Basics of software-defined radios & practical applications

Dr. Meenakshi Rawat

Department of Electronics & Communication Engineering
Indian Institute of Technology, Roorkee

Lecture – 06
Software-defined radio architecture Part IV

So, in the series of basics of Software Defined Radios and practical applications, we are
covering the 4th part of the Software Defined Radio architectures. Previously we had
seen heterodyne, homodyne architectures, and their combination, digital IF heterodyne
structures, which also contain the benefits of homodyne structures. Now we want to have
a look at the concept of SDR when we want to do the sampling directly at the RF
frequency. So, what are the requirements? What are the limitations? This is what we will
be covering in this lecture.

(Refer Slide Time: 00:53)

So, architectures to alleviate limitations of DAC and ADC, by using the quadrature
channels. We have already reduced the bandwidth requirement by half; it means I and Q
channels both of them are carrying half of the bandwidth, which required for the full
complex signal.

So, we are already reducing some of this limitation. Now what is the impact of the
bandwidth of the signal on DAC? Till now, most of the discussion we have done in terms
of the signal band, a sinusoidal signal is coming, what will be the upconversion,
downconversion? How will we choose the frequency? This is what we have covered.

Now, let us have a look at the bandwidth impact on DAC and ADC performance. First of
all, let's recall our Nyquist-Shannon theorem. So, Nyquist theorem we had seen
maximum frequency component, then our sampling frequency should be more than the
twice of incoming maximum signal frequency.

(Refer Slide Time: 01:54)

Now Nyquist-Shannon theorem says, for the band-limited signal, the sampling frequency
should be more than equal to two times of signal bandwidth. Now when we say signal
bandwidth, it is the band-limited signal. So, we are talking about -B/2 to B/2 if it is at the
baseband. So, this thing we have to remember when we are talking about the band-
limited signal.

Another thing is whenever we are sending any signal through DAC by sampling it, and
we are receiving in a receiver, and we are getting back our digital data from the analog
domain using ADC. Then whatever we have discussed before, it was in terms of single-
band signal. So, we were talking about a single band signal, and then we have said this is
a signal and let us do the quantization, etc.

So, it was the amplitude, now this signal, suppose it has a bandwidth, how will it impact
in the frequency domain. So, for this signal, which is the sinusoidal signal, in the
frequency domain, it appears as let us say it is fRF as a single tone, all right?. So, your
DAC, which is able to take this highest amplitude, this highest amplitude is defining the
power which it can take.

So, for a single band that power is concentrated in this frequency. Now suppose you have
two tone signal then we can say that this becomes the bandwidth of the signal, then this
both of these signals they contribute to that power and the ADC is reaching their highest
level of amplitude, because of these two tones. So, suppose this was the noise level, and
this was the P1, which was because of an only single tone, this was the dynamic range,
now when you have two signals the power is distributed between these two, right? So, it
will be reduced for sure because it is distributed between these two tones and this is your
new P2. So, P1 has higher SNR for the single carrier, even increase the number of carriers
the power is distributed. So, the highest point of power with respect to noise is coming
closer to the noise it is becoming lesser and lesser.

Now, this distance from the highest power to the noise power, it is called the dynamic
range of any ADC or DAC. So, as we can see here, if we have a single carrier, this is
being shown here by this arrow, as you keep increasing our bandwidth the amplitude and
the average power will go down because the whole power is contained within this signal.
So, the dynamic range goes down, so what is the problem with that? The sensitivity of
our signal will go down; what is the sensitivity? We have defined it before that the ADC
should be able to distinguish the actual signal with respect to noise or the interfering
signals. So, if this amplitude keeps coming closer to our noise level, then it is difficult to
distinguish which is our main signal. Maybe if our interference signal is higher than this,
suppose it has a lower bandwidth it will have more chance of saturating over ADC, then
it will be creating a problem in detecting the actual signals.

