Tip:
Highlight text to annotate it
X
Hello and welcome to today’s lecture on supply voltage scaling. This is the fourth
and the last lecture on this topic. In the last lectures, we have discussed Static Voltage
Scaling and Multi-level Voltage Scaling. As we have seen in both the cases, the voltages
are assigned at designed time; in case of Static Voltage Scaling the supply voltage
may be scaled to some lower level at design time; and then the entire circuit or module
is assigned a lower level voltage and that remains fixed duration during it is normal
mode of operation. Similarly, in Multi-level Voltage Scaling,
multiple voltage levels are assigned to different modules or clusters, and those voltages remain
fixed during normal mode of operation. They do not change, when they are in use. Now,
we shall discuss another technique in which the supply voltage is dynamically changed
at run time; and today we shall discuss Dynamic Voltage and Frequency Scaling, and special
case of that is known as Adaptive Voltage Scaling, both of these two techniques, we
shall consider in this lecture. Now, this Dynamic Voltage and Frequency Scaling is based
on certain observation.
And these are the important observations; in fact the genesis of this Dynamic Voltage
and Frequency Scaling will come from these observations. Number 1 is the workload of
a CPU is a time varying function, which heavily depends on the application. For example, if
you consider the workload of a CPU, we will find that with time it varies that means,
it may somewhat like this; that means it is a time varying function and not only that,
it will also depend on the class of the system.
For example, this variation of the workload will be different for a server, I mean will
not be same as that of a desktop and it will not be same as that of a laptop. That means
the workload is varying with time, not only it is varying with time it is depending on
the type of application, server application or desktop application like that. And whenever
this is a varying function of time that means this is the workload; workload is varying
over time, whenever this is happening then as you see although the workload is varying
over time, you are maintaining a Fixed Voltage, and also you are maintaining a Fixed Frequency
in a during normal uses.
So, whenever you do so, there is a fixed power dissipation that means, the power dissipation
as you know is dependent on the supply voltage, it is dependent on the frequency of operation
and of course, it also depends on the switching activity and other things. That means the
power dissipation may remain fixed; so this is the power dissipation p that remains fixed
irrespective of the fact that the workload is changing. You may have noticed that in
your desktop or laptop if you look at this CPU utilization, you will find that sometimes
it is 3 percent, sometimes it is 20 percent, sometimes it is 50 percent depending on the
application you are running. So, when you are doing video processing say
MPEG, JPEG and all these things, then the CPU utilization is more; on the other hand
when you are doing word processing, sending e-mail, then you will find that CPU utilization
is very poor. Now, whenever this is the situation, can we really exploit this particular phenomenon
to reduce power dissipation? That is the basic idea behind this Dynamic Voltage and Frequency
Scaling. Now let us look at other observations. The energy drawn from the power supply is
the integration of the power over time. This is particularly important in the contest of
battery operated system as I have already told.
If power dissipation is say P or P1 and if the power dissipation occurs over a period
of time t1 than energy that is being drawn is equal to P1 into t1 that is the area in
covered by this rectangle, that means this is the energy drawn for performing this computation
say computation takes time t1 and during which the power dissipation is taking place is P1.
So, for a battery operates system if we can use power dissipation over time then we can
save energy that I shall discuss how it can be done.
Another important observation is that CMOS circuits can operate over a voltage range
with reasonable reliability that means, if you are already familiar with CMOS circuits
we have seen that in a particular CMOS circuit, you apply supply voltage Vdd let us consider
the simplest example the inverter if we if this is the circuit of an inverter, you see
this Vdd if we change even in the circuit will operate over a certain range not that
it will work from 0 volt or say it will work at a very high voltage that may not be true,
but there is a supply voltage range over which monotonically the voltage range, it will work
lively and not only that there is a region of operation for frequency increases monotonically
over voltage. That means, if we consider say voltage v it can be shown that over a range
voltage over a range of say may be 5 volt or may be this is 2 volt this circuit can
operate and in this particular case as we can see whenever you are reducing the supply
voltage the voltage is reduced, then you will find that the power dissipation will be reduced
and now over this range it can work and the whatever the frequency.
