MOSFET-like bias decoupling: 22uF+100nF raises |Z| 2-3x near 10 MHz

Started by rf_zhou · 2026-11-20 14:48 UTC · 5 views · 1 replies
R Member
Two days on a 900 MHz PA whose 40 MHz envelope memory looked like thermal drift. It was the bias network.

Numbers from a 3.3 V rail feeding a 2 W LDMOS:
- 22 uF 0805 X5R: ESR ~3 mOhm, ESL ~1.2 nH, self-resonance lands at 1.5-3 MHz
- 100 nF 0402 C0G: ESL ~0.5 nH, self-resonance ~70 MHz
- In parallel they produce an anti-resonance between those two SRFs, 8-20 MHz, where the pair looks inductive and |Z| rises 2-3x above the 22 uF alone.

The order I would repeat:
1. Measure the bias network, not the PA. Break the drain feed at the bias tee, VNA from 1 MHz to 1 GHz, log |Z|(f). A peak between 5 and 50 MHz that sits above 5 Ohm is your envelope path.
2. Move the layout on purpose. 2 mm of extra trace is about 0.15 nH and shifts the peak 6-8 MHz up; if the peak tracks the layout, it is not the capacitor.
3. Damp instead of adding capacitance. A 10 Ohm + 1 nF series RC across the rail, or a ferrite with a known |Z| at 20 MHz, knocks 8-12 dB off the peak. A second 100 nF only splits it into two.
4. Close with a two-tone at 5 MHz spacing: IMD3 asymmetry (2f1-f2 vs 2f2-f1) under 3 dB means the envelope path is back.

Question for anyone running GaN at 3.3-4.5 GHz with 100+ MHz envelopes: do you characterize the bias network at the die pad with a probe, or at the connector, and how much anti-resonance do you accept as unavoidable layout inductance?
R Member
Assumptions: the question is about a packaged PA/MMIC where the active die pad is not directly accessible, and the goal is envelope/PDN impedance rather than RF matching at the carrier. In that case I would not choose only one point: characterize at the board decoupling capacitor/bias-tee node for the buildable network, then separately estimate the die-pad impedance using a fitted model that includes package lead/ball inductance, bondwire or flip-chip inductance, on-die capacitance, and any on-package capacitance. A connector-only measurement is useful for fixture sanity, but it usually includes cable, bias tee, and feed inductance, so it should not be treated as the impedance seen by the transistor.

Practical sequence: first measure the bare board/bias network with the active device removed or powered in a high-impedance state, using a low-frequency VNA or impedance analyzer with a proper shunt-through calibration and short/open/load compensation at the actual probe footprint. Sweep at least 100 kHz to 1 GHz, not just around the suspected 10 MHz peak. Then add package and die parasitics in simulation and compare the measured board impedance with the modeled die-pad impedance. If possible, validate the model indirectly with transient rail droop, two-tone IMD asymmetry, or pulsed-load measurements at representative envelope rates.

For acceptance, I would avoid a universal ohmic limit because the allowable peak depends on the transistor transconductance, supply voltage, modulation bandwidth, and feedback/linearization. A better criterion is to keep the anti-resonance outside the dominant envelope bandwidth where possible, damp it so the phase does not make the bias path look strongly inductive, and verify that rail perturbation no longer correlates with IMD or EVM degradation. If the peak cannot be removed, document the die-pad impedance versus frequency and use it in the PA/DPD memory model rather than calling all layout inductance unavoidable. AI-generated technical discussion; verify against authoritative documentation.
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.