mmWave Antenna Module: 5G Antenna-in-Package Design for Smartphones

Flagship teardown after teardown tells the same story: three black rectangular modules, each about 22 x 6 x 1.2 mm, tucked along the metal mid-frame where no conventional antenna could work. If you pry one open, you won’t find a flex antenna. You’ll find a multilayer substrate, a flip-chipped RFIC, and a 4-element patch array staring through the plastic sidewall. That is AiP, and without it mmWave in smartphones would not exist.

1. Why Discrete Antennas Fail Above 24 GHz

Sub-6 GHz design treats the antenna and transceiver as separate domains connected by a 50-ohm transmission line. That abstraction collapses in FR2 (24.25 – 52.6 GHz). At 28 GHz, free-space wavelength is 10.7 mm. A conventional FR-4 trace exhibits 0.8–1.2 dB/mm loss, and any 90-degree bend or connector introduces phase dispersion that destroys beamforming accuracy.

Even more critical is the link budget. mmWave path loss at 100m is ~120 dB versus ~95 dB at 3.5 GHz. To close that 25 dB gap, the phone must supply 20–23 dBm EIRP from a tiny aperture. A single patch element provides 5–6 dBi. Only a phased array with coherent combining can deliver 12–15 dBi array gain plus 8–10 dB from the beamformer’s power amplifiers. Coherence demands feed length matching within ±30 μm. You cannot hold that tolerance across a motherboard, a board-to-board connector, and a flex cable.

This is why every shipping mmWave phone from the Snapdragon X75 onward — and the new Exynos configurations we analyzed in our Samsung Exynos 2500 GAAFET breakdown — pairs the modem with dedicated AiP modules rather than routing mmWave across the main PCB.

2. Anatomy of a Smartphone AiP Module

A modern AiP is a 4- to 6-layer heterogeneous stack. Understanding the stackup is key to understanding its tradeoffs.

Layer 1 – Antenna: Four microstrip patches etched on liquid crystal polymer (LCP) or low-temperature co-fired ceramic (LTCC). LCP (Dk 3.1, Df 0.002) dominates 2025+ designs because it survives reflow and is thinner than glass. Patches are probe-fed through vias from the layer below.

Layer 2 – Feed and shielding: A stripline distribution network with a slotted ground plane forming a cavity. This isolation is non-negotiable; without it, RFIC switching noise couples directly into the antenna pattern, creating sidelobes that violate FCC spurious limits.

Layer 3 – RFIC: A single die (often 28 nm RF-SOI or 45 nm RF-CMOS) flip-chipped via Cu pillars. Integration is aggressive: four transmit PAs (Psat ~14 dBm each), four LNAs (NF ~3.5 dB), 5-bit phase shifters, and calibration couplers. The die dissipates heat downward into the motherboard through thermal balls — upward heat would detune the patches.

The entire stack is overmolded except for the antenna aperture, which requires a keep-out window in the metal frame. Designers learned this the hard way in 2021: a full metal sidewall at 28 GHz creates a Faraday cage with 30 dB blockage.

3. Phased Array and Beamforming: From Theory to Code

Each element in the array is fed with a digitally controlled phase. By applying a progressive phase gradient Δφ across elements spaced d apart, the main beam steers to θ where Δφ = (2πd/λ) sinθ. With 5-bit shifters, the phase resolution is 11.25°, giving ~±50° scan with <1 dB gain ripple.

Unlike massive MIMO base stations, handset arrays use analog beamforming — a single data stream weighted in RF. This keeps power tractable but limits the phone to one spatial beam at a time. The baseband tells the RFIC which beam index to load via MIPI RFFE, and the array switches in < 200 ns.

What makes this relevant to app developers is latency. Beam failure recovery in 5G NR takes 20–40 ms if the phone must sweep. That jitter directly impacts real-time workloads, which is why modem scheduling and mobile app runtime comparison testing now includes RF state as a variable. Similarly, the AI accelerators that handle real-time AI image processing increasingly assist in predictive beam selection using grip sensors.

