When the wavelength fits on the chip
At the frequencies of the rest of this track — say a 2.4 GHz Wi-Fi radio — the radio wave is 12.5 cm long. That is enormous compared to your chip: a 2 mm die is a thousandth of a wavelength across, so the whole circuit behaves like a single point and you can reason with lumped R, L and C. Now crank the carrier to 28 GHz, a flagship 5G band, and the wavelength collapses to about 1.07 cm in air — and to perhaps 3–4 mm inside silicon dioxide where the wave moves slower. Suddenly a metal trace running across your die is no longer a wire. It is a transmission line with its own delay, and the voltage at one end is genuinely different from the voltage at the other.
This is what defines millimeter-wave design: the millimeter-scale wavelength becomes comparable to your wires, your transistors, even your bond pads. Below ~20 GHz that is a nuisance you correct for. Above it, it becomes your toolbox. You stop spending precious area on big spiral inductors and start drawing short, controlled-impedance strips of metal — a few hundred microns long — that act as inductors, matching networks, and power splitters all at once. Geometry becomes the circuit.
Wavelength as the carrier climbs (lambda = c / f) freq lambda (air) lambda in SiO2 (~e_r 4) on a 2 mm die ------- ------------ ---------------------- ------------- 2.4 GHz 125 mm ~62 mm negligible 28 GHz 10.7 mm ~5.4 mm ~0.4 lambda 60 GHz 5.0 mm ~2.5 mm ~0.8 lambda 77 GHz 3.9 mm ~1.9 mm ~1.0 lambda (auto radar) 140 GHz 2.1 mm ~1.1 mm ~1.9 lambda (6G research) Rule of thumb: once a trace is longer than ~lambda/20, treat it as a transmission line, not a wire.
The cruel trade: passives win, transistors lose
Nature never gives a free lunch. The same physics that shrinks your passives also strangles your active devices. Every transistor has a maximum frequency, f_max, beyond which it provides no power gain at all — it is just a lossy resistor. As you approach a sizable fraction of f_max, three things degrade together, and you feel all of them in the same design.
- Gain evaporates. A transistor's gain falls roughly 6 dB for every doubling of frequency. The 20 dB stage you took for granted at 5 GHz might give you 8 dB at 28 GHz and barely 4 dB at 77 GHz — so you need more stages, and each one adds its own loss and noise.
- Noise figure climbs. The first LNA now contends with lossy on-chip routing in front of it, and the transistor's own noise rises. A 0.5 dB noise figure at 2 GHz becomes 3–5 dB at 28 GHz; that loss directly costs you receiver range.
- PA efficiency collapses. A power amplifier that hit 50% efficiency at 2 GHz often manages only 15–25% at 28 GHz. Less RF out, more heat in — and at mm-wave that heat is packed into a tiny die.
- Metal and substrate loss dominate. At 60 GHz the skin effect crowds current into a sub-micron layer at the conductor surface, and the lossy silicon substrate underneath siphons energy away. Interconnect that was free now eats decibels.
Why path loss forces a new architecture
Here is the headache that makes mm-wave sound impossible at first. The energy a single antenna captures scales with its physical area in wavelengths squared — so for a fixed-size antenna, the higher the frequency, the smaller it is electrically and the less it grabs. The free-space path loss between two such antennas grows as the square of frequency. Going from 3 GHz to 30 GHz, all else equal, you lose 20 dB — a factor of one hundred in power — before the signal even reaches the air's molecular absorption peaks (oxygen near 60 GHz, water near 24 and 120 GHz).
Free-space path loss, FSPL(dB) = 20 log10(d) + 20 log10(f) + 32.45
(d in km, f in MHz)
At d = 100 m:
3 GHz -> 82 dB
28 GHz -> 101 dB <- +19 dB just from the frequency term
60 GHz -> 108 dB (+ ~15 dB/km oxygen absorption on top)
But the cure hides in the same physics:
a small wavelength means MANY antennas fit in a small area.
half-wave spacing at 28 GHz = ~5.4 mm
a 16 x 16 element array = ~9 x 9 cm -> 256 antennas!
Array gain ~= 10 log10(N) for the receive SNR boost,
and up to 20 log10(N) for the two-way EIRP of a TX array.
256 elements -> +24 dB receive, up to +48 dB EIRPThe escape is beautiful precisely because it uses the curse against itself. Because the wavelength is tiny, you can fit a forest of antennas — spaced a half-wavelength apart — into the area of a coin. Feed each one a copy of the signal with a carefully chosen phase delay, and the wavefronts add up coherently in one chosen direction and cancel everywhere else. That is a phased array, and it conjures a high-gain, pencil-thin beam you can point electronically, with no moving parts, in microseconds.
