Carnegie Mellon A System Design And Build Course

Carnegie Mellon A System Design And Build Course On Wearable Computers Dan Siewiorek and Asim Smailagic Carnegie Mellon University MSE ‘ 01, Las Vegas, June 2001

Carnegie Mellon Overview 1. 2. 3. 4. 5. 6. 7. Introduction Approach Electronic Design in a Multidisciplinary Project Design Methodology Power Measurements Evaluation Conclusions

Carnegie Mellon Introduction A System - Level design approach to the power and performance of CMU’s wearable computers dedicated to speech processing the Speech Translator Smart Modules Power consumption has to be considered in all phases of system design

Carnegie Mellon Introduction We examine the impact of processor speed, memory size, and type of secondary storage on power consumption and performance We see ahead a world of near-zero energy / weight / cost mobile systems

Carnegie Mellon Approach Experimental framework includes a family of CMU wearable computers dedicated to speech processing - Smart Modules They perform speech recognition, language translation, and speech synthesis

Carnegie Mellon

Carnegie Mellon Multidisciplinary Project Five major factors in a portable electronic system include: • Functionality • User Interface • Physical Form Factor • Power • Sensors

Carnegie Mellon Multidisciplinary Project • Decisions made in one design discipline affect decisions in another discipline • The impact can be measured if the cost of the design decision can be reduced to a common “currency” • In mobile electronic systems that “currency” is power consumption

Carnegie Mellon Major Factors in Portable Electronic Systems and their Relationship to the Design Disciplines

Carnegie Mellon Smart Module Architecture The core of the smart module is the Cardio processor card, combining the processor and motherboard chips The Cardio also supports two serial ports for communication between the modules and a VGA interface

Carnegie Mellon Smart Module Hardware Diagram -Processor CARDIO Keyboard Mouse VGA -Memory -Chipsets IDE PCMCIA HD ISA ESS 1888 Serial (For debugging only) Ports Communications To other Modules

Carnegie Mellon Smart Module Functional Prototype

Carnegie Mellon Optimized Smart Module

Carnegie Mellon Approach The code was profiled and tuned Profiling identified “hot spots” for HW and SW acceleration, and places to reduce computation and storage requirements

Carnegie Mellon Power Measurements Power is consumed by • Processor • Memory • Disk • Sound Chip • Serial Ports

Carnegie Mellon Power Measurements The power and performance measurements were taken using a body of 10 English test sentences for the Speech Recognizer and Language Translator, and their 10 Croatian translations for the Speech Synthesizer

Carnegie Mellon Power Measurements A power profile for the Smart Modules included: • Idle Mode: processor is at nearly 0% usage • Full On Mode: processor is at nearly 100% usage Each state transition has an associated latency value

Carnegie Mellon Power Measurements States for spinning disk drive: • Power Down Mode: power to the disk drive is shut off - when the module is in Suspend Mode • High Spin Mode: disk drive is being accessed • Low Spin Mode: power-saving mode

Carnegie Mellon Power Measurements State diagrams for both processor and disk were combined to produce a new model for power consumption

Carnegie Mellon New Model of Power Consumption Over Time Full on & High Spin Power (W) Idle & High Spin Full On & Low Spin Idle & Low Spin Suspend Time

Carnegie Mellon Power Consumption Profile Power consumption of the Speech Recognizer module over time, using the 586 -based Cardio and a spinning disk drive, was measured

Carnegie Mellon Power Consumption of Speech Recognition Module Over Time Full Spin Up Suspend Spin Down Spin Up Spin Down Close TCP Link & Spin Up

Carnegie Mellon

Carnegie Mellon Performance Comparison The metric for comparison is proportional to the processing power, and inversely proportional to the product of volume, weight and power consumption

Carnegie Mellon Performance Values for Wearable Computers Name TI 6030 Spec. Int Volume (in 3) Weight ( lbs ) Power (watts) R (V*W*P) Log of Normalized 175. 00 260. 00 7. 50 36. 00 70200. 00 TIA-P 55. 00 88. 00 3. 00 6. 50 1716. 00 1. 11 TIA-O 55. 00 45. 00 2. 50 4. 50 506. 25 1. 64 175. 00 45. 00 33. 00 2. 13 1. 50 4. 00 382. 50 198. 00 2. 26 2. 55 SR-SM OPT-SM

Carnegie Mellon Composite Performance of Speech Recognition Wearable Computers 5 4 Norm Log [perf/(vol*wt*power)] 3. 5 3 Goal: 2. 8 2. 5 2 1. 5 Normalized Performance 1 0. 5 0 0 1 2 3 Evolution of Systems 4 5 6

Carnegie Mellon Spot: SA-1110 Based Wearable Computer • New wearable computer as a research platform, including context aware computing • Low power design • 233 MHz Strong. ARM 1110 • 256 MB DRAM • Mobile, throttleable platform that can trade performance for energy

Carnegie Mellon Spot Specifications

Carnegie Mellon Spot Wearable Computer

Carnegie Mellon Spot

Carnegie Mellon Conclusions Results show that there are orders of magnitude improvement from one generation of wearable computers to the next A system-level approach to power / performance optimization improved the metric by over a factor of 300 through the four generations, and over 400 through the five generations

Carnegie Mellon Conclusions Peak demand by an application can often determine the battery life rather than average demand Audio-centric interfaces exhibit high demand “spikes, ” potentially causing significantly reduced battery life

Carnegie Mellon Summary • The options for the design space included: – Two processor architectures (Intel 486, 586) – Three different processor speeds – Two different main memory sizes – Two different secondary storage types • The best configuration for performance is the 586/100/32 Flash, and the 486/75/16 Flash for power

Carnegie Mellon Power Consumption of Speech Recognition Module • Our research indicates that the peak demand of an application can often determine the battery life rather than the average demand • Audio-centric interfaces exhibit high demand spikes • By filling in the valleys, it would be possible to cut the peak demand in half and thus significantly extend battery life