Why GaN PA Nonlinearity Is the Core 5G Emission Problem
GaN high-electron-mobility transistors (HEMTs) are the preferred technology for 5G base station power amplifiers because of their superior power density, breakdown voltage, and thermal performance — but operating them near the compression point, which is necessary for high efficiency, produces significant nonlinear distortions. These distortions are not a minor calibration issue: without compensation, a GaN PA operating at efficiency would produce adjacent channel interference that violates 5G New Radio emission standards and degrades system capacity.
The distortion challenge is compounded by the modulation schemes used in 5G. Wideband OFDM signals carry peak-to-average power ratios (PAPR) of 8 dB or more, which force GaN PAs to handle large instantaneous power swings across a wide bandwidth. Research from Texas Tech University (2023) on a two-stage GaN PA operating across the 5G FR2 band demonstrated that GMP DPD achieved ACLR improvements of 16.9 dB at 24 GHz, 17.3 dB at 28 GHz, 16.5 dB at 37 GHz, and 15.1 dB at 39 GHz — confirming that without DPD, GaN PAs at millimeter-wave frequencies would fail to meet regulatory spectral mask requirements entirely. According to ITU standards and 5G NR emission specifications, these margins are not optional.
AM/AM distortion: amplitude-dependent gain compression. AM/PM distortion: amplitude-dependent phase shift. Memory effects: frequency-dependent behaviour arising from thermal, electrical, and bias-network dynamics. 5G signals employ wider bandwidths, which make memory effects more pronounced and linearization processing correspondingly more complex.
The choice of PA technology also directly influences the complexity of the DPD system required. Research from Teltronic S.A.U. (2019) comparing LDMOS and GaN RF power amplifiers found that GaN PAs exhibit different nonlinearity profiles than LDMOS devices, affecting how much DPD processing burden must be carried by hardware and software resources. This means that the move to GaN — driven by efficiency and power density — has not simplified the linearization problem; it has reshaped it.
A two-stage GaN power amplifier operating in the 5G FR2 band (24–39 GHz) achieved ACLR improvements of 16.9 dB at 24 GHz, 17.3 dB at 28 GHz, 16.5 dB at 37 GHz, and 15.1 dB at 39 GHz after applying generalized memory polynomial digital predistortion, as demonstrated by Texas Tech University in 2023.
DPD Algorithm Architectures: From Memory Polynomials to Neural Networks
The memory polynomial (MP) and its generalization, the generalized memory polynomial (GMP), remain the dominant behavioral models for DPD in GaN PA systems. The MP model captures both static nonlinearity and memory effects using a finite set of polynomial coefficients extracted from feedback signals, making it computationally tractable for real-time implementation while providing sufficient modeling fidelity for most 5G deployment scenarios.
Research from M S Ramaiah Institute of Technology (2021) implemented a memory polynomial DPD in MATLAB/Simulink and showed high performance at polynomial orders 4 and 5, addressing spectral regrowth caused by wideband WCDMA-like waveforms. The indirect learning architecture (ILA) is the classical training method, where the inverse PA model is identified from output observations — though it has known limitations when PA nonlinearity is severe.
“A constrained multi-objective optimization algorithm that simultaneously maximizes RF output power while guaranteeing a prescribed linearity level outperforms classical indirect learning architecture while reusing the same predistorter structure — a critical finding for base station operators who must balance power efficiency with compliance.”
Multi-objective optimization represents a significant departure from simple ILA training. The University of Bologna (2021) demonstrated a constrained optimization algorithm that simultaneously maximizes RF output power while guaranteeing a prescribed linearity level defined by NMSE or ACPR targets. This approach outperforms classical ILA while reusing the same predistorter structure. Neural network approaches are emerging as a powerful complement to polynomial models: MediaTek Inc. (2021) demonstrated a deep neural network (DNN) framework that learns both the PA characteristics and the corresponding DPD simultaneously, incorporating frequency-domain loss functions to improve out-of-band emission suppression compared to naive time-domain mean square error training.
A hybrid predistortion approach that splits distortion correction between digital and analog domains has also been validated for 5G signals. Research from the American University of Sharjah (2022) proposes splitting the DPD function into a digital memory polynomial and an AM/AM-only look-up table suitable for RF analog implementation. Using 20 MHz and 40 MHz 5G test signals, this architecture achieves ACLR levels better than 50 dBc while requiring approximately 20% fewer coefficients than competing two-box models — a meaningful hardware complexity reduction.
For hybrid beamforming massive MIMO systems, research from Belgrade (2023) demonstrated that a single Real-Valued Time-Delay Neural Network (RVTDNN) with two hidden layers outperforms traditional memory polynomial DPD in a 64×64 fully-connected hybrid beamforming massive MIMO system with two RF chains — pointing toward neural network DPD as a viable path for next-generation deployments. Standards bodies including 3GPP continue to tighten emission requirements, making algorithmic advances in DPD directly tied to regulatory compliance.