Now, as we know there is a relationship between delay of a circuit and the voltage and that
relationship is represented by this 1 minus Vt by Vdd square. So, that means delay is
proportional to 1 by Vdd into 1 minus Vt by Vdd, so delay increases linearly as it is
shown here that means if we plot voltage in this direction and instead of writing voltage
here say delay will reduce as you increase the voltage.
So, there is a linear relationship like this, not linear means it is kind of quadratic but
it is monotonic that means monotonically delay increases as supply voltage increase and the
relationship is given by this expression. So, it is a monotonic means as supply voltage
is reduced delay increases monotonically it increases it does not decrease after some
points so there is a range over which this happens. So, in this particular range how
this is or why this is happening, because as you know delay is changing so, the critical
path delay decides maximum frequency of
operation of this circuit that means, what does it means that you can really as you reduce
the supply voltage. Then it will operate at a lower frequency
so this based on this idea we can develop circuit such that the power dissipation can
be reduced by controlling the voltage and frequency dynamically as we shall see how
it can be done. So, based on these observations the technique of Dynamic Voltage and Frequency
Scaling has been developed.
First we shall consider a simpler situation where the frequency scaled down as the workload
reduces. So, in this particular case, as you can see, you were reducing the frequency as
the workload is reduced. So, there is a linear relationship between the energy and the workload;
so we can show you with the help of another plot. Let us consider a situation say this
is the case where workload is 100 percent and power dissipation is P1 that is corresponds
to workload is equal to 100 percent. So, what is the… If time is t1, the energy
drawn for this computation will be equal to P1 if into t1 with 100 percent workload. Now,
let us consider the situation, where the workload becomes 50 percent, now if you do not change
the frequency of operation, power dissipation will remain same however the computation will
be complete within 50 percent of the time that means in time t2, which is 50 percent
of t1 that means, t2 is equal to t1 by 2, then power dissipation will be power dissipation
is remaining same, but time is decreasing. So, in this particular case as we can see
energy drawn is reduced. It is becoming half, because P1 into t2 that
that is reduced however power dissipation remains same during this period t2 now you
can you can in this particular case neither the voltage is scaled nor the frequency is
scaled. So, during this period power dissipation is taking place and you switch off during
the remaining part or you may not if you do not switch it off then power dissipation will
continue to take place in the remaining part, but wiser technique is to use Frequency Scaling
as it is shown in this particular case even if you do not do voltage scaling if you simply
reduce the frequency of operation and make it half then what will happen?
The computation time will become half, but it will continue to take place for the entire
time t1, that means in this case power dissipation will reduce and this is power and this direction
power and time. So, in this case you can see P2 into t1 is same as that of the previous
case, that means energy drawn is reduced is same as this one the previous case that means
P1 into t2 is equal to P2 into t1 energy drawn is remaining same however the power dissipation
is reduced this has definitely has impact on the packaging and cooling of the system.
That means, if the power dissipation is less the lesser heat dissipation takes place and
packaging and cooling will be cooling cost can be reduced. So without voltage scaling
simply by Frequency Scaling, you can achieve lower power dissipation energy remaining same
whenever the workload is changing. Now, let us consider the case, where you can perform
voltage and Frequency Scaling that means in this case, as you can see since the frequency
f2 is equal to f1 by 2 in this particular case that is the reason why it is taking longer
time. So, you have scaled down the frequency to
f1 by 2 and that is why it has taken double time of that 1 and power dissipation has reduced,
because P is proportional to Vdd square into f. So, here it is f2 and there it was f1.
So, power dissipation has reduced because voltage was fixed, but now what you can do,
as you know if the frequency is less then this circuit can operate at a lower voltage.
As we have discussed few minutes back that the observation is as the supply voltage is
reduced, this circuit will operate at a lower frequency or can operate at a lower frequency.
So, in this case there is scope for scaling down the voltage. So, what you have done the
time is remaining same, time is not changing, but instead of P1.
