Use digital predistortion with envelope tracking

Use digital predistortion with envelope tracking

Feature articles |
By eeNews Europe

LTE networks and signals present a range of problems for the RF front end in handsets as the signal characteristics differ greatly from previous 2G and 3G standards. Key metrics such as battery life, antenna performance, network coverage and thermal management are negatively affected by issues in the RF front end. Central to these challenges is the RF power amplifier (PA), its performance and power consumption.

Envelope Tracking (ET) is a new power modulation technique being used to optimise the performance and efficiency of the PA to help overcome the technical challenges presented by complex LTE signals. Indeed ET has largely been accepted by mobile handset manufacturers as the de facto standard RF PA power supply modulation architecture for LTE phones due to enter the market in 2014.

Alongside ET, some chipset companies are looking at a signal processing technique, Digital PreDistortion (DPD), as a way to optimise PA performance. Although many assume the two techniques perform similar tasks, it is important to clarify that DPD offers no efficiency benefits. DPD is a linearisation technique, and so ET and DPD are performing different functions in RF front end design. It is possible to use both techniques in isolation or together.

However, DPD is not an easy technology to develop. For designers there are important questions to resolve, such as: how do ET and DPD interact? How do I implement them in isolation? How do I use them together? If I have ET, do I need DPD, and vice versa?

As a result we have seen several different approaches from LTE chipset vendors. Some are not implementing DPD at all, some have implemented DPD but are not yet using it, others are using ET on its own and some are using DPD in conjunction with ET. Why is this? What might this split in approaches mean for the industry? This article looks at the technical advantages and disadvantages of DPD in the RF front end of handsets, the wider implications of vendors choosing the DPD route or not, and how DPD and ET can be used together most effectively.

Efficiency vs linearity

ET is a very fast power supply modulation technique that improves the energy efficiency of RF power amplifiers (PAs). It replaces the traditional fixed DC supply voltage to the RF PA with a dynamic supply voltage, which closely tracks the instantaneous amplitude, or "envelope" of the transmitted RF signal (See Figure 1).

RF PAs in handsets are typically operated in a classic Class AB configuration, and are only at their most efficient when the RF envelope waveform is close to peak power. This is not a problem with such traditional signals as 2G GSM, where information is encoded only in the phase of the signal – the amplitude is constant, and the PA can operate in this high efficiency mode all the time. GSM PAs consequently have typical efficiencies of 50-55%. However, as data rates increase from 2G to 3G and 4G, the increased spectral efficiency forces information to be encoded in the amplitude, as well as the phase, of the signal. When amplifying RF signals with high crest factors such as 4G LTE waveforms, the average efficiency of the PA drops significantly, with figures of 20-25% being common.

Modulating the supply voltage dynamically, in synchronisation with the envelope of the transmitted RF signal, ensures that the output device stays in saturation – its most efficient operating region – for a large portion of time, by providing just the minimum instantaneous supply voltage to the PA on a sample-by-sample basis. This can restore the PA efficiency to 50-55%, even for high crest factor 4G signals, offering the promise of 4G performance with 2G battery life.

This efficiency gain is a major benefit for product designers. However, if an ET PA is operated in maximum efficiency mode, then it will introduce distortion that compromises the linearity of the PA. So although you may be achieving maximum PA efficiency, some form of linearisation will be needed to correct this distortion (See Figure 2).

Why do we care about distortion?

In an RF Power Amplifier, there are several types of distortion which need to be considered and controlled. Amplifier distortion products falling within the bandwidth of the signal being transmitted will degrade the Error Vector Magnitude (EVM) of the signal at the receiver, reducing coverage and data rate. Higher frequency distortion products outside the transmit channel may cause interference to other users in neighbouring channels, and are usually constrained by Adjacent Channel Leakage Ratio (ACLR) regulatory specifications. For Frequency Division Duplex (FDD) systems, such as FD-LTE, distortion from the PA which spreads from the transmit band into the receive band can also degrade the sensitivity of the handset’s receiver, despite attenuation from the duplex filter – with more than 40 bands in use for LTE, this requires analysis of many different frequency offsets. Additional co-existence requirements, such as WiFi and GPS receivers in the same handset, place yet more constraints on the amplifier performance.

Managing PA distortion, without unduly sacrificing power consumption, is therefore a major consideration for designers of chipsets, PAs, and end-products alike.

So, what is DPD?

