Low Power Systems

advertisement
LOGO
Low Power Solutions:
A System Design Perspective
Nik Sumikawa
Nik Sumikawa
Contents
1
Low Power: Why?
2
Standard Embedded Solutions
3
Innovative Solutions
4
Solutions for Mobile Platforms
www.themegallery.com
Low Power: Why?
Power vs. Performance
Technology Scaling
 VLSI
 Embedded
Technology Trend
 Green Stimulus
 Scaling Size
Company Logo
www.themegallery.com
What You Should Think About
Low power design strategies
 Components: Microcontrollers, peripherals,
ect.
 Low power design with hardware
 Low power design with software
 Low power design in mobile device
Company Logo
www.themegallery.com
Low Power Embedded Systems
TELOS:
 Low power wireless
embedded system
 Low duty cycle
principle
 Minimizes dynamic
power consumption
Company Logo
Nik Sumikawa
Low Duty Cycle Principle
Timer or
Interrupt event
Process
Wake
Up
Sleep Mode
Deep
Sleep
Sleep
Prep
Nik Sumikawa
Low Duty Cycle
Low processing to sleep ratio
 Extended sleep period
Responsively:
 fast wake-up and sleep times
Minimize Interrupts:
 Context switching overhead
Company Logo
Nik Sumikawa
Low Duty Cycle: DMA
Direct Memory Access (DMA):
 Controls bus and transfers data with minimal
processor overhead
Significance
 Transfer data while sleeping
 Minimize processor overhead
Company Logo
Nik Sumikawa
Low Duty Cycle
Fails with significant
processing
Alternatives:
 Dynamic Voltage and
Frequency Scaling
(DVFS)
 Dynamic Power
Management (DPM)
Image:
http://www.domainmagnate.com/wpcontent/uploads/2009/03/failure-success.jpg
Nik Sumikawa
Dynamic Power
Design Variables
Energy Source
Capacitance
Frequency
Dynamic
Power
Voltage
P = CVdd2f
Battery
www.themegallery.com
Reducing Dynamic Power
 Dynamic Voltage and Frequency Scaling
 Scale voltage when sleeping/Idle
 Voltage term quad. proportional to power
 Reduce frequency
 Minimize line capacitance
 Long traces have large capacitance
Company Logo
www.themegallery.com
Dynamic Power Management
Generalize power
management
Multiple policies
 Single-policy
 Multiple-policy
 Task-scaling
Rajami and Brock [2]
Nik Sumikawa
Single-policy Strategy
Idle Scaling (IS)
 Operate at full speed when processing
workload
 Reduce the frequency and voltage when idle
Goal:
 Reduce the CPU and bus frequencies
 Meet continuous DMA requirements
 Provide acceptable latency when resuming
from idle
Rajami and Brock [2]
Nik Sumikawa
Multi-policy Strategies
Load scaling (LS):
 Balance system operating point with current
or predicted processing demands
 Run system with minimal idle time
Other:
 Manage systems state based on status of the
systems energy source
Rajami and Brock [2]
Nik Sumikawa
Task-scaling Strategies
Application scaling (AS):
 Used for workloads that are difficult to power
manage
• Audio and video processing
• Begin processing next sample immediately
 Operate a lower operating point
 Increases to higher operating point when it
begins to fall behind.
Rajami and Brock [2]
Nik Sumikawa
Results of DPM
 IS: Idle Scaling
LS: Load Scaling
AS: Application Scaling
 Frame-Scaling (FS): perfect knowledge of processing requirements of video frame
Rajami and Brock [2]
Nik Sumikawa
Too Many Low Power States
Disadvantages:
 Confusion
 Wrong low power
state
Solution:
 Minimize the number
of state
 Decrease complexity
Image:
http://kunaljanu.files.wordpress.com/2009/02/
ist2_1457667confusion-1.jpg
Nik Sumikawa
Sources of Power Consumption
Microcontroller
Bus architecture
 On chip communication
 External communication
Memory hierarchy
Peripherals
Rajami and Brock [2]
Nik Sumikawa
Communication Architectures
Advanced Microcontroller Bus Architecture
 ARM bus protocol for system-on-a-chip (SOC)
 Advanced High Performance Bus (AHB)
• Pipelined
• Memory mapped
• Up to 16 masters, 16 slaves
 Advanced Peripheral Bus (APB)
• Non pipelined
• Single master, up to 16 peripherals
Rajami and Brock [2]
Nik Sumikawa
AMBA On-chip Bus
Rajami and Brock [2]
Nik Sumikawa
Power Profiling
86% power consumed by logic
14% power consumed by bus lines
Rajami and Brock [2]
Nik Sumikawa
Power Reduction Techniques
Power Management
 Shut down bus interfaces to idle slaves
Bus Encoding
 Reduces # of line transitions, but not bus
transactions
Traffic Sequencing
 Reduce multiple masters interleaving bus
access
Rajami and Brock [2]
Nik Sumikawa
Power Reduction Techniques
No technique achieves large saving alone
Rajami and Brock [2]
Nik Sumikawa
Power vs Energy
Power is amount of energy over an
amount of time (Watts = Joules / second)
Battery provides finite amount of energy
 Goal: minimize energy use, not just power
In mobile systems we care about energy
 Budget energy to prolong battery life
Rajami and Brock [2]
Nik Sumikawa
Static System Optimization
Compiler techniques
 Instruction energy consumption profiling
• Done empirically
 Instruction reordering
•
•
•
•
Without affecting correctness
Improve register utilization
Reduce memory accesses
Reduce pipeline stalls
Nik Sumikawa
Static System Optimization
Code Compression