It is now P3 and which is definitely less than P2 or less than P1. So, you see that
there is a significant reduction in power dissipation not only power dissipation, energy
is drawn from the battery also much less. So, in the previous two cases energy was same,
because only Frequency Scaling was known power dissipation was less, but energy drawn from
the battery was remaining same, but whenever you do voltage and Frequency Scaling Voltage
and Frequency Scaling as you do as you have done here then not only the power dissipation
reduces, but energy drawn from the power sources also significantly reduce that means this
is very important in the contest of battery operating system.
Here energy is significantly reduced and as you can see this is voltage and Frequency
Scaling. This is only Frequency Scaling and in this particular cases neither no voltage
nor Frequency Scaling and here also no voltage and Frequency Scaling, but here the workload
is 100 percent, here the workload is 50 percent, here the workload is 50 percent and here also
the workload is 50 percent. So, this is based on this. Now we shall develop
Now let us consider the realization of a circuit and as I have mentioned energy drawn can be
reduced by dynamically adjusting both voltage and frequency that meets the workload condition.
This is the basic idea of Dynamic Voltage and Frequency Scaling that means during run
time, you will find out at what frequency this circuit should operate to meet the workload
requirement. Then for that part to operate at that frequency, what supply voltage is
required that you can find out. So, you can assign you can the circuit the CPU can work
at a lesser voltage and later lesser frequency there by reducing the power dissipation. This
curve shows how the energy reduction takes place. This is the case where there is no
voltage scaling, but here this is the ideal voltage and Frequency Scaling.
So, I have used the term ideal, later on we shall see it may deviate from this ideal,
because of some other reasons, because of DC to DC converter and other thing that we
can consider later. But you see there is a significant reduction in energy so there is
a cubic reduction in power dissipation and quadratic reduction in energy that will take
place by using Dynamic Voltage and Frequency Scaling. How can it be realized?
This is the model of the Dynamic Voltage and Frequency Scaling system. So, here it shows
you, have got a processor CPU microprocessor mu r, which can operate over a voltage range
that is why the name is variable voltage processor. So, this particular processor can operate
over a voltage range. And you can see the task input is coming from a number of sources
and different sources, this is the task q coming from different sources, this is the
rate of arrival average rate of arrival of different I mean, the task from different
sources lambda 1 is the average rate of arrival from source 1 lambda 2 is the average rate
of arrival from source 2, and in this way lambda n is the average rate of arrival of
task from source n. Out of which you know, although the tasks are coming to from different
sources, there is a scheduler as you know, which is part of the operating system, there
is a task scheduler; the task scheduler will issue tasks, which can be executed by the
CPU, so lambda is the rate at which tasks are issued to this variable voltage processor.
Now the variable voltage processor has got a workload monitor. What is the job of this
workload monitor? It receive, it can find out what is the present workload based on
the task which have been executed by these processor and this workload monitor what it
will do not only it will keep on receiving the information about the workload of the
present workload it will estimate what is the workload required in the next time slot.
A kind of a prediction, workload prediction it will do and based on that prediction it
will send some signal to a DC to DC converter that means, if the workload requirement is
less that it this DC to DC converter will generate a lesser voltage although it will
receive a fixed voltage it will generate a variable voltage V r and this variable voltage
is dependent on this workload and of course, this is related to the frequency. So, there
is a frequency generated variable frequency generator the workload monitor will also send
information about the predicted workload or to the frequency generator and then it will
generate a frequency and that frequency will be applied to the variable voltage processor.
So, this voltage is applied, this frequency is applied and to sustain that frequency a
particular voltage that is required that is being applied by the generated by this DC
to DC converter. So, this is the basic model, which is being
shown and what is written here workload for the next observation interval can be predicted
by the OS kernel based on the workload statistics of the previous N intervals. That is the job
of these workload monitor. So, it can find out, it can predict the workload in the next
time interval. This is the basic model.