Inside a phased-array transceiver
Now everything you learned earlier in this track returns — but replicated per element. Each of the N antennas gets its own front-end: an LNA for receive, a PA for transmit, a shared mixer for up/down-conversion, and a phase shifter that sets this element's contribution to the beam. A single frequency synthesizer and VCO usually feeds all of them, so they stay phase-coherent. The art is making one tiny, low-power, well-matched transceiver slice and then tiling it — because you are about to build dozens or hundreds.
One element slice (x N), beamforming in the RF/phase domain:
antenna_0 --[T/R sw]--+--[LNA]--+
| |
(TX) <--[PA]<-----+ [phase shift psi_0]--+
|
antenna_1 --[T/R sw]-----[LNA]---[phase psi_1]---+--> sum
antenna_2 --[T/R sw]-----[LNA]---[phase psi_2]---+ |
... v
antenna_N-1 -----------[LNA]----[phase psi_N-1]--+ [mixer]--IF/ADC
^
shared LO <--[ PLL / VCO ]
Steering law: psi_k = -k * (2*pi/lambda) * d * sin(theta)
d = element spacing (~lambda/2), theta = desired beam angle
Set the N phase words -> the main lobe points at theta.
Change the words in microseconds -> the beam jumps, no motors.There is a spectrum of beamforming choices, and it is a core architectural decision. RF (analog) beamforming shifts phase at the carrier and sums before a single mixer and data converter — cheapest and lowest power, but one beam at a time. Digital beamforming gives every element its own converter and forms beams in DSP — many simultaneous beams, full flexibility, but the power and area of N data converters running at GHz rates. Most real 5G and radar chips land on hybrid beamforming: analog sub-arrays of, say, 4–16 elements, combined digitally, to get a few independent beams without paying for N full receivers.
Co-design: the chip, the package, and the antenna are one
At mm-wave you cannot finish a chip and then 'attach an antenna' the way you bolt a connector onto a Wi-Fi board. A 28 GHz signal traveling 1 cm down a bond wire or a cheap PCB trace can lose several dB and pick up reflections you cannot undo. The wavelength is so short that the package itself becomes part of the radio. The frontier answer is Antenna-in-Package (AiP): the antenna array is etched directly into the layers of the package substrate, sitting micrometers above the silicon, with flip-chip bumps carrying the RF up from the die.
This collapses three disciplines into one optimization loop. The transistor person tuning the PA output impedance, the passive person drawing the matching strip, and the antenna person shaping the patch and its feed must all solve the same boundary-value problem together — because the PA's load is the antenna, and the antenna's source is the PA, with the package in between. A modern mm-wave AiP module is verified with full 3-D electromagnetic simulation of die, bumps, substrate and radiating elements as a single object.
- Budget the link first. From the target range, data rate and modulation, work out the required EIRP and receiver sensitivity — your link budget tells you how many elements you actually need.
- Design one element slice — LNA, PA, mixer, phase shifter — small, efficient and impedance-matched to the antenna it will feed, not to a generic 50 Ω.
- Tile and route on-chip with controlled-impedance transmission lines so all paths stay phase-equal; mismatched lengths show up as a tilted, blurred beam.
- Co-simulate die + package + antenna in 3-D EM, then iterate — the antenna pattern, the PA match and the substrate loss are not separable problems at mm-wave.
Where the frontier is heading
Three application worlds are pulling mm-wave silicon forward. 5G/6G mobile wants gigabit links and is climbing from 28/39 GHz toward the FR2 extensions and 6G research bands above 100 GHz, where the spectrum is wide and empty — and the path loss is even more savage, demanding still-larger arrays. Automotive radar at 76–81 GHz now uses tens of integrated transceivers and clever FMCW waveforms to measure range and velocity, with matched-filter-style processing pulling a weak echo out of the noise to spot a pedestrian. High-data-rate fixed links — backhaul, data-center wireless, even chip-to-chip — exploit the enormous bandwidth that only opens up this high.
What ties this rung back to everything before it is integration. The reason a 256-element 28 GHz array fits in a phone or a streetlight-sized base station is that one CMOS or SiGe die now holds the LNAs, PAs, mixers, synthesizers, phase shifters and digital control — replicated, calibrated and co-packaged with the antennas. Every block you mastered earlier in this track is here, just multiplied by N and squeezed to its physical limit. That is the frontier: not a single brilliant circuit, but an army of competent ones, steered as one.