WHITE PAPER

Modeling RF Power Amplifiers and Increasing

Transmitter Linearity with DPD Using MATLAB
This paper discusses an approach for modeling and simulating RF power amplifiers (PAs) and digital predistortion algorithms
(DPDs) using MATLAB® and Simulink®. These tools let you rapidly develop ideas, validate designs, and build robust RF systems.
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

This paper discusses an approach for modeling and simulating RF power amplifiers (PAs) and digital predistortion
algorithms (DPDs) using MATLAB® and Simulink®. These tools let you rapidly develop ideas, validate designs, and
build robust RF systems.

Introduction
RF power amplifiers lie at the front end of most RF systems, including wireless communications and radar systems,
and are critical in ensuring the appropriate range of wireless systems. However, these amplifiers contain non-ideali-
ties that can cause performance degradation of the systems they drive when deployed in the real world. To mitigate
this performance degradation, you can apply linearization algorithms including DPD as part of the transmitter or
receiver chain. DPD algorithms aim to preemptively distort the waveform to be transmitted to counter the nonlinear
effects of the power amplifier. Since these waveforms are often wideband, they must typically be implemented on
FPGA/ASIC to achieve real-time performance.
DPD algorithms should be designed in the context of the full system because performance is largely dependent on
the system in which it is placed. Simulation provides an early-stage behavioral model of the full system and makes it
possible to develop the DPD algorithm early without dependency on hardware. System design and simulation with
Simulink also provides a framework to validate the algorithm and other components later in development and can
produce repeatable results that are hard to achieve with real RF hardware in the lab. In addition, you can use
Simulink to create bit- and cycle-accurate models of algorithms as they would run on firmware. You can also gener-
ate production-quality HDL code from these models.
In this paper, we discuss the challenges around RF power amplifier design with a focus on RF imperfections, as well
as a workflow for modeling PAs and increasing linearity with DPD (Figure 1).

Figure 1. DPD development workflow.

W H I T E PA P E R | 2
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Challenges in RF Power Amplifier Design

The design of RF transmitters is becoming more challenging. The emergence of new standards such as 5G and the
demand for multistandard transmitters present a range of contrasting design decisions. This includes tradeoffs on:
• Size: The transmitter might be confined to a small portable device or a limited real estate.
• Weight: Portability of the design might be key in ensuring its usability.
• Power: The device may be solar or battery powered.
• Bandwidth: Transmitters must work with increasingly larger signal bandwidths.
• Linearity: Amplifier gain will increase nonlinearly with input power.
• Frequency: Systems may need to operate in a range of frequencies from 10 Hz to 100 GHz.
• Integrated technology: Materials such as gallium-arsenide have improved high-frequency performance when
used for fabricating transistors.
• MIMO architecture: Hybrid beamforming architectures are considered for good tradeoff of power and flexibility,
for example.
• Configurability: Systems must be designed to be reconfigurable through the use of embedded algorithms to be
able to operate under different conditions.
The design strategy you choose will have an impact on more than one of these tradeoffs. For example, increasing the
bandwidth of the transmitter will typically draw more power and require the use of larger and heavier components
and more specialized IC technologies. Since the job of these components is to ensure that the signal power is large
enough, their requirements can tend to dominate the design.
PAs contain a number of non-idealities including nonlinearity and noise, which must also be assessed. Linearity
describes the behavior of input power vs. output power. Ideally this should be perfectly linear, but in reality, this isn’t
the case, as shown in Figure 2.

W H I T E PA P E R | 3
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Figure 2. Top: Input/output power characteristic (orange) of a nonlinear power amplifier showing compression, and the ideal linear characteristic
(black). Bottom: Input/output power characteristic (orange) of a nonlinear power amplifier showing memory effects, the static DPD characteristic
(blue) to achieve the ideal linear characteristic (black).

When input power increases beyond a certain point, output power begins to increase in a nonlinear fashion through
an effect known as compression. In general, this begins to happen at lower input power as the target carrier frequen-
cy of the system increases, causing high-frequency amplifiers to be less efficient. More crucially, waveform distortion
also takes place. A further complication here is that power amplifiers contain memory effects such that output power
will depend on input power at the current time as well as previous points in time. This is shown in the lower area of
the two plots in Figure 2. Various techniques, including DPD, can be used to linearize the behavior in the PA to
make it more deterministic.
Rapid prototyping is key to reaching an optimal amplifier and linearization design that can meet the many require-
ments listed above. However, this number of complex system requirements causes the design space to explode, which
makes physical prototyping too time-consuming and impossible to scale. Simulation can help in producing and test-
ing many prototypes to reduce development time and reach an optimal RF transmitter design.

