power-aware systems
TRANSCRIPT
Power-Aware Systems
Manish Bhardwaj, Rex Min and Anantha Chandrakasan
Massachusetts Institute of Technology
November 2000
Power-awareness: Intuitive Notions
n Motivation: Maximize lifetime of energy constrained systems ≈ Maximize system-level energy efficiency
n Implication: Given an operating scenario, consume only as much energy as the scenario demands
n Alternately, scale the power consumed in response to changing scenarios (power-awareness)
Agenda
n Key questionsoWhat are operating scenarios?oHow well are these systems tracking their scenarios?oWhat can we do to improve this tracking?oWhat are the costs and benefits?
n AbstractionsoAwareness dimensions, operating scenarios, energy curves,
scenario distributions
n Formalizing Power-Awareness
n Enhancing Power-Awareness
n Examples: oMultipliersoRegister Fileso FiltersoAnalog-Digital Converterso Variable-Voltage ProcessorsoWireless Networks
Abstractions: Scenarios
n Over any specified time interval, the energy consumed by a system is governed by five key dimensions
n Scenarios are characterized by precisely these dimensions
n Scenario ≡ <Input, Output Quality, Latency, State, Environment>
n Choices in specifying scenarioso Number of dimensions to includeo Detail with which the dimension is captured
n Example: Characterizing scenarios in a 16x16-bit multiplier
1. InputStatistics
5. Environment
4. State2. Desired
OutputQuality
3. Tolerable Latency/Desired Throughput
Awareness Dimensions
Scenario Characterization in Multipliers
n Input dimension onlyo Scalar m: Specifies a maximum precision requirementoUnordered pair (m, n): Specifies a mxn-bit multiplicationoOrdered pair <m, n>oOrdered operands <X,Y>
n Input and state oOrdered operands and previous operands <X[n],Y[n],X[n-1],Y[n-1]>
n Input, state and desired precision
n Input, state, desired precision and latency
Abstractions: Energy Curves
n The energy consumed by a system as a function of its scenario, E(H, s)
Abstractions: Scenario Distributions
n The probability that a system will reside in a certain scenario is captured by scenario distributions, dS(s)
Perfect Power Awareness
n Perfect energy curve obtained by constructing dedicated point systems
A system is termed perfectly power-aware iff it consumes only as much energy as its current scenario demands.
Perfect Systems
n A system that would result in Eperfect is termed the perfect system (Hperfect)
n If scenario detection and interconnect costs were zero, the system above would yield Eperfect
H s1
H sS ||
H s2
H si
DEMUX
Scenario DeterminingUnit
Dedicated Point Systems
Input Output
MUX
Quantifying Power Awareness
n The relative energy curve is simply the energy curve of a system normalized to the perfect energy curve
Power Awareness Metric
n Reduce the relative curve to a single number by appropriate weightingoWeigh by probability of occurrence of scenariooWeigh by energy dissipated in the scenario
n Physical interpretation: Expected system lifetime normalized to lifetime of perfect system
n Defined w.r.t scenario distribution and a set of point systems
n Metric leads to complete ordering for a specified distribution and partial ordering otherwise
1
)(),(
)(),(
),()(
),()()(−
∈
∈
∈
∈
=
=
∑∑
∑∑
ScenariosiiSiperfect
ScenariosiiSi
ScenariosiiiS
ScenariosiiiSi
sdsHE
sdsHE
sHEsd
sHEsdsηφ
Enhancing Power-Awareness: Ensemble Construction
n What is the optimal ensemble of point systems?
1x1
2x2
16x16
X Y X.Y
Zero Detection Circuit
1x11x1
2x22x2
16x1616x16
X Y X.Y
Zero Detection Circuit
16x16
X
Y
X.Y16x1616x16
X
Y
X.Y versus
Formal Statement of the Problem
n Given:o Function to be realized (F)oConstraints to be met (C)oA set of point systems (P)oA scenario distribution (d)
n Form of the solution:oAn ensemble of point systemsoA scenario to point system mapping
n Measure of the solution: Power awareness
n Problem: Find the solution with the highest measure
n Appears to be unsolvable in polynomial time
n (Greedy) Heuristics seem to work well
n Can be generalized to temporal and spatial-temporal ensembles
A Near-optimal 4-point Ensemble
Power-Awareness = 0.92
16x16
14x14
11x11
9x9
Zero Detection Circuit
X
Y X.Y
Power-Aware Register Files
n MotivationoArchitecture trends point to increasingly energy-hungry fileso Processors typically access only a fraction of registers over typical
instruction windowsoWhy pay the energy price of full file access?
n Objective: Register access energy must scale with the number of registers being accessed over an instruction window
n Scenario: Number of distinct registers accessed in an instruction window of specified length
n Available point systems: 1, 2, 4, 8 … word register files
Scenario Distributions
>70% of the time, <16 registers accessed in a 60 instruction window
Window Locality
>85% of the time, <5 registers change from window to window
Candidates
32 registers
Bank-0 (4 registers)
Bank-1 (4 registers)
Bank-2 (8 registers)
Bank-3 (16 registers)
Bank Select Logic
Address Data
Address Data
Monolithic File Segmented File
Power-Awareness Comparisons
Power-Awareness Increases by 2-3x
Power-Aware Digital Filters
n Motivation: oAdaptive filters used in communications applications dissipate
significant energyo Filtering requirements change with desired quality and channel
conditionsoWhy run the filter at maximum precision and taps?
