Every RF power amplifier is nonlinear. Push it toward saturation for maximum efficiency and output power, and the output signal is a distorted version of the input. This distortion generates spectral regrowth (energy splashing into adjacent channels), increases the error vector magnitude (EVM) of the modulated signal, and violates the spectral emission masks set by regulators and standards bodies. The brute-force solution is to back off the amplifier input power until the distortion is acceptable, but this wastes 50% or more of the amplifier's rated power and efficiency.
Digital predistortion is the intelligent alternative. It applies a carefully computed inverse distortion to the digital signal before it enters the PA, so that the PA's own nonlinearity cancels the predistortion, producing a clean, linear output at high power. DPD lets you operate a PA at 2 to 4 dB closer to saturation while meeting linearity specifications.
1. PA Nonlinearity: What DPD Fixes
AM-AM Distortion (Gain Compression)
As input power increases, the PA's gain decreases. The transfer function curves from its linear slope toward a flat ceiling (saturation). This compresses the amplitude peaks of the signal relative to the valleys, distorting the signal envelope.
AM-PM Distortion (Phase Shift)
As input power increases, the PA also introduces a signal-dependent phase shift. This means the output phase is no longer a fixed offset from the input phase; it varies with the instantaneous signal amplitude. For high-order modulation schemes (64-QAM, 256-QAM), AM-PM distortion is often more damaging to EVM than AM-AM compression.
Memory Effects
In a memoryless PA model, the output at any instant depends only on the input at that instant. Real PAs have memory: the output depends on both the current input and the recent history of inputs. Memory effects arise from thermal time constants (the die temperature changes with the signal envelope), bias circuit dynamics (the drain supply voltage droops under high current), and trapping effects in GaN devices. Memory effects make DPD harder because the predistortion function must account for the time history of the signal, not just its instantaneous amplitude.
| Distortion Type | Cause | Effect on Signal | DPD Complexity |
|---|---|---|---|
| AM-AM | Gain compression | Envelope flattening | Low (static LUT) |
| AM-PM | Nonlinear capacitance | Constellation rotation | Low (phase LUT) |
| Memory (short) | Bias network, matching | Asymmetric spectral regrowth | Moderate (FIR DPD) |
| Memory (long) | Thermal, trapping | Long-term drift | High (Volterra, GMP) |
2. How DPD Works
The Core Idea: If the PA compresses a signal by applying function f(x), then apply f⁻¹(x) to the signal before the PA. The cascade of f⁻¹ followed by f produces a linear output: f(f⁻¹(x)) = x. The challenge is accurately estimating f⁻¹ in real time as the PA's characteristics change with temperature, aging, and signal statistics.
The Feedback Loop
- Transmit: The predistorted digital signal is sent through the DAC, upconverter, and PA to the antenna.
- Observe: A directional coupler samples a small portion of the PA output. This sample is downconverted and digitized by an observation receiver ADC.
- Compare: The DPD processor compares the observed PA output to the original input signal (delayed to account for pipeline latency).
- Adapt: An adaptive algorithm updates the DPD coefficients to minimize the error between the desired linear output and the actual PA output.
3. DPD Algorithms
- Lookup Table (LUT): The simplest approach. A table indexed by input amplitude that stores the gain and phase correction for each amplitude level. Fast, low complexity, but only corrects memoryless distortion. Suitable for narrow-bandwidth signals (< 20 MHz).
- Generalized Memory Polynomial (GMP): Extends the polynomial model to include cross-terms between the current sample and delayed samples. Handles short-term memory effects. The standard choice for 5G NR signals with 100 to 400 MHz bandwidth.
- Volterra Series: The most general nonlinear model with memory. Captures all orders of nonlinearity and all memory depths. Computationally expensive; used primarily in research and in systems where the PA exhibits strong long-term memory.
4. FPGA Implementation
Production DPD runs on FPGAs (Xilinx/AMD Zynq, Ultrascale+, or Intel Agilex) in the digital baseband processor. The DPD actuator (applying the predistortion) must process samples at the full signal bandwidth (often 500 MHz to 1 GHz for wideband signals). The adaptation engine (updating the coefficients) runs at a slower rate, typically updating the model every millisecond to every 100 milliseconds.
Key FPGA resources consumed by DPD: multipliers (for polynomial evaluation), block RAM (for LUTs and delay lines), and high-speed serial interfaces (for the observation receiver ADC data path).
5. DPD Performance Metrics
| Metric | Without DPD | With DPD | Improvement |
|---|---|---|---|
| ACLR (adjacent channel) | -30 to -35 dBc | -50 to -55 dBc | 15 to 25 dB |
| EVM (256-QAM) | -28 to -30 dB | -38 to -42 dB | 8 to 14 dB |
| PA efficiency (at rated power) | 15 to 20% (backed off) | 30 to 45% (near saturation) | 2x improvement |
RF Essentials provides the passive RF components in the DPD signal chain: directional couplers for the observation path, precision terminations, and waveguide assemblies for the high-power output stage.