Modeling RF Power Amplifiers for Efficient System Design

Power amplifier behavior is affected by nonlinear behavior and memory effects, and hence depends on input signal
characteristics such as signal bandwidth, frequency, peak-to-average power ratio, modulation, and loading condi-
tions. Most PA models that capture these effects are derived from Volterra series, often using complex identification
procedures. Proprietary (“black box”) models of power amplifiers exist, but these tend to be problematic as they lack
transparency, are hard to customize, and are difficult to understand when used in complex system simulations.

W H I T E PA P E R | 4
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

MATLAB offers a simple, fast, and customizable (“white box”) fitting procedure to create a model of a PA that can
then be integrated into a larger system simulation. The model is based on [1]:

where:
a are the PA model coefficients
k is the degree of nonlinearity, K is the number of degrees to which the PA is modeled
m is the memory depth value, M is the memory depth of the PA model

Note that when M is equal to 1, the above model describes purely static characteristics such as AM-AM/AM-PM.
The fitting procedure uses input and output data from the PA to extract the PA model coefficients based on chosen
values of K and M. Once fitting has been performed and a PA model created, the quality of fit is assessed. Values of
polynomial order K and memory depth M are then iterated on until the model performs as desired. Iteration can be
a manual process, or it can be automated within MATLAB.
As an example, consider the following complex IQ data capture collected at the input and output of the PA when an
LTE modulated test signal is injected. The input and output signal magnitudes as a function of time are shown in
Figure 3.

Figure 3. Input/output PA characteristics plotted as a function of time, showing the absolute value of the input waveform (blue),
and the absolute value of the output waveform (red).

W H I T E PA P E R | 5
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

The power transfer characteristic plotted from this data is shown in Figure 4. Here, the x-axis shows the input power,
and each point represents the instantaneous power gain. A straight horizonal line would imply that the PA is per-
fectly linear, and points spreading from this line would imply memory effects in the PA. Hence, in this case it can be
deduced that the PA is mildly nonlinear and has memory effects.

Figure 4. Input/output power characteristics of the PA. The X-axis is the absolute value of the input voltage,
and the Y-axis is the magnitude of the PA power gain.

The MATLAB based fitting procedure is now used to extract coefficients to feed a PA model and measure the quality
of fit. The white-box MATLAB function poweramp _ helper.m is used to compute a least-squares fit solution
with the aid of the MATLAB backslash operator. A snippet of the code is shown in Figure 5.

W H I T E PA P E R | 6
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Figure 5. Example of the identification procedure for the PA model coefficients implemented with MATLAB.

Half of the dataset is used to create the model, and the full dataset is then used to evaluate the quality of the fit by
feeding input data into the PA model and comparing the output data from the model with that recorded from the
real PA. Starting with a relatively low order and memory depth equal to K=3, M=3 results in the fit shown
in Figure 6.

Figure 6. Input/output power characteristics of the PA, showing the measured characteristic (blue), the result
of the fitting with memory depth equal to three and polynomial order equal to three (red).

W H I T E PA P E R | 7
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

The model points (red) show fewer memory effects compared with the real data (blue) and a little more linearity. The
signal standard deviation, which measures the quality of the fit, returns 5.148%. This captures both nonlinear behav-
ior and memory effects.
Increasing the polynomial length to K=4 while keeping memory depth to M=3 improves the PA model, reducing the
signal standard deviation to 4.2384% as shown in Figure 7.

Figure 7. Input/output power characteristics of the PA, showing the measured characteristic (blue),
and the result of the fitting with memory depth equal to three and polynomial order equal to four (red).

However, higher values of K and M don’t always mean an improved fit. To the contrary, an overfitted model might
cause numerical instability. This is especially the case for this data set since it comes from a quite linear PA. For
example, increasing the polynomial length and memory depth to 10 causes the signal standard deviation to rise to
4.6972% and a weaker fit to a more complex model.
When considering the quality of fit, it is important to analyze quality metrics in the frequency domain as well as the
time domain. Figure 8 shows the spectrum of the predicted signal from the fitted PA model alongside the measured
signal when memory depth is M=3 and polynomial length is K=4.