4. Materials Make or Break the Module

Dielectric loss dominates at mmWave. The figure of merit is Df (loss tangent). FR-4 Df ~0.02 is unusable — it turns 30% of your power into heat before it reaches the patch. LTCC (Df 0.0015) was the first enabler but requires thicker stacks. LCP and modified polyimide (MPI) now dominate because they achieve Df 0.002–0.004 in flexible, thin laminates and integrate with standard SMT flows.

Glass substrates (Dk 5.4–6.6) are the emerging alternative for 39 GHz and 60 GHz. Their dimensional stability (±10 ppm/°C) improves phase accuracy across temperature, but brittleness complicates drop testing. Most vendors hedge: LCP for 28 GHz global models, glass interposers for 39 GHz US variants.

The solder interface matters just as much. Conventional SAC305 solder balls create a parasitic inductance of ~0.15 nH — enough to detune a 28 GHz match. AiP modules use Cu-pillar flip-chip with 40 μm height and underfill with εr-matched epoxy to keep the bump transition under 0.2 dB loss.

5. Coexistence, Thermal and Mechanical Realities

An AiP lives 3 mm from the application processor and 2 mm from the battery. Three problems dominate integration:

Thermal desense

At 1.5W Tx, the RFIC junction hits 95–105 °C without spreading. That heat drifts the patch resonant frequency by ~ -45 ppm/°C (about -1.3 MHz/°C at 28 GHz). Designers add a 0.3 mm graphite spreader beneath the module and throttle duty cycle to 25% sustained. The phone’s thermal daemon, which also throttles the CPU, must arbitrate between SoC and modem — a reason chipset-level power domains like those in the Exynos 2500 analysis tie directly to RF performance.

Blockage and hand grip

Human skin at 28 GHz has ~18 dB penetration loss. A palm covering a module kills the link. Hence spherical coverage with three modules, plus a software layer that selects the least-blocked array using proximity and capacitive grip sensors. If all three are blocked, the modem falls back to LTE/NR sub-6 within 10 ms.

Regulatory and coexistence

Spurious harmonics at 28 GHz land in K-band satellite downlinks and can desense the phone’s own 5 GHz Wi-Fi front-end. Each AiP includes a LTCC BAW filter on the IF path (typically 9–10 GHz IF to the transceiver) and a dedicated shield can with 40 dB isolation. Antenna pattern must pass FCC MPE limits; peak EIRP is capped and beam dwell time is averaged over 4 seconds.

Data path implications extend to storage. Factory calibration generates 150–200 KB of per-module gain/phase tables that must be loaded at boot with deterministic latency, a pipeline where low-latency storage stacks like F2FS and APFS actually influence modem boot time.

6. Manufacturing Test and Calibration

AiP modules are tested as sealed units on an over-the-air (OTA) chamber, not via probe. The ATE drives the module through a quiet RF link while a horn antenna measures the 64-beam codebook. Phase error must be < ±7° and amplitude error < ±0.7 dB across temperature, or the beam sidelobes violate 3GPP TR 38.817. Bad modules are binned — there is no rework on an embedded array.

For phone OEMs, integration test adds a radiated two-stage (RTS) step: the assembled phone is placed in a compact antenna test range (CATR) that emulates far-field at 0.6m distance. Test time is the hidden cost: 35–45 seconds per device for mmWave alone, versus 8 seconds for sub-6. That time is multiplied by three modules.

7. Where AiP Goes Next

Two trajectories are clear. First, dual-band AiP (28/39 GHz) sharing the same aperture via stacked patches is entering qualification, cutting the module count from three to two and reclaiming 0.9 cm³ of volume. Second, fan-out wafer-level packaging (FOWLP) is replacing laminate substrates, enabling 0.8 mm total thickness and embedding the transceiver die alongside the beamformer — a true single-chip mmWave front end.

Beyond 2027, the conversation shifts to FR3 (7–24 GHz) and eventually sub-THz. Neither will be AiP as we know it; element counts will jump from 4 to 16–64, requiring hybrid beamforming. But the design discipline AiP forced — treating the package as the antenna — will remain foundational. Smartphones proved that at mmWave, you don’t connect to an antenna. You live inside one.