The 3GPP specification for 5G NR EVM is precise. For PUSCH using 64-QAM modulation, 3GPP TS 38.101 sets an EVM ceiling of 8% RMS at the antenna port. Advance the modulation order to 256-QAM and that ceiling tightens to 3.5%. The PA has to meet both conditions across the full range of instantaneous power levels that an OFDM waveform presents, not just at a single CW operating point.
Most PA datasheets present a CW gain compression curve and a P1dB number. Those two parameters describe a necessary but not sufficient condition for passing 5G NR EVM. Closing the gap between a clean compression curve and a passing EVM measurement requires understanding AM-AM and AM-PM distortion and what they actually do to a complex modulated signal.
What EVM Measures and What the PA Controls
Error Vector Magnitude is the RMS magnitude of the difference between the ideal reference symbol locations and the actual transmitted symbols, normalized to the ideal symbol amplitude. In a 5G NR transmitter the total EVM budget is distributed across the digital-to-analog converter, IQ modulator amplitude and phase imbalance, local oscillator phase noise, and PA nonlinearity. At mmWave frequencies above 24 GHz, the PA tends to dominate the error floor, particularly under wideband modulation where instantaneous signal amplitude cycles over a large dynamic range within each burst.
The two PA-driven mechanisms are AM-AM and AM-PM conversion. AM-AM is gain compression: as input amplitude rises toward P1dB, the actual output amplitude deviates below the linear gain extrapolation. AM-PM is the phase rotation that accompanies that amplitude compression. An ideal linear amplifier produces zero AM-PM. A real device near saturation can produce 5 to 20 degrees of AM-PM at P1dB, depending on the process technology and bias conditions.
For a CW signal, these distortions appear as harmonic products and a saturated output power that is lower than the small-signal gain extrapolation predicts. For a wideband OFDM signal, they produce intermodulation between subcarriers, generating spectral regrowth outside the channel bandwidth and scatter in the IQ constellation. A vector signal analyzer measuring EVM integrates that constellation scatter across all active subcarriers. The result is an EVM floor that reflects both the depth of gain compression and the amplitude of the AM-PM curve at the operating power.
PAPR and the Backed-Off Operating Point
5G NR uplink OFDM signals carry a high peak-to-average power ratio. A 100 MHz NR carrier with 64 active subcarriers routinely reaches PAPR values above 10 dB at the 0.1% probability level in a complementary CDF measurement. This means if the transmitter sets average output power to meet the required EIRP, instantaneous peaks regularly drive the PA to within 1 to 2 dB of its P1dB operating point.
The straightforward mitigation is output power backoff: reduce average transmit power until peaks land comfortably below the compression knee. The cost is drain efficiency. A PA operating 8 to 10 dB below its P1dB is far from its peak power-added efficiency. For a commercial outdoor unit in a 64-element massive MIMO array, that efficiency penalty multiplies across every element. The heat dissipation per array panel scales with the product of per-element efficiency gap and element count, and that product is the fundamental thermal management constraint in 5G active antennas at mmWave frequencies.
GaN-on-SiC addresses this problem through power density. A GaN HEMT at 28 V drain bias delivers 5 to 10 times the power density of a GaAs pHEMT at the same gate periphery. Starting from a higher P1dB means the backed-off operating point still delivers the required output power at a junction temperature that does not demand exotic cooling. A device with a 10 W P1dB delivering 1 W at 10 dB OBO has a fundamentally different thermal problem than a device with a 2 W P1dB struggling to reach the same 1 W at an impractically small backoff. Falcomm's FA-2400 series design targets set drain efficiency above 40% at 6 dB output power backoff across Ku-band, because that is where 5G NR systems spend most of their operating time.
AM-PM Is the Harder Problem
Hardware teams frequently focus on gain compression because it is visible on a standard power sweep. AM-PM appears on the same measurement but receives less engineering attention at the component selection stage, even though it is often the binding constraint for high modulation orders.
Consider 256-QAM: the constellation has 16 amplitude levels and tight angular separation between symbols. An AM-PM curve that shifts 3 degrees between low and high instantaneous power levels introduces phase scatter that overlaps adjacent symbol clusters. At a 3.5% EVM budget, a 3-degree phase variation at the modulation rate consumes a significant fraction of the error margin before LO phase noise, IQ imbalance, and thermal noise have been accounted for at all.