W H I T E PA P E R | 8
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Figure 8. Power spectrum of the output characteristics of the PA, showing the measured characteristic (yellow) and the result of the fitting with
memory depth equal to three and polynomial order equal to four (blue).

Here the yellow line represents the frequency spectrum of the captured PA output data, and the blue line is the spec-
trum of data collected from the PA model output when injected with the captured PA input data. Plotting the differ-
ence between these two spectra gives the signal shown in Figure 9.

Figure 9. Power spectrum of the error signal representing the difference between the measured and predicted output of the PA using a memory
polynomial model with memory depth equal to three and polynomial order equal to four.

W H I T E PA P E R | 9
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

As can be seen, the fit is generally good in the frequency domain for this model. However, it has suboptimal
performance at the main passband and spectral edges, which could be improved.
To further improve the quality of the fit, a PA model that includes leading and lagging memory cross terms
could be used:

This model and associated fitting procedure are also described in more detail in [1] and provided as white-box func-
tions in the poweramp _ helper.m script. It helps to better characterize memory effects while still providing a
causal response. Using a memory depth of M=5 and a polynomial length of L=7 produces a model with a much
reduced signal standard deviation of 2.3833% (Figure 10). Once again, the yellow line represents the spectrum from
the captured PA output data, and the blue line represents the spectrum from the modeled PA when injected with
captured PA input data.

Figure 10. Power spectrum of the output characteristics of the PA, showing the measured characteristic (yellow)and the result of the fitting using a
memory polynomial model (yellow) including cross-terms with memory depth equal to five and polynomial order equal to seven.

W H I T E PA P E R | 10
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

In the frequency domain, the maximum error in the predicted signal is just -50 dB, as shown in Figure 11.

Figure 11. Power spectrum of the error signal representing the difference between the measured and predicted output of the PA using a memory
polynomial model including cross-terms with memory depth equal to five and polynomial order equal to seven.

Hence, memory cross terms improve the quality of the PA model at the cost of additional complexity. Since the
MATLAB code used here is white-boxed, it can be reviewed and modified. The same framework could also be used
to model a completely different PA modeling and fitting procedure such as those based on Doherty models [2,3].
In these examples, the data used was collected by passing a waveform through a PA. The waveform used here must
be representative in order to create a good PA model. It is possible to create these waveforms synthetically in
MATLAB if required. Specifically, for LTE, WLAN, and 5G systems, signal generators exist within LTE System
Toolbox™ and WLAN System Toolbox™ to generate standard compliant waveforms. Instrument Control Toolbox™
can also be used to pass synthetically generated or prerecorded waveforms to and from lab-based test and measure-
ment hardware to exercise the PA for the purposes of automated test.

Designing a DPD Algorithm to Increase Transmitter Linearity

Digital predistortion (DPD) algorithms can be placed before the PA to predistort the signal being transmitted. The
DPD characteristic is designed to be the inverse of the PA characteristic and as such creates a behavior that is linear
overall because the two effects cancel one another. As a result, understanding and characterizing the behavior of the
PA is a crucial part of DPD design.

W H I T E PA P E R | 11
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

However, the practical design and implementation of a DPD algorithm is far more complex than inverting the non-
linear characteristic (Figure 12). Since the PA is affected by memory, the DPD algorithm must be adaptive and oper-
ate in a closed feedback-loop configuration to compensate for PA effects. In addition, DPD algorithms are digital
algorithms implemented in baseband, whereas PAs are RF components operating at high frequency. Upconversion
and downconversion stages are required to make the two systems operate together. Non-idealities associated with
doing this will affect the DPD behavior. Timing imperfections are particularly important as they can make the
system unstable. Loading effects of the antenna on the PA must also be considered as well as the DPD strategy being
used to adaptively compute the DPD coefficients.

Figure 12. Schematic representation of the closed-loop system including the power amplifier and the digital predistortion algorithm.

Once a suitable PA model has been determined, the DPD algorithm can be developed. When multiple signal
domains are being modeled (RF, digital, and analog), Simulink lends itself nicely as the modeling environment.
However, some blocks contain MATLAB code for design clarity. A DPD approach based on [4] is modeled. This
approach describes an indirect learning architecture implemented by a recursive predictive error method (RPEM).
The top-level hierarchy consists of a Simulink block to adaptively calculate the coefficients based on their current
values and the current output from the PA and another Simulink block to apply them to the incoming signal before
feeding to the PA (Figure 13).

