2. Outline
Introduction
Energy-Aware Processors
OS Energy Management
Power Reduction through Architectural
Design
Reducing Power through Circuitry
Smart Batteries
Tools and Packages for Low-Power
Design
3. Introduction
Mobile devices have limited battery
power
Remote areas to charge batteries
Battery life is only expected to grow
10-20% over the next 10 years
Software solutions
Hardware solutions
Batteries
Circuitry
5. Dynamic Voltage/Frequency
Scaling
One of the more popular power
management schemes
Easy implementation
Performance suffers
Most effective when hardware is more than
adequate for the tasks being run
In resource-constrained mobile computing
systems with slower processors there may not
be much room DVFS techniques to save
energy
6. Scheduling Algorithms
Classic computer science problem
Mobile devices add to the complexity when
battery energy utilization is considered
Algorithms are tailored to the specific
needs of the device
Energy saving vs. processing performance or a
combination of the two
7. Computational Offloading
Resource-limited computing device defers
computation to a more capable device
A solution for Mobile Wireless Ad-Hoc
Networks (MANETs)
Distribute computational tasks amongst
devices
Improvement in both energy saving and
processing performance
8. OS Energy Management
Dynamic Software Management (DSM)
Frameworks
Constrained Power Management (CPM)
9. Dynamic Software
Management Frameworks
Goals
Improve battery utilization
Avoid competition for limited energy
resources from multiple applications
Energy demand prediction
Critical component of framework
Uses energy models for prediction of
application energy demand
10. Constrained Power
Management (CPM)
Goals
Manage both system resources and power of
an entire platform
Can optimize energy or performance
Quality of Service (QoS)
CPM collects QoS requirements from
applications and coordinates device drivers to
support requested QoS
11. Power Reduction through
Architectural Design
Caches
Typically 30-60% of total processor area
20-50% of processor power consumption
Filter Caches & On-demand Selective
Cache
Speculative Activation
On-chip Memory
12. Filter Caches & On-demand
Selective Cache
Filter Caches
Reduce dynamic and static power
consumption
Shrink the size of L1 and L2 cache
Significant decrease in performance
On-demand Selective Cache
Shuts down parts of the cache according to
demand
Suffers from same decrease in peformance
13. Speculative Activation
Attempt to make a prediction of where
the required data may be located
If prediction is correct, latency and dynamic
power consumption become similar to direct-
mapped cache
If prediction is incorrect, cache is accessed
twice and results in additional latency
14. On-chip Memory
Minimize external data communication
Eliminates external frame buffers
No off-chip memory support allows for
high-quality processing at a minimal
power level
15. Reducing Power though Circuitry
Near-Threshold Computing
Razor-Thin Margins
Smaller Transistors
All-Digital Phase Locking
Smart Converters
Next-Generation Dynamic Ram
SmartReflix power and performance
16. Near-Threshold Computing
Lower voltage
Pros: Energy consumption
Cons: performance loss, memory and
logic failures
Moore’s law
Energy efficiency problem
32nm chip example
Solution to improve performance
17. Razor-Thin Margins
Run processors at the threshold voltage
Reasons why chips run at higher voltages
Incorporate error detection
Research from U. Of Michigan
Arm Cortex –M3
○ 60% more energy efficient
19. Smaller Transistors Cont.
37% increase in
performance
Moore’s law
Lower voltage
50% less power
Less leakage
20. All-Digital Phase Locking
Phase-locked loops
Lock in and track input signal
Pace car analogy
Used for synchronize the processor and
the clock
Converting from analog to digital saves
power
21. Smart Converters
The analog to digital converts use a
switched capacitor
Oregon State University
Loop of inverters
Chiao Tung University in Taiwan
Separate signal in 2 stages
○ 1 simple processing
○ 2 fine tuning
22. Next Generation of Dynamic Ram
Samsung and Hynix on DDR4 ram
Drop voltage from 1.5V to 1.2V
Includes better clocking
Faster algorithms for encoding data
End result
Consume less energy
Higher bandwidth
23. Smart Batteries
Background Info
Memory effect
○ Solution: lithium ion
Advanced Power Management(APM)
Mobile device uses information
Polling for suspicious activity
Avoid stalling of devices and wasted energy
24. SmartReflix power and performance
Static Techniques
Power-gating
Dynamic Techniques
Dynamic power switching
Adaptive voltage scaling
Dynamic voltage/frequency scaling
25. SimpleScalar (Tools)
Tool set
Simulator to compare two different
machines
Extended to the Power Analyzer
For the ARM platform
Exam the performance/power trade-offs
Extended to the Wattch Project
Web app
By adding parallel processing this could help in preformance
The one side vs 3 side Tri gate transistor
http://www.treehugger.com/gadgets/intel-announces-revolutionary-3d-transistors-50-more-energy-efficient-than-previous-generation.html - ideal for use in small handheld devices, which operate using less energy to "switch" back and forth between states. -200 mV lower voltage compared to the -the lower the voltage the higher the leakage but in this design there is less leakage at lower voltages compared to the 2d http://news.cnet.com/8301-13924_3-20059431-64.html http://www.anandtech.com/show/4333/intels-silvermont-a-new-atom-architecture
Pace car analagy- http://en.wikipedia.org/wiki/Phase-locked_loop The All-digital phase locking are easier to make Paper- http://web.mit.edu/~bdaya/www/All%20Digital%20Phase%20Locked%20Loop%20Design%20and%20Implementation.pdf
Power gating: “The basic strategy of power gating is to provide two power modes: a low power mode and an active mode. The goal is to switch between these modes at the appropriate time and in the appropriate manner to maximize power savings while minimizing the impact to performance.” http://nanocad.ee.ucla.edu/pub/Main/SnippetTutorial/PG.pdf Dynamic power switching- basically the same as power gating as it determines when a device has completed its current computational tasks and, if it's not needed at the moment, then puts the device into a low power state Adaptive voltage scaling-by reducing the voltage you can gain up to 60% less energy consumption but you need to understand the performance at each voltage level. Once this information is gained then the voltage and be scaled based on what the device is doing http://www.ti.com/lit/ml/slyb186/slyb186.pdf Dynamic Vol/Freq Scaling- Similar to the past examples they will be scaled based on the activity of the device
http://www.ecs.umass.edu/ece/koren/architecture/Simplescalar/SimpleScalar_introduction.htm Has high-level function and low –level detailed simulator users can build modeling applications that simulate real programs running on a range of modern processors and systems. Power Analyzer- University of Michigan and University of Colorado Wattch project- Princeton University- would connect to the system and estimate power that the system had consumed. Also has multi meter results from the CPU
http://notrump.eecs.umich.edu/papers/pscope99.pdf http://web.cs.wpi.edu/~emmanuel/courses/cs525m/S06/slides/powerscope_wk4.pdf The data collection phase uses a multi-meter [34] connected by a Hewlett-Packard interface bus (HPIB) connection to a data collection computer running an energy monitor component. A profiling computer running an application and a system monitor component are getting power by being connected to the multi-meter for measurement. The profiling computer also provides the trigger for the multi-meter to start measuring.
The offline analysis phase uses a third component called the energy analyzer, which will take the results produced in the data collection phase by the system monitor and the energy monitor and correlate the two to produce the energy profile of the application.
With this information the user can focus on the process that are consuming the most amount of power. In some cases there were 50% power reduction with minimal loss of performance