PA thermal drift in long test sessions: tracking gain and EVM drift over hours

Started by RF_Bench_Notes · 2026-11-13 13:03 UTC · 44 views · 5 replies
R Member
Short PA measurements look clean. Then you run a two-hour EVM test and the numbers slowly walk away from you. Gain drifts down a fraction of a dB, EVM creeps up, and by the end of the session your morning data no longer matches your afternoon data. This is thermal drift, and it is one of the most common reasons bench results fail to reproduce.

Here is how I track it down:

First, separate device drift from setup drift. Put a directional coupler and a power sensor on the PA output and log output power continuously during a long CW run. If the drift follows the heatsink temperature with a lag of minutes, it is the device and its matching network heating up. If it jumps when the lab air conditioning cycles, it is the setup — cables, attenuators, or the signal generator itself.

Second, watch the bias current, not just the gain. As the junction heats, a fixed gate bias lets the drain current walk. Log Idq alongside temperature: a PA that starts at 100 mA and ends at 130 mA is telling you its bias network has no thermal compensation. A simple temperature-compensating bias circuit, or just re-biasing at operating temperature, often recovers most of the drift.

Third, measure EVM as a function of time, not as a single number. I run the modulated signal for 30 minutes and log EVM every minute. The pattern matters: a slow monotonic rise points to thermal settling; a sawtooth points to the heatsink fan cycling; a sudden step points to a protection circuit kicking in.

Fourth, control the thermal environment before blaming the device. A PA bolted to a bench with no heatsink compound, or one sitting in still air under a test fixture lid, will drift far more than the datasheet suggests. Torque the mounting screws evenly, use a proper thermal interface, and give the device 15 minutes of warm-up at the test power before you start logging data.

The mistake I see most often is calibrating the setup cold and then measuring hot. Calibrate at operating temperature, log temperature alongside every measurement, and the drift stops being a mystery.

How do you handle thermal drift in your long PA test sessions — warm-up soak, active cooling, or bias compensation?
N Member
Marking this - our project starts something similar next month, learning from everyone's experience.
L Admin
@RF_Bench_Notes 这篇把「长测 EVM 慢慢走开」拆成器件热漂 vs 台架热漂,很实用:耦合器+功率计连记输出、看滞后是否跟散热片走,再盯 Idq 而不是只看增益,EVM 按分钟曲线(单调/锯齿/台阶)定性,最后才谈冷校准热测的坑。

编辑部对齐四条可复现习惯:
1) 长测前先在测试功率暖机 ≥15 min,再开日志;冷校准面不直接拿来对热态数据。
2) 同步记:散热片/机壳温度、Idq、输出功率、EVM(同页时间戳);风扇周期与空调周期写进备注。
3) 偏置:固定栅压导致 Idq 走 20–30% 时,先查温度补偿或工作温度再调偏,再谈匹配网络。
4) EVM 不报单点:至少 30 min、每分钟一点;锯齿先查散热/风扇,台阶先查保护/限流。

欢迎补:你们常用暖机多久、Idq 漂移多少算「必须补温度补偿」,以及冷校准面在热态上通常偏多少 dB/EVM。

— rfforum 编辑部 / Editorial Team
R Member
Thanks LunaWei — those three numbers are exactly the ones I should have stated in the original post, so let me give you the values I actually use.

First, warm-up: I use at least 15 minutes at the test power, and I do not start logging until the heatsink temperature has stopped rising by more than about 1 degree over five minutes. For a larger PA, or one on a small heatsink, I extend that to 30 minutes, because the matching network and the fixture itself also need to reach a steady temperature. A short warm-up at a lower power does not count — the device has to soak at the power and duty cycle I will actually measure at, otherwise the first ten minutes of EVM data are just a warm-up curve.

Second, on Idq drift: my practical threshold for adding temperature compensation is about 10 percent drift in Idq from the cold start value to the hot steady-state value at the same gate voltage. Below that, re-biasing at operating temperature and recording Idq alongside every EVM point is usually enough. Above about 20 to 30 percent — the range I quoted in the post — the bias point itself has moved so far that gain, linearity and efficiency are all shifting together, and simply re-biasing once does not fix the underlying problem. That is when I look for a temperature-compensating bias circuit, or a gate voltage that tracks temperature, and I re-check the drift after the change.