Figure 13. Schematic representation of indirect learning architecture used for DPD implemented by a recursive predictive error method (RPEM). The
bottom block calculates the DPD coefficients based on their current values and the current output from the PA; the top block applies the coeffi-
cients and predistorts the incoming signal before feeding to the PA.

W H I T E PA P E R | 12
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

The coefficient calculation block contains the RPEM core, as shown in Figure 14.

Figure 14. Simulink block diagram for adaptively computing the DPD coefficients.

The RPEM block in Figure 14 computes the coefficients based on [2] as shown in Figure 15.

Figure 15. Simulink block diagram for adaptively computing the DPD coefficients following the RPEM algorithm.
An annotation describes the implemented math.

The RPEM core computes coefficients based on a memory length M and polynomial depth K in a similar way to the
strategy used for computing the PA polynomial coefficients. The higher the value of M and K, the higher the level of
computation. This generally results in larger resource and latency requirements to implement it in real hardware.
The DPD Simulink block that applies the computed coefficients to predistort the incoming signal is described in
Figure 16.

W H I T E PA P E R | 13
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Figure 16. Simulink block diagram of the DPD algorithm used to predistort the PA input signal.
A MATLAB function is used to perform the matrix multiplication between the incoming signal and the DPD coefficients.

Basic Simulink blocks are used in this design. Simulink could be used to create alternative strategies for computing
coefficients and applying them to the incoming signal using a similar framework.

System Simulation for Early Verification of DPD Algorithm Performance

The PA and DPD algorithm models should now be integrated and tested together within a system simulation model.
For this, a circuit envelope simulation is performed using RF Blockset™ [5]. This allows for a fast RF simulation while
accounting for nonlinear effects, impedance mismatches, and timing imperfections [6].
Figure 17 shows a simple behavioral model of a transmitter that is fed with two sine wave tones.

Figure 17. Simulink block diagram of a simple behavioral model of an RF transmitter,

an observer receiver path, and an adaptive DPD algorithm tested with a two-tone signal.

The transmitter model contains different amplification stages, a quadrature modulator, and a PA model (orange)
that uses the PA coefficients calculated previously. The PA is loaded by an antenna that is described using
S-parameters. A coupler is used to “observe” the output of the PA, and downconvert it to an intermediate frequency.
All RF component blocks in the transmitter model come from the RF Blockset library and use circuit envelope simu-
lation. In addition, component models contain RF imperfections such as the nonlinearities defined within the input
amplifier as described in Figure 18.

W H I T E PA P E R | 14
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Figure 18. Mask parameters of the nonlinear input amplifier.

The S-parameter antenna model could be captured from a real antenna that has been characterized by a VNA or
using data available from the manufacturer. Alternatively, Antenna Toolbox™ could be used within MATLAB to
build a representative model of the antenna and extract the S-parameters to be integrated in to this model [7].
Antenna Toolbox provides a full EM solver and uses Method of Moments to analyze the performance of antennas
from a prebuilt library as well as custom built antenna structures.
The surrounding testbench injects a two-tone sine wave into the transmitter, and the spectrum analyzer is used to
visualize the output from the PA and compute the third-order output IP coefficient (OIP3). Prior to enabling the
DPD algorithm, the OIP3 value is around -36 dBm with the third-order tone is -32 dBc, as shown in Figure 19.

Figure 19. Power spectrum of the output signal of the RF transmitter tested with a two-tone excitation.
When the DPD is disabled, the third-order intercept point is approximately 36 dBm.

W H I T E PA P E R | 15
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

However, after enabling the DPD algorithm, the OIP3 value drops to -40 dBm with the third-order tone at -41 dBc,
as it shown in Figure 20.

Figure 20. Power spectrum of the output signal of the RF transmitter tested with a two-tone excitation.
When the DPD is enabled, the third-order intercept point is approximately 40 dBm.

In a similar way to the PA modeling exercise, the order and memory length of the DPD algorithm could be modified
and the effect on performance of the DPD strategy at the system level observed. This simulation model facilitates fast
iterative testing of different DPD configurations to quickly and more easily optimize the algorithm. It also provides
an environment for early-stage verification of the DPD approach, perhaps before any physical prototype of the PA
even exists.
The model can also be tested using a modulated waveform. In Figure 21, a QAM signal is used and the effects of the
DPD algorithm are assessed by measuring the difference in EVM before and after the DPD algorithm is applied.