Based on this basic model we can identify what is the hardware and software requirement
for implementing Dynamic Voltage and Frequency Scaling system. First of all as we have seen
we require a variable voltage processor CPU, because you know there are many processors,
which will operate at a fixed voltage which are not designed to operate over a over a
range of voltage, but there are processors which can operate over a voltage range so
that will be the first requirement. Second requirement is variable voltage generator
so V because you know, this to operate at different frequency.
You will require different voltage that has to be generated by hardware which is known
as variable voltage generator you will require a variable frequency generator as you as you
have seen in the basic model f r which will keep on generating variable frequencies and
last, but not the least the workload predictor which is the part of the workload monitor
and may be part of the operating system kernel. So, a part of the operating system kernel
can perform the task of workload prediction.
Now, coming to the first requirement, that is variable voltage processor. There are processors
available, one is your strong ARM 11 00. This allows voltage scaling as you can see it can
operate over a voltage range 1.1 to 1.65 and accordingly it can operate at different frequency
that means when the supply voltage is 1.65 volt the operating frequency is 700 megahertz
and this is the relative power 100. So, this is normalized with respect to this
1 the other power the other power dissipation and then when the frequency is 600 megahertz
then the voltage optimum voltage or minimum voltage that is required is 1.60 and relative
power dissipation is 80.59 of the previous one. And if the frequency is 500 megahertz
it can operate at 1.50 volt and the relative power dissipation is 59 or 59 percent you
can say of the first case and in this way. It can go up to 200 mega-hertz and the voltage
required is 1.10 volt and you can see the relative power is 12.87 percent.
So, this you can see one-eighth of the first case that means if the frequency is varied
from 700 megahertz to 200 megahertz and accordingly the supply voltage is varied from 1.65 to
1.0 than power dissipation can vary from 100 to 12.7. So, that can be significant reduction
in power dissipation and obviously the energy drawn from the battery source will be there,
will be reduction drawn from the battery. Another processor is there, that is Transmeta’s
Crusoe processor there is a manufacturer, transmitter; they have developed a processor
which is very innovative Crusoe processor that allows the following adjustments. It
allows frequency change in steps of 33 megahertz, it also allows voltage change in the steps
of 25 mili-volts and there can be 200 frequency voltage changes per 200 frequency voltage
changes per second. Whenever you go from one particular frequency
to another frequency model operation you cannot change very quickly, the reason for that is
DC to DC converter and the frequency generator will take some time to settle down and that
is the reason why the rate at which you can make this change in the frequency and voltage
cannot be very fast, but you can see that 200 voltage changes per second is quite reasonable
and can be suitable for many applications. So, in this particular case as you can see
this Crusoe processor can operate over the range of 2.2 to roughly 0.7, 0.7 to 2.2 volts
.You can see this portion the shaded portion is operational range and if the voltage is
above 2.2 it is the destructive region that means the power dissipation will be very high
and that may lead to destruction of the device and similarly, as you reduce the voltage you
can see the frequency of operation is also reducing. So, whenever it has to that minimum
maximum frequency decided by this particular. So, as the voltage is reduced from 2.2 the
frequency operation will reduce from 221 megahertz to close to 70 megahertz. So, you can see
this can also operate over a large voltage range and also over a large frequency range
and you can have large number of voltage frequency appears. Although I mean it is shown continuous,
later on we shall see, we will normally use a limited number of voltage frequency pairs.
It will not change continuously. Some discrete voltage frequency pairs are used for operation
of the circuit.
So, this is about the variable voltage processor so in fact this ARM processor is now available
in the form of a I p core and you can use it in many embedded applications and that
allows Dynamic Voltage and Frequency Scaling. Coming to variable voltage generator you can
see here this variable voltage generation can be done by DC to DC converter. DC to DC
converter is commonly used in on all modern digital equipments. You can see it receives
a fix voltage VF is received and there is a pulse-width modulator which controls a switch
S. This switch although is a shown as a mechanical switch it can be it can be an electronic switch.
Normally, it is realized by device known as IGBT insulated gate by polar transistors.
So, IGBT is commonly used as a switch in this DC to DC converter. And so what will happen?