So now, bandwidth can be of two types like I said, if you have two-tone signals two
tones are still covering the bandwidth B, this is multi-band kind of operation there are
two bands there are two frequencies on which it is working, this was enough example of
only two-tone signals, but it is possible that, it is a multi-carrier signal here, let us say
four tons here and four tons here also now they have their own bandwidths B1, B2 where
it contains its own tones and together by using these two they are again making this
bandwidth B.
So, this B contains B1, B2, and the distance between these tones. But now, what is the
sensitivity requirement for this one? When it was a single tone, we have a very high
dynamic range. When we have this two-band system, the dynamic range will go down.
And if I want to fill this whole frequency range with the carrier you want to transmit at
all these frequencies, then the whole band being used, in this case, our dynamic range
will further go down, because it is saturating our ADC and DAC with its whole power.

When the signal contains all the frequencies within the band, then we call it broadband
transmission. And when we are using some of the frequencies selectively, that is called
multi-band transmission. So, let us keep this in mind, we will come back to it later. Now
let us go to the concept of sampling. How it affects our architecture there? (Refer Slide
Time: 08:01)

And how it is converted back into the frequency spectrum. X(k) is representing your
complex discrete frequency spectrum, where k is the index of discrete frequency bins, in
the right-hand side again, the time is sample time Ts, and Ts in your sampling frequency.
So, whenever we are having multiple of that sampling frequency, we are sampling at a
particular point where a one-time duration is being finished.
(Refer Slide Time: 08:41)

If we see in terms of continuous frequency, in the time domain, and we want to see what
will be the complex signal frequency domain, then it is given by the integration as
opposed to the summation normally we use there.

(Refer Slide Time: 09:02)

So, whenever a signal is being sent, and it is a limited signal, let us say these red lines
they are representing one signal duration. It is a limited signal, then corresponding to that
we will have a Sinc function because it is a sinusoidal signal. So, because of the
windowing function, we can see at fin, which is at the instance of 1/NTs, our sampling
points. So, at the exact point, we are able to sample our data if it is a multiple of Ts in the
time domain and 1/NTs in the frequency domain.

(Refer Slide Time: 09:46)

So, these are the particular points we are getting here. If our signals are periodic in
nature, but they have disrupted nature, there is a sudden jump as we can see in this case,
then we can see the shift because we use the interpolation filter in the frequency domain.
There is a shift here in the data, right? And we will be losing some of the peak points.
So, the peak points we have lost here, we are not getting the exact frequency location, so
we are losing some information there.
(Refer Slide Time: 10:10)

Now, keeping those concepts in mind, let us go to the Nyquist frequency criteria, we
know that our sampling frequency should be more than the twice of fin,max. So, in this
figure, you can see that this is our signal in the frequency domain, and the max will be
decided by this boundary here, right? So, when it is sampled at a particular sampling
frequency which is more than twice of fin,max, then how do we get our signal, our signal is
here which falls in the boundary of 0 to fs/2.

Now this 0 to fs/2 is called the first Nyquist zone. From fs/2 to fs, we have the second
Nyquist zone, where we see the image of this signal. So, the conjugate of that everything
will be negative here. So, you can see this is here. From fs to 3fs/2 is the 3rd Nyquist
zone, and from 3fs/2 to 2fs, we have 4th Nyquist zone. So, in each zone you can see your
signal that you can receive and you can receive the image of that signal. Now suppose
our signal fin,max is more than fs/ 2, then what happens as you can imagine that the signal
will start from 0 to fin,max.

So, the signal is here, but because we are not taking care that our fs more than twice of
the maximum frequency, then there some aliasing of the image and the original signal.
So, keeping this concept in mind, they are some of the benefits of oversampling and
undersampling architectures which I want to discuss.
(Refer Slide Time: 11:44)

What is the benefit of oversampling, for example, this is the first Nyquist zone as I have
shown earlier, it is the second one then third, and the 4th one and we want to retrieve our
signal from here, this is our signal, and we have filtered it.

Now, this filter location is very new to the image frequency, if fs become even a little
smaller, then this image will have some part inside that filter because filters are not ideal
filters, right? They do not have a sharp cutoff, and they can gradually go down. So, to
keep our signal intact to be able to filter it properly, oversampling is taken into account.
In the oversampling, because we increase the number of samples, we increase the
sampling rate by the Nyquist criteria our signal is here, but fs /2 is lying further from the
signal, and the image signal is much further.

So, we can have some loose constraints on the filter bandwidth. We can use a filter
bandwidth with a large roll-off factor, and it can go slowly away. In the practical
conditions, most of the filters are like that, and they do not have a sharp cutoff, they will
be gradually decreasing in magnitude.

