The THz Gap Just Met Its Match: Plasma Waves on Plain Silicon

Electrons on a silicon chip are usually thought of as traffic: a slow, orderly crawl of charge pushed along by voltage. But inside a transistor channel they can also move as a wave — a ripple of charge density that travels nearly ten times faster than the electrons themselves. That difference in speed is the difference between a technology stuck at a ceiling and one that just might smash through it.

The ceiling in question is the "THz gap." Terahertz frequencies — roughly 0.1 to 10 THz — sit between the microwaves of your cell tower and the infrared of a heat lamp. They promise vast bandwidth, high-resolution radar, spectroscopy, and data links measured in hundreds of gigabits per second. But nothing good, cheap, and silicon-based has ever worked there. Conventional transistors simply give up: at such frequencies, the time it takes a charge to cross the channel, and the parasitic capacitances that weigh the device down, push the maximum operating frequency, $f_{\text{max}}$, to a hard limit.

For years, engineers have watched that limit barely budge. Shrinking transistors didn't help — smaller dimensions mean lower breakdown voltages and less current-handling ability. The THz gap felt like a geological feature of physics: you could build exotic compound-semiconductor devices to reach it, but not with the cheap, mass-producible CMOS the entire electronics industry is built on.

A team from the University of Washington, the University of Michigan, and STMicroelectronics has now done something that sounds like science fiction but is grounded in careful math: they turned the transistor channel itself into a plasma medium, and used that plasma to amplify a 700 GHz signal — on a standard 28 nm FD-SOI CMOS chip.

The Science

The key insight isn't new, but it has been waiting for the right material and the right manufacturing node.

Plasma waves in transistor channels were first described decades ago by Michael Dyakonov and Michael Shur, who showed that the electron gas inside a field-effect transistor can support oscillations — collective ripples of charge — that travel much faster than the individual electrons. While electrons drift at roughly $10^7\text{cm/s}$, plasma waves propagate at around $10^8\text{cm/s}$, an order of magnitude faster (Otsuji & Shur, 2014). Critically, plasma wavelengths in semiconductors are more than 100 times smaller than electromagnetic wavelengths at the same frequency, which means you can build genuinely tiny THz devices on a chip.

The Naghavi team's contribution is to marry this physics to a practical, manufacturable structure they call a continuum transistor array (CTA): a row of 28 nm fully-depleted silicon-on-insulator (FD-SOI) CMOS transistors whose channels, taken together, behave as a single extended 2D plasma layer. Ten transistors in a row form a plasma waveguide roughly $1.31\mu$m long — and that "strip line" replaces the ordinary metal conductor of a conventional amplifier.

To understand how this works, they built a two-layer theoretical model. The first layer is the hydrodynamic transport model, a simplification of the notoriously expensive Boltzmann Transport Equation that focuses on two conservation laws: conservation of charge (current continuity) and conservation of momentum. The momentum equation, their Equation (1), tracks how the current density $\vec{J}$ evolves under electric fields, collisions, and thermal motion.

Linearizing that equation for small AC signals gives a time-harmonic model of the plasma medium at angular frequency $\omega$. The central result, their Equation (2), shows how energy flows from the DC electron stream into the AC signal. Plasma waves transfer energy from the DC terms — the drift velocity $v_{\text{beam}}$ and DC field $E_{z0}$ — into the AC terms $J_{sz1}$ and $E_{z1}$. In other words, the DC bias of the transistor feeds the amplification, exactly as the power supply of any amplifier does. One essential ingredient for growth: a low effective collision frequency ${\nu }_{m,\text{eff}}$, so the electron gas doesn't dissipate the wave before it can amplify.

The second layer is Pierce's theory, borrowed from the microwave-tube tradition. Slow-wave structures in vacuum tubes amplify by letting an electron beam interact with an electromagnetic wave; Pierce's formalism provides the extra equation linking the AC surface current to the AC electric field (their Equation 3). Combining the hydrodynamic model with Pierce's slow-wave relation gives a transcendental equation (Equation 4) — a third-order polynomial in the wavenumber $k_z$ that must be solved for forward- and backward-propagating waves. Solving it yields two physical modes, and a positive value of $\mathrm{Re}\{j{k}_z\}$ signals a wave that grows as it travels down the CTA.