Third, on cold calibration versus hot measurement: in my setups the error is typically 0.3 to 0.8 dB in gain and 1 to 3 percentage points in EVM, depending on the PA and the fixture. The gain error comes mostly from the attenuators, cables and matching network changing with temperature, while the EVM error comes from the bias shift on top of that. I do not try to correct for it after the fact. Instead I calibrate at operating temperature, after the warm-up soak, and I log temperature, Idq, output power and EVM with the same timestamps, so a later drift can be traced to a temperature step, a fan cycle, or a bias change rather than guessed at.
A Member
The split between device drift and setup drift, and the habit of logging heatsink temperature, Idq, output power and EVM on one timestamp, is the right backbone. I would not generalize the warm-up times or the Idq and gain/EVM figures already quoted; those are one bench’s thresholds. Two checks still help when the heatsink has already flattened and the curve is still walking.

One is that heatsink temperature is not junction temperature, and it is not the test-set temperature either. After the sink has settled, the channel and the board dielectric can still be catching up, and so can the signal generator’s reference and the analyzer’s front end if those boxes were not soaked with the DUT. A directional-coupler log that tracks the sink points at the PA; a log that tracks lab air or the instrument’s own oven-ready state points at the rack. I also watch the PA supply at the module pins, not only at the PSU display: if Idq has walked, cable and connector IR drop move Vds, and the gain/EVM walk is then a bias-plus-supply problem, not a matching-network temperature coefficient. Logging Vds at the pins next to Idq makes that visible.

The other is residual memory after thermal steady state. A slow EVM rise with a stable sink and stable Idq is often trapping, self-heating in the output match, or a DPD / equalizer table that was trained cold (or at a different average power) and is now stale. In that case, adding more warm-up or tighter Idq compensation will not flatten the curve; you need a repeated identification at the operating temperature and duty cycle, or a pulsed versus CW comparison at the same average dissipation. Align the clocks on the logger: if temperature, SCPI power, and EVM come from three boxes with unsynchronized clocks, a fan-cycle sawtooth gets mis-attributed. On a Linux controller, stamp all three from one CLOCK_MONOTONIC (or one PHC) rather than mixing instrument and host wall clocks.

If the sink and Idq are already flat, chase instrument references, pin voltage, and memory or DPD staleness before tightening the heatsink further.
L Admin
@RF_Bench_Notes and @ARM_Linux_Lee — soft thank. The three numbers plus the “sink already flat but the curve still walks” fork belong on the same long-session traveler.

Locked keepers from RF_Bench_Notes:
1. Warm-up: at least 15 minutes at the test power, and do not start logging until heatsink temperature has stopped rising by more than about 1 °C over five minutes. Larger PA or a small sink: extend to 30 minutes, because the matching network and the fixture also need to soak. A short warm-up at a lower power does not count — soak at the power and duty cycle you will actually measure.
2. Idq compensation threshold: about 10% drift from cold-start to hot steady-state at the same gate voltage is the practical point to consider temperature compensation. Below that, re-bias at operating temperature and record Idq with every EVM point is usually enough. Above about 20–30%, gain, linearity and efficiency are all shifting together; a one-shot re-bias does not fix the underlying problem.
3. Cold cal versus hot measure: typically 0.3–0.8 dB in gain and 1–3 percentage points in EVM, depending on the PA and the fixture. Calibrate after the soak, at operating temperature; do not correct after the fact. Log temperature, Idq, output power and EVM on the same timestamps.

Locked keepers from ARM_Linux_Lee, for the case where the sink has already flattened:
4. Heatsink temperature is not junction temperature, and it is not the test-set temperature. After the sink has settled, the channel, the board dielectric, and the generator/analyzer references can still be catching up. A coupler log that tracks the sink points at the PA; a log that tracks lab air or the instrument oven points at the rack. Watch Vds at the module pins, not only at the PSU display — cable IR drop moves bias when Idq walks.
5. Residual memory after thermal steady state: a slow EVM rise with a stable sink and stable Idq is often trapping, self-heating in the output match, or a DPD/EQ table trained cold (or at a different average power). More warm-up or tighter Idq compensation will not flatten that curve; you need a repeated identification at the operating temperature and duty, or a pulsed versus CW comparison at the same average dissipation. Stamp temperature, SCPI power and EVM from one clock — mixed wall clocks mis-attribute a fan-cycle sawtooth.

We will keep “soak at test power, then if sink and Idq are already flat, chase pin voltage, instrument references, and memory/DPD before tightening the sink further.”

— rfforum 编辑部 / Editorial Team
广告 AdRenesas 瑞萨微控制器芯片产品照片Renesas Electronics微控制器、模拟与射频/时钟方案 · MCUs, analog, RF and timing访问官网 →免费展示 · 资料来自公开信息 · 如需撤下请联系 ads@rfforum.com
广告 Ad广告位招商 / Advertise hereads@rfforum.com · 刊例 Rate cardT2 · 728×90
Log in or register to reply.