How does it control the voltage? Here what it does?
Suppose this is the fixed voltage VF and the duty cycle of this at which the switching
is done by this particular device is controlled by this Pulse-width Modulator. Let us assume
the duty cycle is 50 percent. So, in this case that means, 50 percent of the time this
switch will be closed and 50 percent of the time the switch will be off. So, in this way
the switch will operate. And this duty cycle is dependent on the control input that we
shall see; so there is a control input here you can say Reference Voltage; depending on
this reference voltage, this pulse-width modulator adjust the duty cycle.
So, this output of this switch is like a signal like this. So, a kind of what train of pulses;
now this is applied with Low pass Filter. So, as you apply to a low pass filter, then
what will happen? The output of the low pass filter will remove the high frequency component,
and you can get only the DC component, so you can see you will get a voltage, which
is a DC voltage, and which will be equal to VF by 2; since the duty cycle is 50 percent,
it will generate a voltage VF by 2; that is the basic principle of this DC to DC converter.
And here you can see this Low-pass Filter output is applied to a comparator and this
is compared with a reference voltage depending on this comparison it will generate a signal
to the Pulse-width Modulator and accordingly it will increase or decrease the duty cycles
depending on the output is more or less with respect to the reference voltage this is the
basic principle of this variable voltage generator. You will get a voltage VO which will vary
from may be from 0 may not be exactly 0, but very close to 0 to VF and that is the Reference
Voltage VF and that can be applied to the load or the circuit under operation, and this
kind of DC to DC lot of research work is going on efficient DC to DC converter. Nowadays
technique known as buck converter is used which provides very high efficiency may be
up to 90 percent efficiency is achievable from Buck converters. Buck converters are
nothing, but special type of DC to DC to converter and people are also trying to reduce the size
the DC to DC converter so that it can be put on chip.
Obviously all the components cannot be put on chip, but the electronic part can be put
on chip, but a capacitor and some inductor and some other components has to be connected
externally. I am not going into the design of a DC to DC converter and though electric
engineering students will find in their department research work is going on for the development
of efficient DC to DC converter. So, but this is the basic principle I have explained. And
this type of DC to DC converter can be used to generate a variable voltage variable supply
voltage which is required which is one of the primary requirements for implementing
Dynamic Voltage and Frequency Scaling system.
Now, coming to one very important aspect of DC to DC converter, as I have mentioned, it
is possible to realize DC to DC converters having high efficiently say 90 percent and
above. Unfortunately the efficiency drops off as load decreases; that means, when the
load is high 80 percent, 90 percent or 100 percent, then the efficiency of the DC converter
is high, because this DC to DC converter is a electronic circuit, which has some power
dissipation; and whenever the duty cycle is small, then the efficiency of the DC to DC
converter reduces as you can see the relative efficiency degrades as the workload is reduced;
so the efficiency drops off as load decreases. So, you have to this I am raising this point,
because whenever we will be considering the efficiency of Dynamic Voltage and Frequency
Scaling system, you have to not only take into consideration the efficiently of the
processor power dissipation that is taking place in the processor, but also the power
dissipation that is taking place in the DC to DC converter, because DC to DC converter
is now part of your system; and the power dissipation that is taking place in DC to
DC converter has to be taken in to consideration. So, this point you have to keep in mind, when
you are using DC to DC converter with a small workload.
Coming to another important component, that is your Variable Frequency Generator. Variable
Frequency Generator can be generated with the help of Phase Lock Loop. This is the traditional
approach for generating variable frequency as you can see here there is a phase lock
loop and the heart of this device is the high performance phase lock loop core consisting
of a Phase Frequency Detector PFD and Programmable on-chip filter and a VCO, this is part of
this part of this Phase Lock Loop. So, inside this Phase Lock Loop, we have three
important components, one is phase frequency detector, one is Programmable on-chip filter,
another is VCO that is voltage control oscillator. It is here; so it generates a high frequency
then that frequency is divided with the help of a frequency divider to generate the desired
frequency. So you see Frequency Scaling is done in two steps; number one step is you
can change the frequency of the Phase Lock Loop, and also you can change the division
ratio of this frequency divider. The divider generates ultimately you are getting
the variable frequency from the output of the frequency divider, but together the Phase
Lock Loop and the divider together generates the independent frequencies related to the
PLL operating frequency that means, PLL can operate at different frequency say f1 f2 and
so on. Then f1 can be divided by a factor d1 and that ultimately is the f for the that
is being generated that means, the frequency generation is done in two steps; later on
again we shall discuss about it. This is how variable frequency can be generated.