A hybrid DPD architecture combining a digital memory polynomial with an AM/AM-only analog look-up table achieves ACLR levels better than 50 dBc on 20 MHz and 40 MHz 5G test signals while requiring approximately 20% fewer coefficients than competing two-box DPD models, according to research from the American University of Sharjah (2022).
Explore the full patent landscape for digital predistortion and GaN PA linearization in PatSnap Eureka.
Explore DPD Patents in PatSnap Eureka →Scaling DPD to Massive MIMO: Shared, OTA, and Grouped Architectures
The transition from single-PA DPD to massive MIMO base stations — where arrays contain 32, 64, or even 128 antenna branches each driving a GaN PA — is the defining engineering challenge of 5G linearization. A naive approach of deploying an independent DPD for each PA is computationally infeasible, and the industry has responded with three distinct architectural strategies.
Exploiting Spatial Degrees of Freedom
The first strategy exploits the spatial degrees of freedom (DoFs) inherent in massive MIMO arrays. Virginia Tech (2018) proposed a low-order DPD achieved by cascading adaptive zero-forcing precoding with DPD, using a 3rd-order polynomial sufficient because the excess DoFs of the massive array partially cancel intermodulation distortion in the intended beam direction. This reduces polynomial complexity requirements substantially compared to single-antenna DPD. Tampere University of Technology (2018) demonstrated that a single DPD unit placed at the precoder input can linearize an arbitrary number of PAs by minimizing nonlinear distortion of the combined signal at the intended receiver direction, while distortion in unintended directions partially self-cancels due to incoherent superposition.
Ericsson’s Grouped Combined MIMO DPD
Ericsson has industrialized multiple variants of this approach through its per-branch, combined, and grouped combined MIMO DPD patent family (US active, 2022 and 2024). The system provides the flexibility of per-branch DPD — maximizing individual PA linearization — combined DPD (a single linearizer for multiple branches), or grouped combined DPD (partitioned sets of branches). The per-branch DPD uses iterative learning control (ILC) combined with kernel regression. By grouping antenna branches, the patent achieves reduced computational complexity, reduced implementation complexity, and scalability — precisely the properties required for commercial massive MIMO deployment.
OTA Feedback and Crosstalk Compensation
A critical enabler for MIMO DPD is the feedback architecture. Southeast University (2020) presented a real-time single-channel over-the-air data acquisition strategy that captures signals from a fixed location to indirectly identify nonlinear behaviour of all PAs in the array. This avoids direct measurement at PA outputs — wiring that is impractical in a densely packed massive MIMO unit — and can handle mutual coupling between antenna elements, operating in real time without interrupting data transmission. This approach is directly supported by frameworks documented at ETSI for O-RAN compliant radio unit architectures.
Crosstalk between parallel PA channels further degrades shared DPD performance. Beijing University of Posts and Telecommunications (2022) proposed an auxiliary module that isolates crosstalk from per-branch nonlinearity. Experimental results on a 2×2 MIMO transmitter with −20 dB nonlinear crosstalk showed ACPR improvement exceeding 16 dB and NMSE improvement of more than 12.8 dB, with performance comparable to full parallel Hammerstein models but at significantly lower coefficient count.
In a 2×2 MIMO transmitter experiment with −20 dB nonlinear crosstalk, an iterative DPD method from Beijing University of Posts and Telecommunications (2022) achieved ACPR improvement exceeding 16 dB and NMSE improvement of more than 12.8 dB by separating crosstalk compensation from per-branch nonlinearity correction.
For hybrid beamforming systems where digital chains are fewer than antenna elements, the A²-DPD scheme from the University of Electronic Science and Technology of China (2020) uses continuously tunable analog predistortion modules to equalize the nonlinear behaviour of PAs in each channel before applying a single shared digital DPD, enabling one digital predistortion unit to linearize multiple PAs simultaneously.
Hardware Implementation: FPGAs, GPUs, and Low-Power Actuators
Deploying DPD in a 5G base station requires real-time hardware execution across potentially hundreds of parallel transmit chains — and the power budget for that computation is tightly constrained by the fact that 5G base stations already consume four times more power than 4G stations. The implementation landscape spans dedicated FPGAs, embedded multicore CPUs, mobile GPUs, and custom low-power logic.
GPU-accelerated DPD has demonstrated the feasibility of real-time operation for sub-6 GHz 5G NR. Tampere University (2019) presented a DPD implementation on a mobile multicore CPU with on-chip GPU achieving over 400 Msamples/s, sufficient for 100 MHz signal bandwidths in the 5G NR sub-6 GHz band. The learning stage executes on a dedicated CPU core, while the computationally intensive predistortion stage runs on the GPU. RF measurements at 3.7 GHz on two base station PAs confirmed compliance with 5G NR downlink emission requirements. Rice University (2017) also validated parallel DPD on ARM-based processors, exploiting parallel computing on embedded GPUs and multicore CPUs on real radio hardware platforms.
