POWER DISSIPATION
The past
the major concerns of the VLSI designer were area performance cost and
reliability power considerations were mostly of only secondary importance.
In recent
years how ever this has begun to change and increasingly power is
being given comparable weight to area and speed.
Several
factors have contributed to this trend Perhaps the primary driving factor has
been the remarkable success and growth of the class of personal computing
devices portable desktops audio and video based multimedia products and
wireless communications systems.
Personal
digital assistants and personal communicators which demand high speed
computation and complex functionality with In the past the major concerns of
the VLSI designer were area performance cost and reliability power
considerations were mostly of only secondary importance.
In recent
years however this has begun to change and increasingly power is being given
comparable weight to area and speed.
Several
factors have contributed to this trend Perhaps the primary driving factor has
been the remarkable success and growth of the class of personal computing
devices portable desktops audio and video based multimedia products and
wireless communications systems personal digital assistants and personal
communicators which demand high speed computation and complex functionality
with low power consumption
There
also exists a strong pressure for producers of high end products to reduce
their power consumption.
1. Software Level Power
dissipation
The first
task in the estimation of power consumption of a digital system is to identify
the typical application programs that will be executed on the system.
A non
trivial application program consumes millions of machine cycles making it
nearly impossible to perform power estimation using the complete program at say
the RT level.
Most of
the reported results are based on power macro modeling_ an estimation approach
which is extensively used for behavioral and RTL level estimation.
In the
power cost of a CPU module is characterized by estimating the average
capacitance that would switch when the given CPU module is activated.
In the
switching activities on address instruction a n d data b u s e s are used to
estimate the Power consumption of
the microprocessor, based on actual current measurements of some processors
Tiwarietal present the following instruction instruction level power model
Where
Energy p
is the total energy dissipation of the program which is divided into three
parts The first part is the summation of the base energy cost of each
instruction, BCi is the base energy cost and Ni is the number of times
instruction i is executed.
The
second part accounts for the circuit state SCij is the energy cost when
instruction i is followed by j during the program execution. Finally the third
part accounts for energy contribution OCk of other instruction effects such as
stalls and cache misses during the program execution.
In Hsieh
et al present a new approach called profile driven program synthesis to perform
RT level power estimation for high performance CPUs.
Instead
of using a macro modeling equation to model the energy dissipation of a
microprocessor the authors use a synthesized program to exercise the
microprocessor in such a way that the resulting instruction trace behaves in
terms of performance and power dissipation much the same as the original trace.
The new
instruction trace is however much shorter than the original one and can hence
be simulated on a RT level description of the target microprocessor to provide
the power dissipation results quickly Specifically this approach consists of
the following steps
§ Perform
architectural simulation of the target microprocessor under the instruction
trace of typical application programs
§ Extract
characteristic p role including parameters such as the instruction mix
Instruction data cache miss rates branch prediction miss rate pipeline stalls
etc for the microprocessor.
§ Use mixed
integer linear programming and heuristic rules to gradually transform a generic
program template into a fully functional program.
§ Perform
RT level simulation of the target microprocessor under the instruction trace of
the new synthesized program.
Notice
that the performance of the architectural simulator in gate vectors second is
roughly to orders of magnitude higher than that of a RT level simulator.
This
approach has been applied to the Intel Pentium processor which is a super
scalar pipelined CPU with KB way set associative data instruction and data
caches branch prediction and dual instruction pipeline demonstrating to orders
of magnitude reduction in the RT level simulation time with negligible estimation
error.
2. Behavioral Level
Power dissipation
Conversely
from some of the RT level methods that will be estimation techniques at the
behavioral level cannot rely on information about the gate level structure of
the design components and hence must resort to abstract notions of physical
capacitance and switching activity to predict power dissipation in the design.
3. Information
Theoretic models
Information
theoretic approaches for high level power estimation depend on information
theoretic measures of activity .for example entropy to obtain quick power
estimates.
Entropy
characterizes the randomness or uncertainty of a sequence of applied vectors
and thus is intuitively related to switching activity that is if the signal
switching is high it is likely that the bit sequence is random resulting in
high entropy.
Suppose
the sequence contains t distinct vectors and let pi denote the occurrence
probability of any vector v in the sequence obviously
The
entropy of the sequence is given
where log x denotes the base logarithm of x_ The entropy achieves its maximum value of log t when pi log pi For an n_bit vector(t,n)his makes the computation of the exact entropy very expensive. Assuming that the individual bits in the vector are independent_ then we can write
Where qi
denotes the signal probability of bit i in the vector sequence. Note that this
equation is only an upper bound on the exact entropy, since the bits may be
dependent. This upper bound expression is however the one that is used for
power estimation purposes.
The power
dissipation in the circuit can be approximated as
Power=0.5V2CtotEavg
Where Ctot
is the total capacitance of the logic module including gate and interconnect
capacitances and Eavg is the average activity of each line in the circuit which
is in turn approximated by one half of its average entropy havg.
The
average line entropy is computed by abstracting information obtained from a
gate level implementation. In it is assumed that the word level entropy per
logic level reduces quadratic ally from circuit inputs to circuit outputs.
Whereas
in it is assumed that the bit level entropy from one logic level to next
decreases in an exponential manner. Based on these assumptions two different
computational models are obtained
In M a r cu lescu et al derive a closed form expression for the average line entropy for the case of a linear gate distribution(i.e.,)when the number of nodes scales linearly between the number of circuit inputs n and circuit outputs m. The expression for havg is given by
where hin
and hout denote the average bit level entropies of circuit inputs
and outputs respectively hin is extracted from the given input sequence whereas
hout is calculated from a quick functional simulation of the circuit under the
given input sequence or by empirical entropy propagation techniques for pre
characterized library modules.
In Nemani
and Najm propose the following= expression for havg
where Hin
and Hout denote the average sectional word level entropies of
circuit inputs and outputs respectively.
The
sectional entropy measures Hin and Hout may be obtained by monitoring the input
and output signal values during a high level simulation of the circuit.
In
practice however they are approximated as the summation of individual bit level
entropies hin and hout. If the circuit structure is given the total module
capacitance is calculated by traversing the circuit netlist and summing up the
gate loadings.
Wire
capacitances are estimated using statistical wire load models. Otherwise Ctot
is estimated by quick mapping for example mapping to input universal gates or
by information theoretic models that relate the gate complexity of a design to
the divergence of its input and output entropies.
One such
model proposed by Cheng
and Agrawal in for example estimates
This
estimate tends to be too pessimistic when n is large hence in Ferrandi et al
present a new total capacitance estimate based on the number N of nodes i.e.,to
multiplexors in the Ordered Binary Decision Diagrams OBDD representation of the
logic circuit as follows
The
coefficients of the model are obtained empirically by doing linear regression
analysis on the total capacitance values for a large number of synthesized
circuits.
Entropic
models for the controller circuitry are proposed by Tyagi in where three
entropic lower bounds on the average Hamming distance bit changes with state
set S and with T states are provided.
Related Topics
Privacy Policy, Terms and Conditions, DMCA Policy and Compliant
Copyright © 2018-2024 BrainKart.com; All Rights Reserved. Developed by Therithal info, Chennai.