Figure 21. Simulink block diagram of a simple behavioral model of an RF transmitter,

an observer receiver path, and an adaptive DPD algorithm tested with QAM signal.

W H I T E PA P E R | 16
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Before the DPD algorithm is enabled, the EVM has an RMS value of 17.5% (Figure 22).

Figure 22. Constellation of the output signal of the RF transmitter tested with a QAM signal. When the DPD is disabled, the EVM is 17.5%.

After the DPD algorithm is enabled, this drops to 4.0% (Figure 23).

Figure 23. Constellation of the output signal of the RF transmitter tested with a QAM signal. When the DPD is enabled, the EVM is 4%.

W H I T E PA P E R | 17
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

The constellation points are more compacted after the DPD algorithm is applied.
Once the DPD algorithm has been developed and tested in simulation, a next step could be to test it with a real PA.
This could be done by keeping the DPD algorithm in Simulink and interfacing it with test and measurement hard-
ware to drive a baseband signal through the DPD, out to the real PA and then back in to Simulink for analysis and
DPD adaptation. This can be achieved using Instrument Control Toolbox [8]. This approach was employed as a part
of the workflow for DPD design at Huawei using Simulink. In [9] the authors describe the benefits of using
MATLAB and Simulink for DPD design, which enabled them to:
• Perform closed-loop simulation of designs containing both analog/RF and digital components, such as digital pre-
distortion (DPD) for RF power amplifiers
• Quickly develop a flexible, high-performance hardware development platform at the beginning of the R&D pro-
cess using a seamless interface to RF instruments
• Quickly build an automatic verification platform between software and hardware
• Use a single platform for hardware development, including reference models, fixed-point conversion, and automat-
ic C and RTL code generation
• Reuse models for bit-true verification of floating-point, fixed-point, and RTL code

Implementation and Verification of the DPD Algorithm Using Hardware

After experimenting with different DPD strategies in simulation, the next step is to prototype the algorithm outside
of Simulink. As transmitter bandwidth increases, the DPD algorithm must operate at a faster rate to achieve real-
time performance and meet latency requirements. Because of this, FPGAs are common prototyping architectures for
DPD algorithm testing and verification. Algorithms are generally deployed to FPGA as fixed-point, so the first step
to performing rapid prototyping is to convert the DPD algorithm from floating point to fixed point and to analyze
the impact of fixed-point precision on the performance of the algorithm.
Fixed-Point Designer™ [10] provides a guided workflow to assist in converting a model from floating point to fixed
point. A bit-true simulation model can be created, and the performance tested in simulation, in advance of deploy-
ment to FPGA. This is particularly significant in models with feedback configurations, such as the DPD model
under discussion here, since using finite precision arithmetic can affect stability and locking time. This is notably
difficult to detect and overcome when the algorithm is on FPGA. Simulation can help diagnose these issues more
easily and earlier in the design cycle.
To convert the DPD algorithm modeled within the system simulation in the previous section, a data type conversion
block is placed in advance of the DPD algorithm (Figure 24).

W H I T E PA P E R | 18
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Figure 24. Simulink block diagram of a simple behavioral model of an RF transmitter,

an observer receiver path, and an adaptive DPD algorithm using finite precision arithmetic.

The Data Type Conversion block parameters are configured to convert the input signal from floating point to a
12-bit signed signal with a fractional length of 16 bits (Figure 25).

Figure 25. Parameters of the Data Type Conversion block to convert the input signal from
double floating-point precision to fixed-point finite precision arithmetic.

W H I T E PA P E R | 19
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

As shown in Figure 25, this implies a representable range of approximately ±0.03 and signals can be represented to
the nearest 1.5e-5. The floor operator will be used to achieve the rounding required to do this. Floor is a good opera-
tor for FPGA implementation since it requires only simple logic to implement. In addition, signals exceeding the
limits of the maximum and minimum values will saturate rather than wrap. This is a more complex behavior to
implement but is safer when there are multiple feedback loops. These two strategies to rounding and saturation will
be adopted throughout the DPD design.
The recursive predictive error method uses the transmitter input and PA output to compute the coefficients. The
error between the PA input and the predistorted PA output is closed in a feedback loop to compute future DPD coef-
ficients, as shown in Figure 26.