n Objective: Energy consumed by a filter must scale with the word-length precision and taps
n Scenarios: <Desired Taps, Desired Precision>
n Point systems: All <m taps, n bits> filters
Scenario Distribution
Candidates
64-tap, 24-bit FIRX[n] Y[n]
51-tap, 10-bit FIR
58-tap, 20-bit FIR
64-tap, 24-bit FIR
64-tap, 15-bit FIR
Arbiter
X[n] Y[n]
51-tap, 10-bit FIR
58-tap, 20-bit FIR
64-tap, 24-bit FIR
64-tap, 15-bit FIR
Arbiter
X[n] Y[n]
43-tap, 23-bit FIR
43-tap, 13-bit FIR
64-tap, 7-bit FIR
30-tap, 17-bit FIR
Monolithic Filter
Optimal 4-point Ensemble Optimal 8-point Ensemble
Monolithic Filter
Power-Awareness = 0.51
4-point Ensemble
Power-Awareness = 0.82
8-point Ensemble
Power-Awareness = 0.90
Perfect System
Power-Awareness = 1.0
Power-Aware Processors
n Motivation:o Processor workloads vary significantlyo Tremendous energy savings by spreading workload to occupy
all available time (by lowering Vdd and operating frequency)oWhy pay the energy price of a full workload?
n Objective: Energy consumed by a processor should scale with its workload requirement
n Scenarios: Workload (∈ [0,1])
n Point systems: Processors with Vdd, frequency customized for a workload
Candidates
0ddV
SA-1100
maxddV
DEMUX
Workload DeterminingUnit
Input Output
MUX
1ddV
SA-1100
SA-1100
SA-1100
VddBuck
Regulator
Controller+
Prog. Logic
Vddmax
µ-OS
Desired Supply Voltage (Digital Value)
VariableVdd
SA-1100SA-1100
VddBuck
Regulator
Controller+
Prog. Logic
Vddmax
µ-OSµ-OS
Desired Supply Voltage (Digital Value)
VariableVdd
SA-1100
Vdd
Fixed Voltage Processor
Dynamic Voltage Processor
Power-Awareness Comparisons
DVS 1.6x more power-aware than fixed-voltage system
Analog-Digital ConvertersContributed by Kush Gulati, MIT [ISSCC’01]
n Motivation:oA/Ds have non-trivial system-level power-budgetsoUser/algorithms might be able to tolerate low quality
(resolution)o Signal statistics might allow variable sampling rates
n Objective: Conversion energy must scale with the desired sampling rate and resolution
n Scenarios: <Rate, Resolution>
n Point systems: All <Rate, Resolution> converters
Candidates
Conventional A/D Power-aware A/D
ResolutionSamplingRate
ReconfigurableCore
AnalogInput
DigitalOutput
Scenario Diversity in A/Ds
Power versus Sampling Rate for different Resolutions
0.01
0.1
1
10
100
1.00E+03 1.00E+04 1.00E+05 1.00E+06 1.00E+07 1.00E+08
Output Data Rate (Hz)
An
alo
g P
ow
er C
on
sum
pti
on
(m
W)
(6)
(12)
(10)
(16) (14)
(8)
Power-Awareness Comparison
Power-Awareness increases from 0.31 to 0.81
0
1
2
3
4
5
6
7
8
9
10
70 75 80 85 90 95
SNR (dB)
Po
wer
(mW
)
Reconfigurable Converter
Unaware Converter
Wireless Data-Gathering Networks
B
R
S0
S1
S2
123
4
5
6
789 10
n Energy constrained nodes deployed to observe a source in a specified region
Power-Aware Wireless Networks
n Motivation: oKey challenge in data-gathering networks is energy efficiencyoNetworks exhibit tremendous operational diversity (topology,
source behavior, desired quality, environmental conditions, instantaneous state)
n Objective: Data gathering energy should scale with desired quality, environmental conditions and internal state
n Scenarios: <Environmental Noise, Energy Vector>
n Point systems: All <Noise, State> protocols
Environmental Awareness
Protocol is potentially 10x more power-aware!
Awareness to State
Protocol 2x more power aware than unaware versions
Summary
n Power-aware design can significantly enhance lifetime of battery constrained systems
n Power-awareness is a system-wide design philosophy
n Systematic methodology for power-aware design:oCharacterize scenarios by understanding the awareness
dimensions of a domainoGather statistics and construct scenario distributionsoConstruct optimal ensemblesoMeasure power-awareness o Iterate
n Power-aware design is NOT low-power designo Low power design focuses on engineering point systemso Power-aware design focuses on characterizing and harnessing
diversity by actively adapting the system