What Changes in a 5G RF Front End at 3.5 GHz vs mmWave?
What Changes in a 5G RF Front End at 3.5 GHz vs mmWave?
2026-10-27
Below 6 GHz a 5G front end is still a board problem. At 26 or 28 GHz it becomes a package problem. A 3.5 GHz 64T64R radio can be built as discrete PA, LNA and filter chains on a multi-layer laminate, one power amplifier per antenna path and a cavity filter behind it. At mmWave the antenna moves into the package, the filter follows it in, and the PCB keeps only power, control, heat spreading and mechanical registration, because the matching and the radiating element are already inside the module you bought.
What changes between the two bands
Item
3.5 GHz (n78)
26 / 28 GHz (n258, n257)
Wavelength in air, and a quarter wave
86 mm, 21 mm
11 mm, 2.7 mm
Carrier width in normal use
100 MHz
400 MHz
Path loss against 3.5 GHz
Reference
18 dB higher
PA per antenna path
5 to 10 W, GaN Doherty, 40 percent or better
1 to 3 W, 20 to 30 percent
LNA noise figure
0.5 to 1.0 dB
2 to 3 dB
Filter
Cavity or dielectric, mounted on the board
Inside the package with the antenna
Antenna
Separate array, fed by the board
Antenna in package, no board feed
Typical board
8 to 16 layers, mixed FR-4 and hydrocarbon
4 to 8 layers, thin PTFE or LCP, 0.1 mm to 0.25 mm
Line width tolerance
Plus or minus 25 to 38 micrometres
Plus or minus 25 micrometres, tighter on request
Assembly
0201 passives, QFN, connectorised ports
Antenna in package BGA or LGA at 0.4 mm pitch, flip chip, underfill
>
Read the table as a change in where the loss budget lives: at 3.5 GHz most of it sits in parts you can measure one at a time, at 28 GHz in transitions and in the package to board interface.
Why the loss budget moves from the board to the package
Free space loss scales with frequency, and 28 GHz costs about 18 dB more than 3.5 GHz over the same distance. The array has to earn that back with element gain, which is why a mmWave radio uses many small elements instead of a few large ones. On the board the arithmetic is worse. A 100 mm run of 50 ohm microstrip on standard FR-4 loses on the order of 0.5 to 1 dB at 3.5 GHz and several dB at 28 GHz, and most of that second figure is copper loss, so a lower loss laminate only removes part of it.
Geometry does the rest. A quarter wave at 28 GHz is 2.7 mm, so a 10 mm trace is already more than three quarter waves long: every discontinuity becomes a phase error, and ground vias spaced more than about 1 mm apart stop acting as a return path. The same trace at 3.5 GHz forgives almost all of it.
Why the PA and the LNA stop being symmetrical problems
At 3.5 GHz a GaN Doherty stage delivers 5 to 10 W per path at 40 percent efficiency or better at 8 dB backoff, so a 64 path radio can carry its heat into a conventional heatsink. At 28 GHz the same device family gives 1 to 3 W per element at 20 to 30 percent, and the array answers by using more elements. Heat per square centimetre goes up while the area to spread it goes down, which pushes mmWave carriers toward metal cores and vapour chambers.
On receive the pressure runs the other way. A 0.5 to 1.0 dB LNA at 3.5 GHz is routine, a 2 to 3 dB part at 28 GHz is normal, and 0.2 dB saved at the input is worth more than any gain added later. The board cannot add loss in front of the module either: a filter placed on the carrier ahead of an antenna in package receiver eats the whole noise budget.
Three things that go wrong when a 3.5 GHz build is moved up
Reusing the launch and the via structure. A connector that measures well at 3.5 GHz can be 1 dB lossy at 28 GHz with a resonance inside the band, and a plated through via in a 1.5 mm board leaves a stub that pulls the matching off at mmWave. Thin laminate, blind or buried vias, or back drilling fix it, and all three belong in the stack-up decision.
Tuning on the board what belongs in the package. Discrete matching added between an antenna in package module and the transceiver moves the reference plane away from where the module was characterised, and every extra millimetre adds loss and phase drift over temperature. If the module datasheet gives a 50 ohm interface, the board job is to arrive at 50 ohm and stop.
Treating the antenna in package module as an ordinary BGA. A 0.4 mm pitch package on a 0.1 mm laminate is a warpage problem before it is a soldering problem: the package expands at about 3 ppm per degree Celsius, the laminate at 14 to 17, and the joints take the difference across a 240 degree reflow. Paste volume per joint is small, the ground pad in the middle is large, and underfill is part of the mechanical joint rather than an optional extra. Builds that skip the underfill qualification pass first article and fail thermal cycling.
One more item is not electrical: at 28 GHz the antenna is inside the package, so acceptance is over the air and the fixture becomes part of the product definition.
How SUPERB supports both bands
SUPERB PCB supports RF and microwave boards with 50 ohm single ended impedance held to plus or minus 5 percent, and to plus or minus 2 percent on critical runs, with a TDR coupon on every panel rather than a sample. Mixed material stack-ups combine FR-4 with hydrocarbon or PTFE layers, layer counts run from 20 to 68, back drilling is available for stub removal, and panel size goes to 622 by 1092 mm. Capacity is 2.05 million square feet a month.
On the assembly side, SUPERB Assembly runs seven SMT lines in a 5,000 square metre ESD controlled area, from 01005 passives up to 57 mm components at 40,000 CPH, with twelve zone nitrogen reflow held below 500 ppm oxygen, 3D SPI before reflow and 3D AOI after, plus X-ray and CT for the ground pads under a PA or a module. Every build is keyed to the serial and to the panel, so a batch traces back to its paste lot and its reflow profile. Systems are certified to ISO 9001, with ISO 13485 for medical work.
See the RF module assembly range for PA, LNA and front end boards, and the RF and microwave PCB capability page for laminate and tolerance detail. Send the band plan, the stack-up and the module datasheet, and the reply will state what has to change before the first panel is made.