Figure 26. Fixed-point implementation of the RPEM algorithm for computing the DPD coefficients using finite precision arithmetic.

The DPD coefficients are computed using multiple matrix multiplications, inversions, and feedback loops, so con-
verting the algorithm to fixed point isn’t straightforward. Specifically, it isn’t clear what word length and fractional
lengths are required for each of these internal signals to allow the algorithm to work correctly and efficiently without
using excessive resources. The Fixed-Point Tool in Simulink will be used to help in the conversion of the coefficient
estimation subsystem illustrated above.
Initially, all signal lines are set to override as double. The Fixed-Point Tool then runs the whole model and collects
and analyzes data from the part of the system to be converted at each signal line, as shown in the report in Figure 27.

W H I T E PA P E R | 20
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Figure 27. Fixed-Point Tool report of the simulation dynamic range of all the signals in the DPD subsystem.

As can be seen for the output of the block “Product 5,” the range of outputs produced during the simulation was
around -66.5 to 54. A histogram (Figure 28) of the data values that were represented during the simulation run is
also shown, to give an idea of the dynamic range required by the signal line.

Figure 28. Histogram of the simulation dynamic range of one of the internal signals used to compute the DPD coefficients.

W H I T E PA P E R | 21
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Here, the signal dynamic range is large, and the output is both positive and negative. This analysis can help to
inform decisions on appropriate data type choice.
Given a fixed word or fractional length, the tool can recommend a suitable fractional or word length for each of the
signal lines in the system under design. For example, you can configure word length to 16 bits and request an opti-
mal fractional length to be computed, as shown in Figure 29.

Figure 29. Automatically calculating fraction length in Fixed-Point Tool.

Requesting a proposal under this configuration results in the proposed data types shown in Figure 30.

Figure 30. A Fixed-Point Tool report after automatic calculation of the signal fractional length for a specified word length.

W H I T E PA P E R | 22
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

The Fixed-Point Tool proposes that the output from the “Product 5” block should be signed with 8 fractional bits.
This results in a representable range of ±128 with a precision of around 0.004. All data types proposed can be accept-
ed and then the simulation rerun with these new types. In this case, the EVM now increases to 4.9% (Figure 31) from
the 4% that was achieved when using double precision.

Figure 31. Constellation of the output signal of the RF transmitter tested with a QAM signal using finite precision arithmetic for the DPD algorithm.

In the histogram of simulation ranges captured for the “Product 5” block, a number of underflows occur, as shown
in Figure 32.

Figure 32. Histogram of the simulation dynamic range of one of the internal signals used to compute the DPD coefficients, showing the dynamic
range covered by a 16-bit word length (blue) and the data affected by underflows occurring due to lack of sufficient precision (yellow).

W H I T E PA P E R | 23
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

The fixed-point configuration can be easily modified within the Fixed-Point Tool and the system respecified. The
word length is now increased to 32 bits, which provides more coverage of the simulation data. The fractional length
of the output of the “Product 5” block proposed by the Fixed-Point Tool is now 24, and the precision now drops to
almost 6e-8. The updated coverage on the histogram of logged data points is shown in Figure 33.

Figure 33. Histogram of the simulation dynamic range of one of the internal signals used to compute the DPD coefficients, showing the dynamic
range covered by a 32-bit word length (blue)and the data affected by underflows occurring due to lack of sufficient precision (yellow).

In addition, the EVM also reduces to 4.3%. As well as designing signal line data types for the entire design, individu-
al signal lines can be isolated and analyzed. Through this process, the whole model can be converted by selectively
focusing on critical blocks and estimating the impact on system performance. By selecting a larger word length, even
if just for one of the internal blocks, the EVM can reduce and the algorithm converge more rapidly. A decision must
then be made as to whether the additional cost is worth the improved performance.
The Simulink subsystem to compute the DPD coefficients has now been converted to a fixed-point equivalent. The
Simulink subsystem to apply the DPD coefficients to the incoming signal must also be converted since this will also
be targeted to the FPGA. For this, an alternative, new and fully automated approach to the assisted conversion work-
flow described above will be used. Here, the tool will suggest an optimum word length and fractional length for each
signal line by using an optimization approach to meet an acceptable conversion error defined by the user.
The workflow is MATLAB based but can drive a Simulink model. First, an acceptable range of word lengths to be
used in the system under design must be defined as well as the maximum relative and absolute errors that are
acceptable at the output of the system. These constraints feed the optimization algorithm, which suggests appropri-
ate data types. For the DPD subsystem, acceptable word lengths are specified to be between 8 and 32:

%% make the options object

options = fxpOptimizationOptions;

options.AllowableWordLengths = 8:32;

W H I T E PA P E R | 24
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Acceptable relative and absolute tolerances at the subsystem output are specified as 1e-3 and 1e-4, respectively:
%% set the behavioral tolerances to be used when comparing

% fixed-point designs to the original baseline model

addTolerance(options, sud, 1, 'RelTol', 1e-3);

addTolerance(options, sud, 1, 'AbsTol', 1e-4);

The optimization is then started:

%% set max time to 30 minutes

options.MaxTime = 60 * 30;

options.Verbosity = 'High';

%% run optimization

result = fxpopt(model, sud, options);

%% explore the best fixed-point design

bestSolution = explore(result);

Successive iterations allow the proposal of fractional and word lengths for each signal line until the desired absolute
and relative tolerances are achieved. For the DPD subsystem, this takes approximately 10 minutes. After conversion,
the Simulation Data Inspector shows the simulation outputs from the DPD block before and after fixed-point data
types are applied as well as the difference between the two signals (Figure 34).

Figure 34. Plot of the error signal representing the difference between the simulation results of the output signal of the
DPD block computed with floating-point and fixed-point accuracy.

W H I T E PA P E R | 25
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Figure 34 shows that the differences between the final fixed- and floating-point results are well within the acceptable
tolerances. Following this, Fixed-Point Tool could be used for fine-tuning of the conversion.
Both the DPD coefficient calculation subsystem and the DPD subsystem to apply these coefficients to the incoming
signal have now been converted to fixed-point, and it has been verified in simulation that these subsystems still
operate as expected within a system-level simulation. With HDL Coder™, it is now possible to generate readable and
synthesizable VHDL® or Verilog® code that could then be deployed to FPGA. The example below will do this for the
DPD subsystem. The first step is to ensure that the Simulink subsystem being converted only contains blocks that
support HDL code generation. This can be achieved by running the HDL compatibility checker, as shown in
Figure 35.

Figure 35. Report of the HDL compatibility checker for the DPD block.

HDL code can then be generated by right-clicking on the subsystem and selecting “Generate HDL Code for
Subsystem.” Generated code is displayed in an HTML code generation report (Figure 36).

Figure 36. Report including the automatically generated HDL code for the DPD block.

Additional HDL specific options can be selected within the model to optimize resource utilization, speed, and area
of the design being implemented. Configuration options can be modified and the design and resulting generated
code iterated on. Figure 37 shows resource utilization, which might be of interest when iterating on the design to
ensure it fits on the FPGA being used.

W H I T E PA P E R | 26
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Figure 37. Report of the resources required for the automatically generated HDL code for the DPD block.

The generated HDL code could be used for rapid prototyping on FPGA but is also production quality and could be
used for FPGA or ASIC implementation of the final design. The generated HDL code can be verified by generating a
cosimulation testbench using HDL Verifier™. This product interfaces with a third-party HDL simulator such as
Mentor Graphics® ModelSim® or Cadence® Incisive® and reuses the Simulink testbench to verify that the HDL code
running in the HDL simulator and the Simulink model have the same behavior. This is particularly useful when the
system being converted is part of a closed-loop system because the interface allows the input being passed to the
HDL simulator to be changed adaptively over the course of the simulation dependent on the output from the HDL
simulator in previous time steps. This wouldn’t be possible if the HDL testbench stimulus was prerecorded. Figure 38
shows the automatically generated cosimulation testbench, which is run from Simulink.

W H I T E PA P E R | 27
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Figure 38. Simulink testbench for cosimulating the HDL implementation of the DPD algorithm with Mentor Graphics ModelSim.

Clicking the blue box launches the HDL simulator, which is ModelSim in this case. It then automatically loads in the
HDL code and configures the simulator for cosimulation. Once this is done, hitting the play button in Simulink
starts to stream data to the ModelSim environment and sends outputs back to Simulink on each timestep.

W H I T E PA P E R | 28
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Figure 39. Simulation results of the cosimulation testbench as shown in Mentor Graphics ModelSim.