Coming to the workload prediction, as I said this is also very important thing, an adaptive
approach for workload prediction by filtering the trace history. So, this is a technique,
which has been developed proposed by Amit Sinha and Anantha P Chandrakasan of MIT. This
is the approach they have provided; so essentially, these workload prediction is somewhat similar
to you know the way the filters works, digital filter works. So, workload for the next observation
interval can be predicted based on the workload statistics of the previous N intervals that
means, it is a n Nth order Markov process.
So, whenever you say it is a Nth order Markov process that means it present output present
the presents status is dependent on the past N status so if you know the workload in the
previous N time intervals you can predict the workload of the present interval that
is the basic idea of this workload prediction and h and k represents an N-tap adaptable
FIR filter. So, it can be generated with the help of FIR filter and there are three basic
approaches. I shall discuss, one is moving average workload MAW, where this h and k is
equal to 1 by M. You are finding out the workload say W p n plus 1, which is equal to 1 by N
of say W 1 plus W 2 in this way W n that means, what you are doing? You are taking the average
of these previous n workloads. This is how you are doing the prediction that
means, the workload in the n plus 1th instant is the average of the workload of the previous
N time intervals, in fact that is what you are generating in this particular expression.
That means in this case h n is workload prediction h n function is equal to 1 by N, and this
is the summation of the w k n minus k that is n 1 W 1, W 2 up to W n. So, this is the
first case moving average workload. Why it is called moving average? Because suppose
this was the workload during first time interval W 1, and next time interval it was W 2, and
next time interval it was W 3. So, in this way workload say W n.
The workload during this period can be predicted from this then workload for this next period
will be predicted from this window that means, first you are considering this window next
you are considering this window that means, this window is moving from as the time is
progressing and that is how you are keeping on taking the average of the previous n workload
histories and that is you are doing the prediction. Now, this particular workload prediction works
very well, when the workload changes very slowly.
So, when the workload changes very slowly, since it is changing very slowly increasing
or decreasing whatever it may be if you take the average of the previous n values, it will
become very close to the next workload, but whenever it changing little fast, then of
course, it will not work very well. In such a situation, what you have to do? You have
to give more importance to the you know that not the past workloads increasingly more importance
to the workloads, which are nearer or you know closer to the present workload.
So, that is what is being done in that Exponential Weighted Averaging workload so here what you
are doing here the. Here you are multiplying that h n k, this function is equal to 1 by
a k. So, this you are multiplying this workload W1 with a factor 1 by a k and this value is
large, this 1by a k is a k value of k is large whenever you know W is equal to 1 and for
W is equal to n it is very small that means, you are dividing W1 by a large number and
Wn by a small number then taking the sum of that.
And that is how you are giving more importance to the most recent history and the most past
history you are giving lesser importance. So, this is the basic idea behind this Exponential
Weighted Averaging, where the workload function is equal to 1 by a to the power k. Another
important workload prediction function is the Least Mean Square LMS here h n plus 1
k is equal to h n k plus mu into W e here you were computing the error W e n is the
error, W n k into W k n minus where W e n is equal to W n minus W p n so every time
what you are doing the predicted workload and actual workload in a particular time slot
that difference you are computing and that difference you are using for the computation
of the next workload. So, here the function is you know h this h
n this function is h n plus 1 k is equal to h n k plus mu W e n where mu W e n is equal
to W n minus W p n and mu is the step size of course, you can choose the step size this
value constant and that constant you can adjust to, so that this work this error is less.