Scalar predistortion, which boosts gain at lower input amplitude levels, partially corrects AM-AM compression. Correcting AM-PM requires full complex predistortion that tracks both the amplitude-to-phase mapping and its frequency and temperature dependence. GaN devices on SiC substrates exhibit lower AM-PM at a given fractional P1dB operating point compared to GaAs pHEMTs, in part due to different charge trapping dynamics and the absence of the low-frequency dispersion traps that are more prevalent in GaAs buffer layers. This is not a universal statement about all GaN processes; it depends on epi stack design and surface passivation quality. But it is a real advantage when the application calls for meeting 5G NR EVM at 256-QAM without the complexity of full per-element DPD.
Digital Pre-Distortion: Complement, Not Replacement
DPD is a standard subsystem in 4G and 5G base station radios. A lookup-table or Volterra-series predistorter running in the digital domain can reduce EVM by 10 to 15 dB under steady-state conditions. The question is not whether DPD works. The question is where DPD leaves the system exposed.
DPD convergence takes time. An adaptive training loop typically requires tens of milliseconds to converge after startup, after a power level change, or after a thermal transient. During that convergence window, the predistorter operates on a model that no longer matches the PA's current state. A PA whose AM-PM shifts by 5 degrees across a 20-degree ambient temperature change creates a tracking error that shows up as transient EVM violations at startup, during cell handover events, or after sleep-cycle recovery.
We are not arguing that DPD is avoidable in a 256-QAM system; for most deployments it is not. We are saying that the PA's inherent AM-PM characteristic determines the dynamic range demand placed on the DPD system, and how brittle EVM compliance becomes during the intervals when the predistorter is re-converging. A PA with better native linearity reduces the magnitude of what DPD must correct and shrinks the transient window during which EVM compliance is at risk.
mmWave-Specific Nonlinearity at Ku and Ka Band
Above 24 GHz, several effects compound the standard nonlinearity picture. First, the thermal time constant at the transistor junction is short enough that junction temperature partially tracks the signal envelope at baseband rates. A 100 MHz NR burst modulates the device's self-heating on a timescale comparable to the envelope variation, creating a memory effect where AM-AM and AM-PM depend not just on instantaneous amplitude but on the recent amplitude history. A memoryless DPD model does not compensate this correctly.
Second, at 28 to 40 GHz, package and interconnect parasitics represent a larger fraction of the transistor's impedances. A bond wire inductance of 0.5 nH that contributes modest reactive loading at 10 GHz produces a meaningfully different load at 35 GHz. If a package resonance falls within the signal bandwidth, AM-PM slope varies across the channel, producing frequency-dependent EVM degradation that a narrow CW measurement would not reveal.
A concrete planning scenario: a 100 MHz NR FR2 carrier at 28 GHz operating 256-QAM in a 48-element active antenna unit, where each PA element must contribute below 2.5% EVM at the element port. Achieving this without per-element DPD requires AM-PM below 4 degrees at the 6 dB OBO point and in-band gain flatness better than 0.3 dB across 100 MHz. These are the kinds of design targets that shape how the FA-2400 is being specified for 5G applications. Whether the full system meets the 256-QAM EVM threshold depends on chain calibration beyond the PA alone, but the PA's nonlinearity budget is where the calculation starts.
Reading a PA Data Sheet for EVM Compliance
When evaluating a PA for 5G NR compliance, look beyond the headline P1dB and efficiency numbers. Check which operating point the AM-AM and AM-PM curves are plotted at: CW at P1dB, or at the backed-off power level relevant to the target modulation scheme? Check the test waveform: CW measurements at compression look very different from modulated measurements at backoff. Verify the measurement temperature: efficiency and AM-PM both degrade at elevated junction temperatures, and a datasheet measured at 25 degrees Celsius ambient does not describe behavior at 85 degrees.
Requesting modulated EVM data, AM-PM curves across temperature, and pulsed I-V sweeps alongside the standard small-signal parameters costs little in the evaluation process and reveals substantially more about whether the PA closes the link budget. The distance between a PA that looks good in a datasheet and one that performs in a deployed 5G NR system is largely determined by the quality of that backed-off, modulated, temperature-characterized dataset.