The RF Challenge at Millimeter Wave Frequencies
Building chips for 5G's mmWave bands (24-71 GHz) is a completely different game than sub-6 GHz design. At these frequencies, signal propagation gets brutal — free-space path loss increases with the square of frequency, meaning a 28 GHz signal loses about 20 dB more than a 3.5 GHz one over the same distance. That's not a rounding error. It's the difference between a working system and a dead one.
The RF front-end module (RFEM) for mmWave has to compensate for those losses, and that means cramming a lot of gain into a very small package while keeping noise figures reasonable. Qualcomm's QTM545 module, for instance, packs antenna elements, phase shifters, power amplifiers, and low-noise amplifiers into a single package that's about the size of a fingernail.
Beamforming: Steering Signals Without Moving Parts
At mmWave frequencies, the antenna elements are small enough — roughly half a wavelength, so about 5mm at 28 GHz — that you can fit 16, 32, or even 64 elements into a practical form factor. This is where phased array beamforming comes in.
Each antenna element gets its own phase shifter, and by adjusting the relative phase between elements, you steer the beam electronically. No mechanical parts. The beam can be redirected in microseconds, which is critical for tracking mobile users or switching between multiple devices.
There are two main approaches to the beamforming architecture:
- Analog beamforming — phase shifting happens in the RF domain before the signal gets digitized. Cheaper, lower power, but you can only form one beam at a time per panel.
- Hybrid beamforming — splits the work between analog phase shifters and digital baseband processing. You get multiple simultaneous beams at the cost of more ADCs/DACs and higher power consumption.
- Full digital beamforming — every antenna element has its own complete RF chain and ADC. Maximum flexibility, maximum power draw. Mostly used in base stations where power budget isn't as tight.
I'd argue that hybrid beamforming has won the practical tradeoff for handset and small-cell designs. Pure analog is too limiting for multi-user MIMO, and full digital is too power-hungry for battery-operated devices.
We covered a related topic in Semiconductor IP Licensing: ARM, Synopsys, and the IP Block .
Power Amplifier Design: The Efficiency Problem
The power amplifier (PA) is where most of the RF front-end's power goes. At mmWave frequencies, PA efficiency drops significantly compared to sub-6 GHz designs. A typical sub-6 GHz PA might achieve 40-45% power-added efficiency (PAE). At 28 GHz, you're looking at 25-30% if things go well. At 39 GHz, it's even worse.
Why the drop? Several factors compound:
- Transistor gain decreases as frequency approaches fmax
- Matching network losses increase
- On-chip interconnect losses become significant
- Back-off requirements for high PAPR signals (5G NR uses OFDM with PAPR of 8-12 dB) force the PA to operate well below its compression point
The process technology matters enormously here. GaAs and SiGe BiCMOS have traditionally dominated mmWave PAs because their transistors have higher fmax and better breakdown voltages than bulk CMOS. But there's been a steady push toward CMOS integration — if you can build the PA in the same CMOS process as the digital baseband, you eliminate expensive multi-chip packaging.
TSMC's 16nm FinFET and Samsung's 14nm processes have demonstrated mmWave PAs with respectable performance at 28 GHz. The trick is using stacked transistor topologies and transformer-based power combining to work around the lower breakdown voltage of CMOS devices. Qualcomm's been shipping CMOS-based mmWave front-ends since the X55 modem generation.
Antenna-in-Package: Where the Chip Meets the Air
One thing that makes mmWave design fundamentally different from lower-frequency RF is that the antenna becomes part of the chip package. At 28 GHz, a patch antenna element is roughly 5mm × 5mm. You can't route these signals through board-level traces and connectors without unacceptable loss — the signal would be half gone before it reached an external antenna.
For a related perspective, see ARM Architecture Evolution: From ARMv8 to ARMv9 and Custom C.
The solution is antenna-in-package (AiP), where the antenna array is integrated directly into the package substrate or built on top of it. The RFIC die sits underneath, connected to the antennas through short vertical transitions. This keeps the RF path loss manageable and allows the entire module to be treated as a single drop-in component.
Package material selection is critical. Standard organic substrates have loss tangents around 0.005-0.01 at mmWave frequencies. Low-loss materials like liquid crystal polymer (LCP) or modified polyimide bring that down to 0.002-0.004, which translates directly into better antenna efficiency and wider bandwidth.
The Thermal Challenge
A 5G mmWave module in a smartphone might dissipate 2-4W when transmitting at full power. That's concentrated in a package maybe 15mm × 10mm. The thermal density rivals some laptop processors.
Phone designers typically place mmWave modules at the edges of the device, behind the display or near the frame. Thermal spreading layers — often graphite sheets or vapor chambers — help distribute the heat. But there's a hard physical limit: skin temperature can't exceed about 45°C for comfortable handling, and the module junction temperature needs to stay below 125°C for reliability.
This is why mmWave throughput in phones tends to throttle after sustained use. The chips are capable of more, but the thermal envelope won't allow it continuously.
For a related perspective, see Automotive Chip Requirements: ISO 26262, Temperature Ranges,.
Testing and Calibration at Scale
Testing mmWave devices is expensive. You can't just probe the RF ports with a network analyzer the way you would a sub-6 GHz chip. At 28+ GHz, probe-based testing becomes impractical for high-volume production because the parasitics at the probe interface are comparable to the DUT's own characteristics.
The industry has moved toward over-the-air (OTA) testing, where the packaged module radiates into a measurement antenna inside a shielded chamber. This tests the entire signal chain — IC, package, and antenna — as a system. It's more representative of real-world performance, but the test fixtures are expensive and measurement time is longer.
Each module also needs per-unit calibration. Manufacturing variations in the antenna elements and phase shifters mean that the beam patterns won't match the design without individual tuning. This calibration data gets stored in non-volatile memory on the module and loaded during operation. It adds production time and cost, but there's no way around it if you want consistent beam quality.
What's Coming Next
FR3 bands between 7-24 GHz are getting attention as a middle ground — better propagation than mmWave with more bandwidth than sub-6 GHz. Chip designers are already working on front-ends that can cover both sub-7 GHz and FR3 with a single transceiver, which would simplify phone designs significantly.
On the silicon side, the push toward 3nm and below processes will help digital beamforming become more practical for mobile devices. Lower supply voltages mean lower power per ADC/DAC, which is the main bottleneck for full digital architectures. Whether that's enough to displace hybrid beamforming in handsets remains to be seen, but it's heading that direction.