A 160 MHz bandwidth 5G DPD system requires a feedback ADC operating at approximately 900 Msamples/s — an extremely demanding and costly specification, as quantified in patent filings by the Chengdu Institute of Industry (2022). This has motivated Ericsson’s reduced-bandwidth observation approach, which linearizes a PA over a band extending to approximately five times the signal input bandwidth using a reduced-rate observation receiver.
Ericsson’s dominant patent-protected low-power hardware approach replaces standard fixed-point multipliers in the DPD actuator with approximate multiplication functions using look-up tables that store shift-and-add representations of complex values. Each LUT entry encodes a sum of power-of-two-weighted sign values, enabling approximate multiplication through only bit shifts and additions — drastically reducing power consumption while maintaining the linearization performance required for 5G NR. This approach is covered by active US, WO, and IN patents filed between 2019 and 2023.
Samsung Electronics has addressed adaptive DPD resource management through a channel-bandwidth-aware DPD circuit architecture (US active, 2023). The DPD circuit dynamically selects the number of active DPD units based on channel bandwidth and allocated resource blocks, while unused units are deactivated. This allows power consumption to scale with actual signal bandwidth — an important efficiency mechanism in TDD massive MIMO base stations operating with variable traffic loads. For O-RAN deployments, Sterlite Technologies (2024) has disclosed a method for enhancing DPD model performance in Open Radio Access Networks, where the O-RU reconstructs a linear distortion-free signal by comparing baseband output with input and compensating for missing spectral components.
Ericsson’s low-power approximate DPD actuator (US active, 2023) replaces fixed-point multipliers with look-up tables encoding shift-and-add representations of complex values, enabling approximate multiplication through only bit shifts and additions to reduce DPD power consumption in 5G base stations that consume four times more power than 4G equivalents.
Track hardware implementation patents for 5G DPD — from approximate computing to O-RAN integration — in PatSnap Eureka.
Search DPD Hardware Patents in PatSnap Eureka →Key Players and the Direction of Innovation
The competitive landscape for DPD in 5G GaN PA systems is concentrated among a small number of infrastructure vendors and a broader set of research institutions that are advancing the algorithmic frontier. Based on patent filing frequency and research output across more than 50 sources spanning 2013 to 2026, five organisations stand out.
Telefonaktiebolaget LM Ericsson holds the largest patent portfolio in this space, with multiple active US, WO, IN, and CN grants covering low-power approximate DPD actuators for 5G NR, per-branch/combined/grouped combined MIMO DPD using ILC and kernel regression, and wideband DPD with reduced observation bandwidth. This portfolio reflects Ericsson’s commercial deployment focus: minimising DPD power overhead while maintaining scalability to hundreds of MIMO antenna branches.
Qualcomm Incorporated is active in DPD at the component level, with patents covering DAC DPD circuits for RF digital-to-analog converter linearization and composite DPD systems combining memory polynomial and memoryless LUT predistorters. Qualcomm’s focus spans handset and infrastructure applications. Samsung Electronics contributes adaptive DPD circuit management for variable-bandwidth 5G NR systems. Nanjing Howking Technology holds patents in China and the US for dual-feedback-path DPD architectures targeting 5G MIMO systems.
On the research side, Tampere University has produced multiple high-impact papers on GPU-accelerated DPD and large-array beamforming DPD. Southeast University leads in OTA-based MIMO DPD feedback architecture. Virginia Tech and Beijing University of Posts and Telecommunications have advanced DoF-exploiting and crosstalk-aware DPD respectively. The University of Bologna contributes multi-objective optimization approaches for DPD training. According to WIPO patent trend data, 5G base station linearization has been among the fastest-growing RF patent categories since 2018.
The overarching trend across the dataset is the transition from single-PA DPD to system-level DPD: algorithms increasingly co-design the DPD with precoding, beamforming, and feedback acquisition strategies rather than treating the PA as an isolated component. Simultaneously, hardware complexity is being reduced through approximate computing, neural network pruning, and bandwidth-efficient feedback receivers. The IEEE Transactions on Microwave Theory and Techniques and related journals have documented this shift across multiple landmark papers in the dataset. PatSnap’s innovation intelligence platform tracks over 2 billion data points across 120+ countries, enabling R&D teams to monitor this rapidly evolving patent landscape in real time. Organisations building 5G infrastructure can use PatSnap’s R&D intelligence tools to identify white spaces and freedom-to-operate risks in the DPD patent space before committing to a hardware architecture.