So, this is how you can these are the three different work functions which you can use
and here essentially you are using a N-tap adaptable FIR filter those N-taps are essentially
storing the workload function of the previous n time slots. So, this is how workload prediction
can be done and you can refer to if you want to learn about it more you can you can get
the get it from the paper by Anantha P Chandrakasan.
And this is the prediction performance of different filters is shown here. As you can
see the RMS error root means square error is minimum for this LMS approach, this l m
s approach. I have showed means least means square approach, where the error is minimum,
error is maximum for this moving average window and this exponential weighted averaging gives
you lesser than moving average window, but this LMS definitely gives you much less error
compare to the other two schemes.
So, coming to another very important design parameter that is the discrete processing
rate we have seen that we will be using a number of voltage and frequency error pairs.
Question is, how many 2, 3, 4 or 100 or 200? Here the number can be large. So, depending
on how many voltages you are using, there is a kind of quantization is taking place.
So, you may find out that your workload requirement is W 1, but because of quantization error
because we have got only two workloads, I mean two different voltages in that case you
have to use either V1 or V2 so error will be more. So, in this way some simulation has
been carried out to check. How many voltage frequency pairs will give
you reasonably good performance 2 or 3 or 4 or 5. So, here is the particular impact
of this Discrete Processing Rate is given here you can have number of levels can vary
from 2 to 10. So, in this case that means the voltage frequency pair that has that can
be used is varying from 2 to 10. So, Whenever too few frequency points, I mean say you are
using 2 or 3 in such a case what can happen too few frequency points may result in ramping
between two levels for workload profiles and too many levels may result in hunting of the
power supply for a new supply voltage most of the time. So, in these so, there are a
trade up, trade up between two small and two large you have to choose an optimal number
such that these kind of problem does not arise. So, it has been found that for L is equal
to 10, that means number of voltage level is voltage level 10 or in other words number
of voltage frequency appear is equal to 10 you get roughly, you know I mean the accuracy
is very good and almost 90 percent. For example, for L is equal to 10 which is
provided by all practical processors the ESRdegradation due to quantization, ratio quantization noise
is less than 10 percent that means if you increase beyond 10 say if you use 15, 20,
30 obviously the error will be less, but even with 10 voltage frequency pair the accuracy
is, I mean the efficiency that you are getting, the reduction that you are getting is close
to 90 percent than the ideal. So, this is the Discrete Processing Rates that you have
discussed.
Another important parameter is latency overhead. As I have already told there is a latency
overhead involved in processing rate update. This is due to finite feedback bandwidth associated
with the Phase Lock Loop and the DC to DC converter, because this Phase Lock Loop and
DC to DC converter these are analogue systems. So, they will take time to settle down so
even if you apply some signal for a particular voltage for a particular frequency the new
voltage and new frequency will be available after certain time, that is why you will have
some latency overhead and switching to higher processing rate you have to you have to be
little careful and you have to follow some steps. The voltage should be increased first
followed by the increase in the frequency and the following steps will be followed.
Why this is necessary? Suppose, if the frequency increases before
the voltage increases what will happen chip will not operate, chip will not function properly
as we know we are trying to give I mean you know, there is a relationship between voltage
and frequency. So, as the voltage is increased, if the frequency is increased voltage has
to be increased first. If the frequency is increased first then voltage the chip will
fail to operate at that frequency that is the reason why you have to follow this step.
Set the new voltage, allow the new voltage to settle down, then set the new frequency.
Then secondly switching to lower processing rate in this particular case, set the new
frequency, then set the new voltage the CPU continues to operate continues operating at
the new frequency while voltage settles at the new voltage level. So, in this particular
case restriction is less, because when the voltage is high the chip will operate at high
frequency as well as low frequency, but whenever you are switching from you know you are changing
from low to high frequency, there you have to be careful.
So, after discussing Dynamic Voltage and Frequency Scaling now it is time to discuss Adaptive
Voltage Scaling. What is the difference between Dynamic Voltage and Frequency Scaling and
Adaptive Voltage Scaling?