Following the simulation, results can be analyzed within the HDL simulator (Figure 39) or within Simulink. The
difference between the Simulink model and the generated HDL code can also be compared in Simulink by using the
scopes inserted by HDL Verifier as part of the testbench generation process. The scope outputs in Figure 40 below
show the real and imaginary signal outputs from the DPD subsystem. In each case, the top plot shows the Simulink
model output, the middle plot shows the output from the HDL code running in ModelSim, and the bottom plot
shows the difference between the two. In both cases, it is clear that the Simulink model and HDL code are
equivalent.

W H I T E PA P E R | 29
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Figure 40. Comparison of the output of the reference Simulink model and the HDL implementation of the DPD block. The top scope plots the real
part of the output data, and the bottom scope plots the imaginary part. In each scope, the top trace is the cosimulation output, the middle trace is
the Simulink output, and the bottom trace is the error signal representing the difference between the two.

HDL Verifier can also be used for HIL simulations involving an FPGA by following a similar workflow. This vali-
dates that the algorithm behavior is the same following HDL synthesis.
In summary, the Simulink model is used to directly create and verify the implementation. With Model-Based
Design, the same environment can be used to design, test, prototype, and implement your algorithms. More details
and examples and videos of HDL workflows can be found on the HDL Coder and HDL Verifier product pages [11].

W H I T E PA P E R | 30
Modeling RF Power Amplifiers and Increasing Transmitter Linearity with DPD Using MATLAB

Conclusion
In this paper, we have discussed the top-down design process of a DPD algorithm and shown how you can use simu-
lation and Model-Based Design to accelerate this process and reduce risk by providing early-stage understanding of
the performance of the RF system and a workflow from simulation to implementation. This starts by modeling the
RF power amplifier using actual component characteristics. We explored RF component imperfections as well as
strategies to create a model by fitting to collected RF power amplifier data.
The PA model was then placed alongside a model of a DPD algorithm within a closed loop, and we performed sys-
tem-level simulation and analysis using OIP3 and EVM metrics to assess the quality of the linearization technique.
The DPD algorithm was then converted from a floating- to a fixed-point implementation. The effect on quantiza-
tion, timing, and other system imperfections was then assessed. Finally, we generated HDL code from the DPD algo-
rithm, and we covered techniques to optimize and verify the generated HDL code.
MATLAB and Simulink can be used to develop DPD algorithms, using the same models throughout the develop-
ment process to maintain consistency, reduce risk, accelerate design and implementation, and complement lab test-
ing and prototyping. You can find more information about this on the RF Blockset product page.

References
[1] Morgan, et.al., “A Generalized Memory Polynomial Model for Digital Predistortion of RF Power Amplifiers,”
IEEE Transactions on Signal Processing, Vol. 54, No. 10, October 2006.
[2] A. Zhu, “Decomposed Vector Rotation-Based Behavioral Modelling for Digital Predistortion of RF Power
Amplifiers,” IEEE Transactions on Microwave Theory and Techniques, Vol. 63, No. 2, February 2015.
[3] O. Hammi, “Augmented Twin-Nonlinear Two-Box Behavioral Models for Multicarrier LTE Power Amplifiers,”
The Scientific World Journal, Volume 2014, Article ID 762534.
[4] L. Gan, “Adaptive Digital Predistortion of Nonlinear Systems,” Ph.D. Thesis, Electrical and Information
Engineering, Graz University of Technology, Austria, April 2009.
[5] RF Blockset. mathworks.com/products/simrf.html
[6] Circuit Envelope Fundamentals: Simulate High Frequency Components with RF Blockset.
mathworks.com/help/simrf/gs/minimize-computations-for-rf-simulations.html.
[7] Antenna Toolbox. mathworks.com/products/antenna.html.
[8] Instrument Control Toolbox. mathworks.com/products/instrument.html.
[9] E. Zhu, “Developing a Radio Frequency System for Wireless at Huawei”, September 2018.
mathworks.com/content/dam/mathworks/case-study/huawei-customer-case-study-portrait.pdf.
[10] Fixed-Point Design Optimization.
mathworks.com/products/fixed-point-designer/features.html#data-type-exploration-and-optimization.
[11] HDL Code Generation and Verification. mathworks.com/solutions/hdl-code-generation-verification.html.

© 2018 The MathWorks, Inc. MATLAB and Simulink are registered trademarks of The MathWorks, Inc. See mathworks.com/trademarks for a list of additional trademarks.
Other product or brand names may be trademarks or registered trademarks of their respective holders.

W H I T E PA P E R | 31