Low Power Design For

Embedded Systems
CPE112: Embedded Systems
Why Low Power?
• Growth of battery-powered systems
• Users need for
• Mobility
• Portability
• Reliability
• Cost
• Environmental effects

❑ Save Energy; it’s limited after all!

Sources of Power Dissipation
❑ Dynamic power dissipation
❑ Switching power dissipation
❑ Short-circuit power dissipation
❑ Static power dissipation
❑ Leakage power
Power vs. Energy
❑ Power-Critical Applications
▪ Heat Dissipation Requirement
▪ Power/Ground Metal Line Width
▪ Power/Ground Bounce due to IR drop
❑ Energy-Critical Applications
▪ Battery Lifetime
▪ Heat Dissipation Requirement
Applications for Low Power Technology
❑ Medical : Implantable hearing-aid, cardiac pacemaker
❑ Mobile Devices : cellular phone
❑ Military Devices
❑ Hard-to-access points : Space
❑ Too-many-to-access points : Sensors/Actuators in Ubiquitous World
Power Management
❑ is a prediction problem
❑ seeks to forecast whether an idle period will be long enough to
compensate for the overhead of power state changes
❑ System-level power management saves power of subsystems.
❑ If there were no overhead, power management would be trivial.
❑ a device should sleep only if the saved energy justifies the overhead.
What is Dynamic Power Management (DPM)?

❑ is a design methodology for dynamically reconfiguring systems to

provide the requested services and performance levels with a
minimum number of active components or a minimum
load on such components
❑ selective shutdown of system components that are idle or
underutilized – has proven to be a particularly effective technique
for reducing power dissipation in such systems
Break - Even Time (Tbe)
❑ minimum length of an idle period to save power
❑ depends on individual devices and is independent of requests
Power Management VS. Performance
❑ Many policies focus on power saving and do not care the
performance impact.
❑ Occasional delays are unavoidable.
 Dynamic Power Management Policies for
Embedded Systems
1. Timeout Policies
2. Predictive Policies
3. Stochastic Policies
Predictive Policies
❑ predict the length of the upcoming idle period
❑ Little knowledge of future input events.
❑ Dynamic Power Management (DPM) based on uncertain
predictions.
❑ Overprediction.
❑ Underprediction.
Timeout Policies
❑ have a timeout value, τ
❑ put the device into sleep if it is idle for than τ
❑ The basic assumption is that if the device remains idle for τ, then it
should further stay idle for at least break-even time.
❑ Drawback: waste energy while waiting for the timeout to expire
Stochastic Policies
❑ Stochastic policies model request arrival and device power state
changes as stochastic processes
❑ Minimizing power consumption and performance delays then
become stochastic optimization problems.
Techniques for
Low-Power Operation
Static Techniques
1. Fixed Timeout
2. Predictive Shutdown
3. Predictive Wakeup
Fixed Timeout
❑ Most common predictive power management (PM) policy
❑ has timeout value (τ) equal to the break-even time (Tbe) of the
device it is managing
❑ guarantees that it would not consume more than twice the energy
of an ideal offline policy
Fixed Timeout
❑ Advantages
▪ Generality
▪ Safety improved increasing timeout values.
❑ Disadvantages
▪ Tradeoff efficiency for safety.
▪ waste power waiting for the timeout to expire.
▪ Pay a performance penalty upon wakeup.
Predictive Shutdown
❑ tried to predict the length of idle time based on the computation
history and then shut down the processor if the predicted length of
idle time justified the cost
❑ exploit sleep mode operations for energy savings
❑ Solve waste power
❑ Scheme 1: Non linear regression equation
❑ Scheme 2: Based on a threshold.
Predictive Wakeup
❑ Solve the performance penalty.
❑ Wakeup when the predicted idle time expires.
❑ Increase power dissipation if TIDLE has been underpredicted.
❑ Decreases delay for servicing the first incoming request after an idle
period.
Adaptive Techniques
❑ dynamically modify timeout value (τ) based on certain parameters
❑ Several adaptive predictive techniques have been proposed
❑ Set of timeout and an index .
❑ Assigns a weight to timeouts.
❑ Only one timeout.
Stochastic Control
• Low power problem can be cast as a stochastic optimization
problem.
• PM is modeled as a Markov process.
• complex systems with many power states.
• compute PM policies that are globally optimum
• explore tradeoffs between power and performance.
Model
• Generality
• High level of abstraction
• Non deterministic
Elements of the Model
❑ Service Requestor (SR) ❑ Service Provider (SP)
▪ Send request to the SP. ▪ Serves incoming requests from a
▪ Modelled as a Markov chain workload source.
▪ Assume that the process and all its ▪ Each state is characterized by a
relevant parameters are know. performance and
power consumption level.
Elements of the Model
❑ Queue ❑ Power Manager
▪ When service requests arrive ▪ communicates with the service
during one period, they are provider and
buffered in a queue. attempts to set its state at the
▪ Several queus disciplines, most beginning of
common FIFO. each period.
▪ contains all proper specifications
and collects
all relevant information.
▪ Small power consumption
Elements of the Model
❑ Decision ❑ Cost Metric
▪ PM observes the history of the ▪ expected power consumption
system and controls the Service level performance penalty per unit
Provider by taking a decision. time (depends on the queue
▪ deterministic decision taking a length).
single action on the basis of the
history of the system.
Static Techniques
❑ The performance and power obtained by a policy are expected values.
❑ Policy optimization requires a Markov model for SP and SR.
❑ Can be safely assume that the SP model can be pre-characterized, but not
always the SR.
❑ Power consumption of the PM is not always negligible.
❑ Policy implementation may not be straightforward.
❑ Assumes a complete a priori knowledge of the system and its workload
(SR) and workloads are often nonstationary.
Adaptive Techniques
❑ based on three simple concepts
▪ policy pre-characterization
▪ parameter learning
▪ policy interpolation
❑ Two parameter Markov model for workload
❑ Policy pre-characterization constructs a two-dimensional (2D) table.
❑ The table is filled by computing optimum policies under different
workloads.
Adaptive Techniques
❑ Average techniques to estimate workload parameters based on past
history.
❑ Parameter values are used to obtain the power management policy.
❑ If estimated parameter do not correspond to value in the table gt
policy interpolation.
Stochastic VS. Predictive Techniques
❑ Captures the global view of the system.
❑ It enables the exact solution.
❑ Exploits the strength and optimality of randomized policies.
❑ But more complicated.
Thank you for listening.
Reference
❑ https://www.powershow.com/view1/1a48db-
ZDc1Z/Low_Power_Design_For_Embedded_Systems_powerpoint_ppt_pr
esentation

❑ https://cseweb.ucsd.edu/~gdhiman/Gaurav_files/CSE-
237A/TopicResearch/DPMPolicies.htm