Now, the voltage scaling techniques discussed so far are open-loop in nature. Voltage-frequency
pairs are determined at design time keeping sufficient margin for guaranteed operation
across the entire range of best and worst process, voltage and temperature conditions
that means you know those voltage frequency pairs are decided at designed time, although
they keep on their applied at run time, but they are decided at design time.
And here the design has to be very conservative, so that the chip operates under best and worst
condition of process voltage and temperature variations. So, under best and worst conditions
of process variation, voltage variation and temperature variation chip has to work and
that is the reason why those voltage frequency pairs are chosen in a very conservative way
and as a consequence, the gain or reduction in power dissipation that you really achieve
is not very I mean, less than what can be achieved if the design is not that conservation,
but how can it be done? It can be done if you use a close loop technique, instead of
open loop technique. So, in this case it is a open loop technique
as the workload is changing, you are blindly applying a new voltage and new frequency without
measuring whether the voltage and frequency requirements are appropriate or not? And that
is the reason why a better alternative that can overcome this limitation is the Adaptive
Voltage Scaling, where a close-loop feedback system is implemented between the voltage
scaling power supply and the delay sensing performance monitor at execution time that
means, as the execution is taking place the CPU performance is monitored, it is delay
is monitored, its power dissipation is monitored and by doing that you can adjust the voltage
and frequency by using a close loop technique and for that purpose what you have to do?
You must have an on-chip monitor. That will not only check the actual voltage
developed, but also detects whether the silicon is slowing typical or fast and the effect
of temperature on the surrounding silicon. So, that means it is taking care of the PVT
variation. PVT variations that take place since you are doing at run time the time of
execution whatever variation has taking place that is monitored; and based on that, you
are doing the design.
So, let us look at the system, what it does. So, here I have taken an example of Adaptive
Voltage and Frequency Scaling from a paper, Dynamic Voltage and Frequency Scaling by Masakatsu
published in IEEE transactions on you know Journal of Solid Set Circuits, January 2005.
So, here you can see apart from these functional units the ARM processor, the RAM, the Peripheral
Blocks, Blocks control circuits. This is the main block, which is shown as a small part
of the circuit. Here you have got a dynamic voltage control, which does delay synthesize
that means, the delay of the circuit can be the delay information is generated based on
real life measurement by this. And similarly, the voltage and dynamic voltage and voltage
dynamic frequency control circuit finds out at what frequency it should operate. Based
on the actual activity measurement that is taking place in the processor that means,
whatever the actual performance required is based on that frequency generated and accordingly
clock generator generates clocks and so on.
As it is explains here the DVC emulates the critical path characteristics of the system
by using a delay synthesizer and controls the dynamic supply voltage in the range of
0.9 to 1.6 volt at 5millivolt step in real time. Similarly, the dynamic frequency control
adjusts the clock frequency at the required minimum value by monitoring the system activity
in the range of 20 megahertz to 120 megahertz. So, you see this DFC and DVC Dynamic Voltage
Control and Dynamic Frequency Control. These are on-chips hardware, which does the actual
measurement of delay and activity level. And based on that, it sends information to
the PLL and clock generator and then the clock generator generates the system clock and similarly,
here the based on the delay information, information goes to DC to DC converter information block
and then voltage is generated by the DC to DC converter and applied to the processor.
So, you see the voltage and frequency are predicted according to the performance monitoring
of the system the Dynamic Voltage and Frequency Scaling system can track the required performance
with a high level of accuracy over the full range of temperature and process deviations.
So, it is a kind of close loop control system that has been adopted, so it is a Dynamic
Voltage and Frequency Scaling, where the close loop technique has been used instead of open
loop technique.
So, with this we have come to the end of todays lecture and here is the summary we have discussed
Static Voltage Scaling techniques, we have discussed Multilevel Voltage Scaling techniques.
And today, we have discussed Dynamic Voltages and Frequency technique and special case that
is Adaptive Voltage Scaling technique and in the next lecture we shall discuss about
the minimization of switch capacitance. Thank you.