Digital PreDistortion (DPD), also known as Digital Precorrection, is a signal processing technique that compensates for nonlinearities in a transmission system. It works by inverting the measured gain and/or phase distortion of an amplifier, and “pre-distorting” the input signal to compensate for the PA distortion, with the goal of achieving a combined response which is linear (See Figure 3).

It is important to clarify that DPD does not by itself increase efficiency. DPD can in some circumstances enable a higher average output power from the PA, if the inherent nonlinearity of the PA is limiting the achievable output power due to EVM, noise or ACLR specifications. This increase in average power would also increase the PA efficiency, but DPD does not in itself improve PA efficiency. It is purely a linearisation technique.

DPD flavours

DPD also comes in several different varieties. There are two key distinctions: whether the DPD is “open loop” or “closed loop” (adaptive), and whether it is “memoryless”, or also corrects “memory effect” distortions.

Open loop vs closed loop DPD

Open loop DPD is a relatively straightforward technique that makes signal corrections based on a static model of the PA distortion, which may have been characterised at design time, or perhaps at factory calibration. Open loop DPD is typically implemented as a lookup table indexed by the instantaneous amplitude of the IQ sample, containing correction coefficients derived from the measured gain (AM) and phase (PM) distortion of the PA.

A significant disadvantage of open loop DPD when used with fixed-supply PAs is that the RF performance of the PA can vary significantly with power levels and temperature, and also as the PA device ages over its lifetime. Over time, this means that the PA behaviour can deviate significantly from the static model used to precorrect the signal, resulting in distortion creeping back in.

To overcome this, closed loop (adaptive) DPD involves capturing the output signal from the PA with a measurement receiver, comparing the measurement with the desired signal, and then updating the pre-distortion coefficients based on the measured response. The computational requirements of closed loop DPD are dependent on the linearity specifications for the target application, but can be significant as the algorithms needed to adapt the coefficients typically require high precision floating-point matrix mathematics. Although closed loop DPD offers significant advantages over open loop DPD, particularly in tracking variations in the PA due to temperature and ageing effects, it is significantly more difficult to implement, often requiring a high bandwidth A/D converter and memory buffer, and significant computational overhead. The frequency of adaptation is another parameter to explore – some systems can require constant adaptation to keep them stable, while others may only require infrequent adaptations every few minutes.

Memory and non-memory DPD

The most simple form of DPD is “memoryless” correction, which applies the gain and phase correction to a single IQ sample, typically implemented as a straightforward lookup table.

However, some PA distortion mechanisms also exhibit a “memory effect”, where the response of the PA depends not just on the instantaneous signal amplitude, but also on the amplitude of immediately preceding samples. To correct memory effects, the DPD must therefore model and correct several sequential samples, and also their interdependencies (crossterms). This increases the computational complexity exponentially as the memory depth increases – even correcting for a few nanoseconds of memory effect may involve 10x-20x the number of calculations needed for memoryless DPD, and will typically be implemented as a multi-tap digital filter, with multiple coefficient tables.

Both memory and non-memory DPD may be implemented as open loop or closed loop/adaptive systems, although memory DPD is almost always implemented as a closed loop system, as characterising the memory distortion of the PA at design time and decomposing it into individual correction components would be a very complex task.

So, while memory DPD is far more effective in correcting distortion in high bandwidth PAs, it is also far more computationally intensive – maybe by as much as 100x compared to memoryless DPD. Both closed-loop and memory DPD implementations add to silicon cost, development complexity, and engineering resource requirements.

DPD – advantages and disadvantages

To date no chipset vendors have implemented memory DPD in handsets, although it is commonly used in basestation design. For handsets, the typical option is memoryless DPD in combination with either open or closed loop operation.

In isolation, DPD can be a useful technique for RF designers. Where PAs have a relatively “soft” compression characteristic, such as in CMOS PAs with a fixed supply voltage, memoryless open loop DPD can help to compensate for the compression of the PA at its peaks when transmitting a signal with high Peak-to-Average Power Ratio (PAPR) such as LTE. In this mode, DPD is quite ‘light touch’ and adds only minimal complexity and power consumption. However, the variation of PA characteristics with temperature and under load mismatch can prove problematic.

