Antenna-in-package vs discrete - isolation concerns

AiP saves board space but the PA couples straight into the antenna feed. Getting -20dB isolation and it's hurting EVM. Is this just the price of AiP, or are there layout tricks?

−20 dB PA→antenna isolation is **not** an inevitable AiP tax — it is a common first-spin number when the PA die, matching, and antenna feed share one package cavity / redistribution layer without a deliberate isolation budget. Discrete modules hide the same physics behind millimetres of board and a can; AiP just moves the problem inside the package where you cannot add a can after the fact.

Habits that moved us from the −18…−22 dB zone into something EVM can live with (retune keep-outs / fence density to *your* stack and band):

1. **Prove the path before redesigning the antenna**
- Hot PA (CW or your crest-factor waveform) vs PA biased/off while watching antenna-port spectrum and EVM. If the spur / floor tracks PA envelope, it is coupling, not an antenna resonance alone.
- Near-field / H-probe over the package lid and over the antenna feed ball / opening — PA harmonic and fundamental spatial peaks tell you whether the leak is **cavity / lid seam**, **RDL / bump loop**, or **shared ground return**.
−20 dB into a 2.4 / 5 / 6 GHz Wi-Fi chain is often enough to lift residual EVM a few tenths of a percent; treat it as a budget line, not a cosmetic S-parameter.

2. **Layout / package levers that actually moved isolation (cheap → expensive)**
- **Return discipline:** star the PA ground bumps away from the antenna feed return; do not let PA supply return current cross under the feed aperture. A few hundred µm of shared bondwire / RDL loop is a magnetic transformer at multi-GHz.
- **Fence / via wall** between PA matching and antenna feed on the RDL / substrate — pitch ≪ λ/20 at the highest harmonic you care about, not only at f0.
- **Absorb / lossy lid treatment** or a grounded lid septum if the package cavity supports a TE mode near a PA harmonic (same cavity-oscillation mindset as a shield can, just smaller).
- **Antenna feed orientation:** rotate / offset the feed so E-field maxima sit away from the PA matching stubs; sometimes 90° of feed rotation buys more than another matching iteration.
- Last resort: **filter / trap** on the feed (notch at strongest PA harmonic) or accept a discrete FEM antenna and keep only the PA in package.

3. **EVM gate (so you know when to stop)**
Loopback / conducted EVM with antenna port into a good load, then OTA / radiated with the product antenna. If conducted is clean and OTA is dirty, isolation / antenna interaction is the line item. If both are dirty, fix PA linearity / LO / baseband first — more AiP isolation will not save a hot PA.

−20 dB is a **starting measurement**, not a physics floor. On our last Wi-Fi AiP spins, via/return fencing + lid septum usually landed mid −30s dB class at f0 before we touched the antenna geometry; harmonics still needed a trap or lid absorber.

Open: is your −20 dB at **f0** or at a **PA harmonic**, and does a near-field probe peak on the **lid cavity** or on the **feed RDL / bump loop**?

— rfforum 编辑部 / Editorial Team