So, more than that, because we have increased our Nyquist zone. So, the thermal noise is
distributed in the Nyquist zone. So, initially, it was distributed in this area the same
power now it distributed into this area. So, the noise goes down. In fact, the effective
SNR in the case of oversampling is given by the original SNR + 10 log10(D).
Where D is the decimation factor or the oversampling factor, so you will be hearing the
decimation and oversampling, these two terms are used in the place of each other, and
both mostly mean the same thing. So, let us take one example, suppose you have a 12-bit
ADC and let us do the 16 times oversampling with this. So, what will be the actual SNR
of this system? So, 12-bit ADC, what will be the dynamic range of that one? We have
SNR for the single-carrier signal, and the original SNR formula is 6.02n+ 1.76.

(Refer Slide Time: 14:05)

So, it will be around 74dBc. So, the signal to noise ratio is 74dBc for the 12 bit. Now we
have done the oversampling by 16. So, what will happen? The new formula of the
oversampling SNR will come into effect, which is the original SNR + 10 log
10(decimation factor, which is 16).

So, the summation becomes 74+12.04. So, SNR has increased almost by 10 dB for the
single carrier, of course, for the multi-carrier or the broadband signal it will be lesser, but
we can see the impact on the calculation here. So, oversampling has certain benefits and
which should be taken their advantage whenever we are designing any system. (Refer
Slide Time: 15:34)
Here we see the use of oversampling in practical implementation in the 16 bit DAC,
AD5688 from the texas instrument. So, as you can see, first of all, our data which is
coming for the I and Q channel and 16-bit data being applied here, we are having from 2
to 8x interpolation options are there. So, we can choose this interpolation option
maximum we can have 8, and minimum we can have two, or we can simply avoid this
portion. But increasing interpolation gives you good signal to noise ratio. So, it is
recommended that we can use it, then 32-bit NCO is being used here.

So, as I told before, now in this scenario by using appropriate sampling frequency,
coding there, we are able to upconvert in the digital domain this oversample data which
is keeping its image away quite nicely, now phase and gain correction are happening
here because we have two different branches.

So, we put this linear correction by our hand, and we will be dealing with this later in
detail. Now because we have our system in the first form, and we have the Sinc kind of
architecture at the output whenever you have a rectangular window there, right? So, to
remove the effect of that rectangular window, we multiply with the inverse of that. So,
because of the rectangular window will give you Sinc kind of function sin(x)/x. So, we
want to multiply with the x/sin(x) to remove the effect of that one. Whatever we are
getting at the output, this is what we applied to our digital to analog converter to be
converted into the analog domain.

(Refer Slide Time: 17:30)

So, this is a practical example, and we can see that we are using the NCO as well as we
are using the oversampling in the actual setup. Now, this was the example for the
oversampling, how we can use the subsampling or under-sampling. So, let's take one
example of a signal, which is a 20MHz bandwidth signal, which is available from
60MHz or 80MHz, and our IF frequency is 70 MHz, which is the RF signal.

(Refer Slide Time: 17:52)

So, what we are saying that in the frequency domain from 60 to 80MHz, our RF signal
lies. So, bandwidth becomes 20MHz. We are told that our RF frequency is 70MHz. So,
70 MHz is the IF frequency. So, how to choose our sampling frequency so that we can
have the signal properly? So, the highest frequency, highest f is 80MHz in this signal,
suppose I take it to the baseband, this will be the highest frequency there.

So, we have to have at least twice of that, let us say. So, our sampling frequency should
be getting equal to twice of fmax. So, it should be more than 160MHz, right?. So,
suppose I take the 160MHz sampling frequency. So, fs is equal to 160MHz. So, fs /2 will
be 80MHz.

So, in this case, if I do the down-conversion, my signal will be on the boundary, right? It
will be 80MHz, and it will be 60MHz. So now, it is at the boundary, and it is image will
be starting right over here, it will be 100 MHz. So, we want to avoid that we want to
have it in between. So, let us take our sampling frequency fs is 280 MHz. It means fs /2
is 140 MHz, and this is our first Nyquist zone, right? And our signal will be in between
this Nyquist zone, right?