There's a subtlety worth flagging. The hydrodynamic model assumes electrons follow a DC drift. But for the model to hold, the plasma wave velocity must dominate — and at zero DC current ($V_{DS}=0$), the plasma region can be treated with the Drude model instead, a simpler picture of free charges responding to fields with a characteristic collision time. This simplification lets the team extract the effective plasma parameters from full-wave electromagnetic simulations in HFSS, extracting $k_z$ from the simulated propagation and back-computing ${\sigma }_{0,\text{eff}}$, ${\nu }_{m,\text{eff}}$, and the effective gate distance $d_{\text{eff}}$ (Equations 5a and 5b).

What They Found

The simulation results establish that forward plasma waves can deliver gain at THz frequencies, while backward waves cannot. For a back-gated PWA, gain is achieved when the effective gate-to-channel distance $d_{\text{eff}}$ is below about 10 nm. The team mapped out gain as a function of electron mobility $\mu$ and drain-source voltage $V_{DS}$, showing that raising $V_{DS}$ increases the forward-mode gain.

Crucially, they found that achieving net gain requires $V_{DS}>2$ V — a value constrained by device breakdown limits. This is the honest tension at the heart of the work: the physics demands more voltage than the fragile thin channel can always supply. Their simulations suggest a path forward — thinning the buried oxide (BOX) layer to below 10 nm — as a way to relieve that constraint.

The proof-of-concept chip is the headline. The team designed and fabricated a back-gated plasma wave amplifier for 700 GHz on 28 nm FD-SOI CMOS, using ten transistors per unit cell biased to a uniform $V_{GS}$ and $V_{DS}$, with kilo-ohm resistors and quarter-wave bias lines keeping THz energy from leaking out through the power supply. Forty unit cells were cascaded, and the whole amplifier fits in a die area of $0.31{\text{mm}}^2$ including pads, consuming at most $1.29$ W of DC power.

Measured S-parameters show the amplifier achieves gain improvement along the plasma wave propagation path, though the peak shifted from the designed 700 GHz down to roughly 650 GHz — a mismatch the authors attribute to minor inaccuracies in the HFSS material modeling. The input return loss $S_{11}$ is also reported. It's a working THz plasmonic amplifier on standard CMOS — the first such demonstration on 28 nm FD-SOI, according to the authors.

The comparison table puts the achievement in sharp relief. Previous solid-state traveling-wave amplifiers used exotic compound semiconductors: n-GaAs, GaAs/AlGaAs, GaN/AlGaN. The GaN/AlGaN device (Naghavi et al., 2023) reached a gain of $21.8\text{dB/mm}$ at 75–110 GHz with a channel length of $8.3\mu$m. This work, by contrast, uses plain 28 nm FD-SOI CMOS, needs just $1.3\mu$m of channel (ten transistors), runs at a much higher frequency (650 GHz), and reports a gain per millimeter of $62\text{dB/mm}$ — far beyond anything in the prior art, even accounting for interstage finger-cap losses.

Plasma-wave amplifier gain density vs. frequency: CMOS leapfrogs compound semiconductors

Label	Value
n-GaAs (Lyubchenko 1994)	0.4
GaAs/AlGaAs (Pousi 2008)	0.8
GaN/AlGaN (2023)	21.8
This work (28nm CMOS)	62

Source: Table 1, Naghavi et al. (2026). This work's gain is 62 dB/mm including interstage finger-cap loss.

Electric field strength: tiny CMOS channels carry extreme fields

Label	Value
n-GaAs	120
GaAs/AlGaAs	15
GaN/AlGaN	964
This work (28nm CMOS)	7,600

Source: Table 1, Naghavi et al. (2026).

The striking point is the electric field: at $7,600$ V/mm (with $V_{DS}=1$ V across 131 nm), this design operates at a field strength an order of magnitude higher than the GaN device's $964$ V/mm — a consequence of the tiny dimensions that CMOS can now deliver.