Post compilation static optimization
Reduces storage size of instructions
Can have a large impact
Requires complex design space exploration
Goal for mobile system: reduce power
consumption while preserving performance
Nik Sumikawa
Code Compression Challenges
Random access decompression
 Defining decodable block beginnings
 Jump to new locations in program without
decoding all blocks between
Solutions
 Begin compressed blocks on byte boundaries
 Store translation table
• More efficient the compression, larger the table
 Recalculate branch offsets to compressed
addresses
Nik Sumikawa
Code Compression Requirements
Additional hardware
 Additional memory to store table
 Decompression unit
Design decisions
 Table generation/lookup
 Compression technique
Nik Sumikawa
Code Compression Implementation
SPARC ISA
Optimize consumption of complete SOC
Multiple iterations on binary
Instructions split into 4 categories




Group 1: immediate instructions (code = 0)
Group 2: branch instructions (code = 11)
Group 3: dictionary instructions (code = 100)
Group 4: uncompressed instr (code = 101)
www.themegallery.com
Diagram
Update branch offsets
Optimized Binary
Phase 4
Branch compression
Phase 3
Immediate compression
Phase 2
Markov model
Compiled
Binary
Phase 1
Company Logo
Nik Sumikawa
As a Result…
Bus Compaction
 Instructions transmitted no longer require
entire bus
 Use the extra lines to transmit the next
compressed instruction
Nik Sumikawa
Decompression Architecture
Pre-Cache
 Decompression engine between
memory/cache
Post-Cache
 Decompression engine between cache/cpu
Nik Sumikawa
Simulation
Full SOC simulation
7 sample apps run
Nik Sumikawa
Results
Nik Sumikawa
INCLUDE?
Nik Sumikawa
Results
Net energy saving observed
 22-82% power savings from code
compression
 What about additional hardware?
Bonus
 Increased performance
 Reduced area
Nik Sumikawa
Verdict
Static power optimization
 Potentially large payoff for little preprocessing
Still more sources of consumption
 We’ve observed SOC savings
 What about peripherals?
Nik Sumikawa
Energy Budget
Voice Call
SMS
Energy
Budget
Emails
Pictures
localization
Nik Sumikawa
Energy Budget: Localization
How much of the energy budget
should be given to localization?
 Depends on the user
Grant increase allotment when
localization is a higher priority
Nik Sumikawa
Localizations Methods
1
2
3
GPS
GSM
WiFi
• Very accurate
• Power Hungry
• Lower accuracy
• Lower power
requirement
• Mod. Accurate
• Mod. Power
requirement
Nik Sumikawa
Power vs. Precision
Localization
Power:
Precision:
amount of energy
required by peripheral
in order to determine
location
Accuracy of the device
used for localization
Constandache, Gaonkar, Sayler,
Choudhury, Cox [3]
www.themegallery.com
Power Consumption
 30 Second sampling intervals
 Power Consumption:
 GPS: High baseline
 WiFi: Low baseline with high spikes
 GSM: Low baseline with varying spikes
Constandache, Gaonkar, Sayler,
Choudhury, Cox [3]
www.themegallery.com
Power Consumption
 30 Second sampling
intervals
 Results:
 GPS: increased
baseline
Company Logo
www.themegallery.com
Localization Accuracy
 Accuracy varied
based on location
ALE: Average
Location Error
 Wifi and GSM
oversampled
Company Logo
www.themegallery.com
Diagram
Add Your Text
Add Your Text
Add Your Text
Add Your Text
Add Your Text
Add Your Text
Add Your Text
Add Your Text
Company Logo
www.themegallery.com
Diagram
Add Your Text
Add Your Text
Add Your Text
Company Logo
www.themegallery.com
Diagram
Add Your Title
Add Your Title
Add Your Title
ThemeGallery is a
Design Digital
Content &
Contents mall
developed by
Guild Design Inc.
ThemeGallery is a
Design Digital
Content &
Contents mall
developed by
Guild Design Inc.
ThemeGallery is a
Design Digital
Content &
Contents mall
developed by
Guild Design Inc.
Company Logo
www.themegallery.com
Diagram
Text
Text
Add Your Title
Text
Text
Company Logo
www.themegallery.com
Cycle Diagram
Add Your Text
B
Microcontroller
A
C
Add Your Text
Sources
Add Your Text
E
D
Add Your Text
Company Logo
www.themegallery.com
Diagram
2001
2002
2003
2004
Company Logo
www.themegallery.com
Progress Diagram
Phase 1
Phase 2
Phase 3
Company Logo
www.themegallery.com
Diagram
Company History
2004
2001
2003
2004.01 Add Your Text
2004.03 Add Your Text
2004.05 Add Your Text
2001.10 Add Your Text
2001.10 Add Your Text
2001.10 Add Your Text
2002
2003.10 Add Your Text
2003.10 Add Your Text
2003.10 Add Your Text
2002.10 Add Your Text
2002.10 Add Your Text
2002.10 Add Your Text
Company Logo
www.themegallery.com
Map
Company Logo
www.themegallery.com
3-D Pie Chart
Text2
Text3
Text1
Text4
Text5
Company Logo
www.themegallery.com
Block Diagram
Add Your Text
Add Your Text
concept
Concept
Concept
Concept
Company Logo
LOGO
www.themegallery.com
Download