DPD also significantly increases I/Q bandwidth requirements. DPD requires a wider bandwidth transmission, which includes a ‘correction bandwidth’ either side of the transmit channel, resulting in higher sample rates in the D/A converters and signal processing paths. If a closed loop DPD system is used, the observation receiver A/D also requires a similarly high bandwidth. This correction bandwidth is typically 3-5 times the transmit channel bandwidth, with a corresponding increase in signal processing load and dynamic power consumption of the digital logic. Although you can correct nonlinearities within this extended bandwidth, outside of the correction bandwidth DPD can degrade system performance by introducing additional noise. In FD-LTE systems, this can impact the receiver sensitivity due to the introduction of additional noise from the transmitter.

This represents a major increase in complexity for RF designers and the hardware and power penalties that result from this complexity are not to be ignored. Nujira estimates that closed loop DPD, assuming processing resources are shared with other parts of the system, probably costs around 1.5 mm2 of silicon and 50 mW of power dissipation in a typical handset implementation. Using a fully dedicated DSP core to support adaption of the coefficients in a closed loop memory DPD would introduce significantly higher silicon area and power consumption penalties, as illustrated in Figure 4.

Figure 4 – Estimated area and size impact of DPD

Developing DPD expertise in-house

A significant challenge for programme managers, RF designers and chipset developers is that there are no “standard” DPD algorithms, or readily available sources of IP cores. DPD adaptation algorithms rely on building a mathematical model of the PAs distortion behaviour; this is the subject of significant academic efforts, with even a cursory search of the literature revealing tens or hundreds of different approaches. Some algorithms exploit knowledge of the specific PA distortion mechanisms of the target device, and the optimum approach can change significantly with PA device technology. This approach may be practical for wireless infrastructure applications, where a single PA device may be used for many years, but may be harder to manage in handset applications, where PA devices are often sourced from multiple different vendors and used on the same PCB. A DPD solution designed for use with GaAs PAs may not work well with CMOS PAs.

A further limitation is the specialist engineering resource needed to understand, design, implement and test DPD. DPD development requires engineers with excellent knowledge of mathematics, practical real-time software skills, digital hardware, RF systems, and PA transistor behaviour. Debugging a DPD system requires engineers who can distinguish between a hardware limitation, a software bug, an algorithmic deficiency, or a poor PA design. This is not a skill set that is easy to hire, or quick to develop, and is typically limited to a few ‘gurus’ in each company.

These are not insignificant hurdles to overcome and the sheer engineering effort of DPD can dramatically increase time to market for RF vendors.

DPD – have and have not

The development and optimisation effort required to implement DPD could have severe consequent effects for the wider industry. DPD is a difficult technology to develop, and there are no easily licensed sources of DPD for the fast follower chipset vendors. The risk of not being able to reuse DPD algorithms across different PA technologies is exacerbated by the fact that there is a limited pool of engineers who can understand and design the complex algorithms required to implement DPD. As a result, we could see the industry segment into DPD “haves” and “have-nots”, with only the market leaders having sufficient resources to develop and test high performance DPD implementations.

This “two tier” split would be bad for business and bad for handset users. Reliance on DPD would force handset manufacturers to choose between a limited number of chipset suppliers, reducing competition and raising silicon prices.

However, this can be avoided. If we go back to the design of the ET power supply modulator itself, then properly implemented high performance ET can dramatically simplify, or even eliminate the need for DPD.

ET – making the problem worse?

Envelope Tracking is primarily being adopted as a power saving measure. However, there is a wide variation in performance of different ET systems, primarily dependent on the architecture and implementation of the ET power supply chip. Whilst bandwidth remains one of the key differentiators between solutions, ET swing range is an equally important performance metric, since simultaneously maintaining high efficiency, high bandwidth and a high swing range in the ET power supply is challenging. Bandwidth and swing range are also inter-related, since a wide-swing-range ET solution must track the waveform down into the “troughs”, where the bandwidth is highest. To avoid catastrophic distortion, caused when a low bandwidth ET power supply fails to keep up with the rising peaks of the waveforms, a higher ET voltage must be used to provide more headroom. This results in a higher average supply voltage to the PA, reducing efficiency.

A reduced swing range ET solution may give some increase in PA efficiency compared to a fixed supply, but at the expense of introducing additional distortion, which will then require DPD to maintain ACLR, EVM and receive sensitivity requirements. Figure 4 compares the performance of a fixed supply PA, a high swing range IsoGain ET system, and two intermediate ET systems with reduced swing range.

The ET systems with reduced swing range provide some efficiency benefit, but this comes at the cost of increased distortion, here shown as a degradation in ACLR. This additional distortion increases the workload on the DPD system. Only when increasing the swing range to achieve high accuracy IsoGain does the ACLR recover to the point where DPD is no longer required.