Why This Changes Things

The THz gap has never been a physics impossibility; it's been a manufacturing economic problem. Compound semiconductors like InP and GaN can reach THz frequencies, but they're costly, hard to integrate, and don't play well with the silicon logic that surrounds them. CMOS is everywhere and cheap, but its transistors have been thought too slow at the top end.

This work reframes the problem. Instead of fighting the transistor's transit time — insisting the electron physically cross the channel fast enough — the plasma wave approach uses the channel as a medium for a much faster collective excitation. The electrons barely move; the wave does the traveling. That's why the plasma wave, at $10^8\text{cm/s}$, sidesteps the $10^7\text{cm/s}$ drift-velocity wall that limits conventional distributed amplifiers and traveling-wave transistors (Tarar et al., 2024).

The practical implications are large. A THz amplifier on standard CMOS would slot directly into the existing semiconductor ecosystem — no exotic materials, no special fabs. It could unlock on-chip terahertz for high-resolution radar imaging (the kind that sees through clothing and fog), spectroscopy (chemical fingerprinting without ionizing radiation), and the ultra-high-data-rate wireless links that 6G and beyond will demand. If a 700 GHz amplifier can be fabricated in a standard process, then the THz gap starts to look less like a gap and more like a frequency band that silicon never got around to claiming.

The numbers justify the optimism. A gain of $62\text{dB/mm}$ on CMOS at 650 GHz is not incremental; it's a different league from the $0.4$–$21.8\text{dB/mm}$ of the compound-semiconductor devices in the comparison table. If the frequency shift can be corrected and the $V_{DS}$ constraint relaxed by thinning the BOX layer, a THz amplifier chain becomes plausible.

There is also a satisfying intellectual elegance here. The team has unified two traditions that rarely meet: the century-old physics of vacuum-tube electron beams (Pierce's theory) and the modern language of CMOS device modeling (the hydrodynamic and Drude models). The microwave tube that powered early radar finds a grammatical descendant in a silicon chip smaller than a fingernail.

What's Next

The honest caveats matter as much as the headline. Net gain remains elusive at the moment — the measurements show gain improvement along the propagation path (i.e., the wave grows as it travels), but not yet a fully net-positive amplifier that overcomes all input losses. The $V_{DS}>2$ V requirement collides with device breakdown limits, and the team explicitly notes that realizing net gain "may necessitate thinning the BOX to less than 10 nm."

The frequency shift from 700 to 650 GHz is a reminder that the material modeling still has inaccuracies — the HFSS simulations don't perfectly capture the real physics of the channel. And the boosted gain figure carries a footnote: it includes the interstage finger-cap loss, so the raw channel gain is even more impressive, but the system-level accounting is still being refined.

What's promising is that the levers to pull are known and physical. The gate distance $d_{\text{eff}}$ can be tuned; the mobility $\mu$ and bias $V_{DS}$ can be adjusted; the BOX thickness is a fabrication choice. This is not a "maybe some day" paper; it's a demonstration that the mechanism works on standard silicon, with a clear roadmap for the engineering that remains.

The next steps, in the authors' framing, are to thin the buried oxide, correct the material models, and push toward net gain at (or above) 700 GHz. Beyond that lies the question of whether the plasma-wave approach can scale to higher frequencies still, and whether it can be co-integrated with the digital logic that would make a full THz transceiver on a single chip — a "terahertz system-on-chip."

Channel length: CMOS shrinks the active medium by three orders of magnitude

Label	Value
n-GaAs	2,000
GaAs/AlGaAs	2,600
GaN/AlGaN	8.3
This work (28nm CMOS)	1.3

Source: Table 1, Naghavi et al. (2026).

There is something quietly revolutionary about the direction this points. For decades, the answer to "how do we go faster" in electronics has been "make the transistor smaller." This work asks a different question: what if the electrons don't cross the channel at all, but the wave does? The transistor, in this view, is not a switch but a medium — and the fastest thing in silicon is not the electron, but the ripples it makes.

The THz gap may not be a wall. It may be a wave waiting for a surfboard.