies A product conception and design team need expertise and solutions in all these areas Each partner in the production/supply chains need expertise and solutions in all these areas Unfortunately, low-power solutions normally conflict with the low-cost requirement 10 ISLPED Keynote, B. Dennington, October 2006 Holistic approach: system first Audio DAC SRAM 2% 3% HDD Drive 15% DC/DC 15% Audio 6% Understand the trade-offs SDRAM 0% Power [rel] vs. Application DataRate[kbps] for Different Video Sources 1600 Video 19% 1500 1400 Flash 0% LCD 40% 1300 1200 Identify where to act !!! 1100 1000 0 512 1024 1536 2048 HDD 2560 802.11b 3072 35084 4096 UWB 11 ISLPED Keynote, B. Dennington, October 2006 Holistic approach: define the problem P = (1-AF) Pidle + AF • Pdynamic Optimization space Application dependent !!! 12 ISLPED Keynote, B. Dennington, October 2006 Holistic approach: AF < 50% The system is mostly idle. Thus, minimize stand-by power! – For example: pagers and mobile phones. Minimize software activity in stand-by – Make stand-by a real stand-by Switch off power from unused modules, ICs and cores – Use MSV or similar techniques Use high Vt to minimize Ioff – Minimize the intrinsic leakage Choose a process with a high Ion/Ioff ratio – Basically any currently named Low Power process should do 13 ISLPED Keynote, B. Dennington, October 2006 Holistic approach: AF > 50% The system is mostly active. Thus, minimize dynamic power! – For example: DVD players, Sony PSP, etc. Use Software Power Manager to use just-enough performance and power – Do not waste performance when not needed. Make your system adaptive (e.g. voltage/frequency scaling) according to the nature of your application – Use all the time every task has to complete. Choose low-power IOs, memories, libraries, etc. Use a multiple-Vt design style and clock gating. Choose a process with a low Ion/Ioff ratio – This is not what is typically called an LP process!!! 14 ISLPED Keynote, B. Dennington, October 2006 Holistic approach: AF ~ 50% The system behavior is not constant. It’s the low power nightmare! – For example: a pocket PC or a Smartphone (used as such) Make your system really adaptable using aggressive voltage/frequency scaling, back biasing and a process with tunable Ion/Ioff ratio coupled with Software Power Management wherever possible! Use prediction of the system loading to better tune it. Final power budget will be worse when comparing the same function in such a system with respect to the previous two cases!!! 15 ISLPED Keynote, B. Dennington, October 2006 Holistic approach: solution space Reduce cost & improve scalability Optimize system and software for minimum power consumption Tuneable Ion/Ioff processes Tuneable multi-process SiP System & Software Power Management Optimize design for both dynamic and stand-by power Top-Bottom Power Estimation Flow Dynamic Voltage/Frequency Scaling Multiple Supply Voltage / Power gating Body-biasing technology Multiple Vt design Clock gating 16 ISLPED Keynote, B. Dennington, October 2006 Holistic approach: design technologies ? Logic is “Connected” Power is Not “Connected” Verification Scripts File translation Errors Parser Synthesis Parser Parser Logic Information (Verilog) Parser Silicon Virtual Prototype Parser Simulation Test Verification Synthesis Simulation Silicon Virtual Prototype Power Information Test (no consistency) Parser Libraries P+R Libraries P+R IP Can be Automated IP Very Difficult to Automate 17 ISLPED Keynote, B. Dennington, October 2006 Holistic approach needs co-operation! Is this an opportunity for collaboration or an area in which to compete ? No one company can do it alone 18 ISLPED Keynote, B. Dennington, October 2006 Low-power design: eChip Starting from the system issues CPU 23% Hard disk Spin down timer Voltage/ Frequency Scaling other 31% LCD 16% DC/DC 13% Memory 17% LCD backlight dimming Low Power DDR memory 20 ISLPED Keynote, B. Dennington, October 2006 Voltage/Frequency Scaling basics 50% CPU usage, MSV 100% CPU usage Performance Performance Full speed Power High f,V Time 50% CPU usage, DVFS Performance 50% speed Idle How to predict the required performance Power in advance? Energy used Optimal ! Power Low f,V Stand-by power Time Time Power savings are achieved by executing a workload at a lower frequency. 21 ISLPED Keynote, B. Dennington, October 2006 eChip: Block diagram Monitors Supply Noise Temperature ARM1176 Main facts: 0.065um Taped-out in 2005 Linux-based system Peripherals INTC, Timers, Watchdog, RTC, UART, I2C, DMAC Clock Reset Power Mngmnt AXI Control & Memory Access Networks Memory Controllers Embedded SRAM LP DDR & Static 0.5 MByte Tunnels 22 ISLPED Keynote, B. Dennington, October 2006 eChip Power Management Architecture Reg Always -On Domain V1 V2 V1 V2 Reg SOC Domain 299MHz PLL 33MHz OSC Clock Generation Unit Reg CPU Core Domain F current F target Reg Power Supply Unit IC CPU SRAM Domain AXI Interface Mode Clocks 399MHz PLL V1 V2 Operating Point Transition Control Clamp Contro l V1/V2 select LP IF Power Mode Ctrl VDD_OK timer I2C: PMU control Power Modes I2C FiFo 23 ISLPED Keynote, B. Dennington, October 2006 eChip: Example of MPEG4 operations 120 100 % 80 60 40 20 0 Time CPU usage DVFS level Simulated workload 24 ISLPED Keynote, B. Dennington, October 2006 Designing low-power products Implementation Example • 6.8M Gates + Analogue • Including memories and macros • Aggressive die-size target • 43mm2 in 90nm • 110/220 MHz target speed • Low power • Dynamic and Leakage • Multiple 3rd Party IP • Including different graphics IP • Reduced power consumption up to 35% 26 ISLPED Keynote, B. Dennington, October 2006 Implementation Example 2 This Media Processor is a complete Audio/Video/Graphics system on a chip capable of high quality software video, audio signal processing, as well as general purpose control processing. The architecture is memory centric, as every data communication occurs through writes and reads to background memory. The SoC is therefore build around the central data bus, the main memory interface, and the background memory. 27 ISLPED Keynote, B. Dennington, October 2006 Implementation Example 2 Original design based on fixed supply voltages but suited for voltage/frequency scaling. Optimisation step includes: – Partitioning in voltage domains – Closed-loop voltage/frequency scaling based on on-chip activity monitors and off-chip voltage regulators – Closed-loop process spread control. – Adaptive Back Biasing. As reference: “ideal case” assumed when we can scale voltage/frequency irrespective of the use cases. 28 ISLPED Keynote, B. Dennington, October 2006 Implementation Example 2 700 Power [mW] 639 Original pnx1500 600 Optimized pnx1500 491 Ideal min power 500 386 320 400 300 196 200 147 120 83 66 100 0 Mpeg4 -23% Mpeg2 -39% MP3 -31% 29 ISLPED Keynote, B. Dennington, October 2006 Conclusions Power is a pervasive problem Power is a problem due to technology scaling coupled with an increasing integration of features on new products, which are expected to run as usual on our old batteries for the usual low cost. Designing for low power affects all parts of the product conception and design cycle. Design teams needs experience in low-power design Cost of low-power need to be well explained and (maybe) accepted Low-power requires co-operation in the industry, nobody can do it alone! 31 ISLPED Keynote, B. Dennington, October 2006 Thank you for your attention