Reducing distortion without DPD

A key feature of Envelope Tracking systems is that the mapping of instantaneous RF signal amplitude to supply voltage is controlled digitally, using a programmable shaping table. The shaping table typically uses 128 entries, interpolated to 14 bit precision, allowing the supply voltage to be accurately defined for every possible I/Q sample amplitude.

At any given RF drive level, increasing the PA supply voltage will generally increase the gain of the PA, whereas decreasing the PA supply voltage will increase the efficiency of the PA.

The ability to independently control the gain at every possible I/Q amplitude also enables the designer to control the linearity of the PA, since the AM:AM linearity can also be thought of as “variation of gain with input power level”. An “IsoGain” shaping table can therefore be designed which selects the voltage required to deliver a constant gain at every power level (Figure 6).

The instantaneous supply voltage also affects the phase response of the PA, so the shaping table can be used to influence phase (AM:PM) distortion, as well as efficiency, gain, and additional characteristics such as the noise transfer function of the PA.

As the behaviour of the PA is now defined by a digital lookup table, the performance can also be optimised dynamically under software control, for example in response to changes in frequency band, operating environment or power level.

Using IsoGain ET with a high swing range ET modulator can therefore linearise the PA without requiring any DPD, offering chipset vendors a viable alternative to in-house development of complex DPD solutions.

Combining ET and DPD

Although ET can eliminate much of the need for DPD, there are instances where it is advantageous to use the techniques in tandem.

Where an ET modulator suffers from a low dynamic swing range (the variation in ET supply voltage), ET introduces unavoidable PA distortion and in these ‘sub-optimal’ implementations of ET, this must be corrected by DPD. As previously discussed, this constraint could limit adoption of those ET devices to those chipset vendors able to develop DPD in-house.

With a high swing range ET modulator that is able to track the waveform accurately enough to follow an IsoGain contour, DPD can still offer advantages by correcting the phase response of the PA (AM:PM distortion). DPD can also correct the residual amplitude (AM:AM) distortion that occurs at the minimum ET voltage, where the PA is no longer in compression (see Figure 7).


As already discussed, it can be problematic to use open loop DPD in systems without ET. A PA’s behaviour can change – small signal RF response varies particularly in response to temperature, or as a function of the age of the device. With open loop DPD alone, it is not possible to correct for this behaviour. By adding ET, much of this variability is eliminated, as the PA operates in compression over much of the envelope cycle, greatly simplifying the DPD implementation.

In an ET system, the PA is operated in compression throughout the modulation cycle. This not only improves the efficiency of the PA, but also provides direct control of the PA output from the power supply pin – when the PA is operating in compression, the amplitude of the input signal does not affect the output. In this circumstance, the linearity is no longer dependent on the small-signal RF performance of the PA, but is only dependent on the accuracy of the ET supply voltage. Since it is much easier to control the accuracy of the ET supply (which has a few tens of MHz bandwidth) than the accuracy of an RF amplifier operating at several GHz, the system is much more stable over temperature and process variation. As a result, when using ET it is possible to use open loop DPD in order to deliver the maximum linearity, whilst also maintaining excellent temperature stability.

The importance of high performance ET

Regardless of how designers approach the combination of ET and DPD, it is critical that the starting point is a high performance ET implementation that allows system engineers to find the perfect trade off between linearity, noise, gain and efficiency for each chipset and PA.

While a low swing range ET solution can deliver some of the promised efficiency gains, it will require DPD to fix the resulting distortion. A high bandwidth, high swing range ET solution can deliver a higher efficiency gain, and simultaneously minimise distortion.

By eliminating or drastically simplifying the requirements for DPD, ET can enable fast followers and cost-driven chipset vendors to deliver performance which competes with, or even exceeds, that of the market leaders, overcoming the complexity barrier traditionally associated with DPD, enabling consumers to benefit from both technologies on a wide range of mobile platforms.

About the Author

Jeremy Hendy is VP Sales & Marketing at Nujira; he has worked in semiconductor sales and marketing across multiple technologies for wireless communication and digital video. Previous positions include Marketing Director of wireless USB start-up Artimi, VP Marketing for Aspex Semiconductor, and Strategic Technology Director of Cadence’s Wireless and Multimedia business unit. He started his career with Texas Instruments, and holds a first class honours degree in Electronic Engineering from the University of Liverpool.

If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :    eeNews on Google News


Linked Articles