6G Research: Terahertz Communication and Next-Generation Mobile Networks

If 5G was about scaling millimeter wave, 6G is about taming terahertz. The 100 GHz to 1 THz frontier promises contiguous 50 GHz channels and terabit-per-second links, but it breaks nearly every assumption we use to design a modern smartphone radio. This is where physics and packaging collide.

Standards bodies have already drawn the line. ITU-R's IMT-2030 Framework (Recommendation M.2160) sets 6G targets at 1 Tbps peak rate, 0.1 ms air-interface latency, and 1e-9 reliability for industrial control, with integrated sensing and communication (ISAC) as a native capability. None of that is achievable below 100 GHz. The contiguous bandwidth simply does not exist. Our analysis of current mobile network interception and signaling architecture suggests even the control plane will need re-architecting to handle THz beam management at scale.

1. Why Terahertz, Why Now?

The usable spectrum below 100 GHz is fragmented and auctioned to exhaustion. Even with carrier aggregation across FR1 (410 MHz - 7.125 GHz) and FR2 (24.25 - 71 GHz), a 5G-Advanced UE tops out at ~1 GHz of aggregated bandwidth with complex RF front-ends. THz offers a different order of magnitude.

A single window at 140 GHz, 220 GHz, or 300 GHz contains atmospheric attenuation minima where propagation loss is manageable over 10-100 meters. At 140 GHz, absorption is ~2 dB/km, comparable to 28 GHz in rain. At the 183 GHz water vapor line, it spikes to >30 dB/km, creating natural isolation for dense indoor cells. This physics defines the use case: THz is not for wide-area mobility. It is for whisper-range, ultra-dense hotspots, device-to-device, and wireless backhaul that replaces fiber inside factories and data centers.

2. The THz Transceiver: A Block-Level Breakdown

A THz UE radio looks nothing like a Snapdragon X75 modem. The RF chain we tore down in the iPhone 16 Pro Max PCB teardown uses discrete mmWave IF modules with wafer-level fan-out. THz pushes that to heterogeneous integration: RFIC, antenna-in-package (AiP), and digital baseband must be co-designed as a single thermal and electromagnetic system.

The local oscillator is the bottleneck. Generating a clean 140 GHz carrier from a 23 GHz PLL via multipliers adds 20 log(N) phase noise — roughly 15.6 dB degradation for N=6. Recent ISSCC 2025-26 papers from TSMC and Samsung show 140 GHz VCOs with -88 dBc/Hz at 1 MHz offset in 28nm CMOS, but only just. Phase noise directly caps the usable QAM order; beyond 64-QAM, EVM collapses under THz phase noise even with digital compensation.

Silicon Choices for 140 GHz

• 40nm / 28nm CMOS (FD-SOI preferred): Cheapest, integrates baseband. fmax ~350 GHz. PAE <10%, output power <12 dBm per element.
• SiGe BiCMOS (90nm): Best balance today. fT 350 GHz / fmax 500 GHz. PA Pout 16-18 dBm, better noise. Used in Nokia/Keysight testbeds.
• InP HEMT: Highest fmax (>1 THz), Pout >20 dBm, but impossible to integrate with digital SoC. Likely for infrastructure only.

For a handset, the most credible path is heterogeneous: SiGe or InP front-end chiplets flip-chipped onto a CMOS beamformer with through-mold vias, similar to today's mmWave modules but with air-cavity packaging to control dielectric loss. We examined similar chiplet partitioning in our vapor chamber cooling analysis, where thermal interface material selection became the limiting factor — the same will happen at THz due to power density.

3. Propagation, Beamforming, and RIS

Free-space path loss at 140 GHz is punishing: 20 log(4πd/λ) yields ~102 dB at 10 meters before any atmospheric or blockage effects. A single isotropic radiator is useless. The system relies on extreme array gain.

A 256-element array (16x16) with λ/2 spacing at 2.14 mm wavelength fits in 17 x 17 mm — smaller than a current mmWave module. That array yields ~24 dBi, and with coherent combining at both ends, you claw back ~48 dB of link budget. But beamwidth drops to ~6 degrees. Mobility now becomes a beam tracking problem with sub-millisecond coherence times at walking speed.