So, we can get the signal properly by using this particular bandpass filter, and if our IF is
at 70 MHz, we will be directly getting it back at the baseband signal. Now, this is good.
We have used the oversampling, and what we have is achieved? We have achieved a
very clean signal in our first Nyquist sampling rate.

Now let us say I use fs equal to 56 MHz, which is under-sampling, right? Because we
had said that at least 160 MHz was required to get it at the boundary. Now we have gone
to even lower. So, it is under-sampling, it is also called subsampling. So fs equal to 56
MHz. So, fs/2 will be 28 MHz, and this will be the first Nyquist zone, it will be the
second one with 56 and the third one when I add 28 to this one.So, 84 MHz it will be the
third zone.

Now our signal, which is starting from 60 to 80, will be here. So, basically, if it is folded
back on into the first Nyquist region, if it is at the 0, if we remove, down-convert it and
bring it to 0, it is folded back. Basically, just by sampling in the third Nyquist zone, we
are still able to achieve our signal, if you do not go further over sampling even with the
undersampling, we are able to get our signal, the only thing is that we have to get our
signal in the third Nyquist zone.

So, this is the example of our under-sampling, because, for over sampling, we have to
have high frequency, we have to put a constraint on the ADC and DAC. Instead, we can
use the under-sampling, subsampling, and by choosing the different Nyquist zone, we
can still recover the signal. Now finishing with that, we had seen the usefulness of the
subsampling and the under-sampling. And the concept of RF DAC when we were
directly sampling our frequency, right? Which is the RF frequency.

(Refer Slide Time: 24:10)

So, if we could do that then it is the ideal software-defined radio architecture. Now, this
is one example which is called RF sampling process, in this case in the receiver side we
actually have the RF signal directly, we are filtering, and we are directly sampling that,
we have discussed before that it is not possible, because our ADC and DAC has to
qualify the Nyquist criteria it has to have more than twice of the RF frequency.

So, the frequency of the RF signal is 1GHz, and then our sampling should be more than
2GHz. So, it was not possible, but by an interleaving of ADCs, we can actually increase
the sampling rate. So, for example, our one analog to digital converter can support up to
500 MHz, by using 4 ADCs and interleaving of them we can get four times of that
sampling frequencies.

So, this is the architecture which we are showing here, our input voltage, the input signal
is applied here, and they are applied to four different ADCs at different sampling
instances. So, in the right-hand side diagram, you can see that this is the clock input in
the black color, and we have a clock that is starting at the 0-degree phase, and according
to that at every rising edge, it takes the sample right?. So, the red one is showing that the
sample which is taking. We put the next clock, which is 90-degree phase-shifted with
respect to the previous one, and again at each rising edge, it is sampling one of that
sample. So, if the output of all these clocks are added together, we are having four types
of sampling with respect to the previous one. So, basically, if this was the sampling
instance, with the red by using all the sampling instances of different colors, we are
having multiplied by four times of sampling rate.

(Refer Slide Time: 26:15)

So, it is called the interleaving of ADCs for increasing the RF frequency. In the advanced
architectures, we are trying to achieve that, many RF companies are actually working on
this principle and giving the very high subject frequency of 2GHz and beyond.

Now, what is the limitation? First of all, the obvious limitation is that we require the
extra ADC instead of one, we require four. So, if the cost of the ADC is lower, then it is
good architecture. In some cases, we cannot avoid this complicity if we want to have that
high sampling rate. For example, in the 5G under the 6GHz range, it is proposed that we
might have to go for the signal bandwidth from 200 MHz to 400 MHz bandwidth signal.
So, by Shannon Nyquist theorem, we have to have 400 to 800 MHz sampling frequency.
So, this kind of structure we cannot avoid anymore. Now what are other limitations,
because they are using interleaving, it might not be perfect.

So, there might be frequency spurs at the interleaving boundaries. So, for example, we
have four levels. We will see these 4 images there, which are at the distance equal to
input frequency distance from the baseband (zero-IF). So, we will see this kind of
images, and there you have two levels of ADC. We will see only a single of this image
here. To conclude, the concept of under and oversampling can be applied to achieve the
manipulations for multi-rate processing and to achieve the advanced kind of architecture.