This is where Reconfigurable Intelligent Surfaces (RIS) change the link. An RIS is a passive metasurface of 1024-4096 tunable unit cells that reflects and steers the beam around blockers — a hand, head, or torso that would otherwise impose 30 dB loss. Field trials by NTT DOCOMO and LG (2025) demonstrated a 14-18 dB link recovery in NLOS using a 400-element RIS at 140 GHz. Rather than increasing phone transmit power, the network deploys cheap, wall-mounted RIS panels that act as controllable mirrors.

4. Comparative Evolution: From 4G to 6G THz

5. Waveforms, Link Budget, and a Spec Sketch

OFDM, the workhorse of 4G/5G, suffers at THz: high PAPR (10-12 dB) forces the PA deep into backoff, wasting already scarce efficiency. Research is converging on DFT-spread-OFDM and single-carrier with frequency-domain equalization (SC-FDE) for uplink, and OTFS (Orthogonal Time Frequency Space) for doubly-selective high-Doppler scenarios.

The link budget at 140 GHz for a 10-meter indoor NLOS link reveals the razor-thin margin:

# 6G THz Indoor Link Budget Example — 140 GHz, 20 GHz BW, 64-QAM
TX_EIRP = 18 dBm (PA) + 24 dBi (256-el array) = 42 dBm
Path_Loss = 102 dB (FSPL @10m) + 8 dB (O2/H2O + wall) + 25 dB (hand blockage)
RIS_Gain = +16 dB (1024-el RIS, if deployed)
RX_Gain = 24 dBi (256-el array + LNA 6dB NF)
RX_Power = 42 - 102 - 8 - 25 + 16 + 24 = -53 dBm
Noise_Floor = -174 + 10*log10(20e9) + 8 (NF) = -63 dBm
SNR = 10 dB → 64-QAM requires ~18 dB → FAIL without RIS, MARGINAL with RIS
# Conclusion: THz requires LOS or RIS, ultra-wideband coding, and burst transmission.

This is why AI-native air interface proposals matter. Instead of sweeping 256 beams exhaustively (1.2 ms of overhead), a learned predictor using history and UE inertial sensors narrows the candidate set to 4-6 beams, cutting acquisition to ~80 µs. Samsung's AI-CE (AI-based channel estimation) demo at 140 GHz showed 2.1x spectral efficiency gain over LS estimation in NLOS.

6. Thermal and Power Delivery Constraints

The most underreported THz challenge is not spectrum physics but heat. Current premium phones sustain ~5 W continuous SoC + modem power before skin temperature hits 44°C. A THz TX burst of 4 W DC for 50 µs repeated every 1 ms is a 4% duty cycle — manageable. But sustained 1 Tbps streaming would require near 100% duty cycle, exceeding the envelope by 2x. The solution is architectural: THz is not for continuous browsing. It is for offloading — a 300 GB holographic asset downloaded in 2.4 seconds at 1 Tbps, then radio sleeps.

Power delivery also forces innovation. At 64 GS/s, the data converters alone draw >1 W. Optical or sub-THz interconnects between the modem and AiP will replace copper traces to avoid PCB loss >0.8 dB/mm at 140 GHz. Expect co-packaged optics and through-glass vias in 2029-era flagships.

7. Timeline: From Lab to Pocket

1. 2024-2026 (Now): Channel sounding, testbeds (Keysight, Rohde & Schwarz UXR 110 GHz scopes), IEEE 802.15.3d amendment for 100 Gbps THz.
2. 2027-2028: 3GPP Release 21 study items on 7-24 GHz (FR3) + THz channel models. First ETSI ISG THz documents. SiGe front-end sampling.
3. 2029-2030: ITU-R IMT-2030 specs finalized. WRC-27 identifies 102-109.5 GHz, 151.5-164 GHz, and 167-174.8 GHz for IMT. Early fixed-wireless THz CPEs.
4. 2031+: First integrated THz-Capable UE (likely foldable/AR glasses with larger AiP area). Mass adoption requires RIS infrastructure density — ~1 RIS per 2 small cells.

For SoC engineers, the takeaway is narrow: start designing for heterogeneous integration and AI-driven beam management now. The transistors can reach THz — TSMC N3 and Samsung SF2 with fmax >400 GHz are already there. The system — packaging, thermal, algorithms, and network planning — is what will determine whether 6G THz delivers a terabit or just a terahertz headache.