🔬 optics

Dispersion-Engineered Terahertz Silicon Interconnects Enabling Terabit-Scale Data Links

This paper demonstrates a CMOS-compatible, dispersion-engineered silicon waveguide platform operating from 220 to 500 GHz that achieves terabit-scale data transmission (up to 1.004 Tbps) with low loss and group-velocity dispersion, offering a scalable solution for high-bandwidth on-chip and chip-to-chip interconnects in next-generation AI and 6G systems.

Original authors: Bodhan Chakraborty, Wenhao Wang, Nikhil Navaratna, Thomas Caiwei Tan, Pascal Szriftgiser, Hadjer Nihel Khelil, Guillaume Ducournau, Ranjan Singh

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Bodhan Chakraborty, Wenhao Wang, Nikhil Navaratna, Thomas Caiwei Tan, Pascal Szriftgiser, Hadjer Nihel Khelil, Guillaume Ducournau, Ranjan Singh

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your computer's brain (the processor) and its memory are like two busy cities that need to exchange massive amounts of information every second. Right now, the roads connecting these cities are getting clogged. The old "roads" (electrical wires) are too slow and get too hot, while the "highways" (fiber optics) are powerful but too complex and energy-hungry to build right inside the tiny chip itself.

This paper introduces a new kind of "super-highway" built directly onto the silicon chip, using invisible waves called Terahertz (THz) waves. Think of these waves as a magical bridge that sits right between the slow electrical wires and the complex fiber optics, offering the best of both worlds.

Here is how the researchers built this highway and why it's a big deal:

1. The Problem: The "Traffic Jam" of Light

Usually, when you try to send light (or THz waves) through a silicon chip, the waves hit tiny bumps and holes in the material. Imagine trying to drive a car down a road where the pavement has a repeating pattern of potholes. At certain speeds, the car gets stuck or bounces back. In physics, this is called a Bragg stopband. It creates a "no-go zone" where data can't pass, limiting how much information you can send.

2. The Solution: A Smooth, Custom-Paved Road

The team designed a special silicon waveguide (a pipe for light) that acts like a dispersion-engineered, unclad highway.

"Unclad": Imagine a road that doesn't have guardrails or fences on the sides, but instead relies on the natural slope of the road to keep cars (the waves) in the center. This keeps the road simple and easy to build.
"Dispersion-Engineered": This is the magic trick. The researchers changed the shape of the "potholes" (the holes in the silicon) from perfect circles to semi-circles. It's like repaving the road with a specific texture that cancels out the bumps. This stops the waves from bouncing back, allowing them to flow smoothly across a huge range of frequencies (from 220 to 500 GHz).

3. The Results: A Terabit-Per-Second Express Lane

Because the road is so smooth and wide, they were able to send a massive amount of data at once.

The Speed: They achieved a data transfer rate of 1.004 Terabits per second. To put that in perspective, that's like downloading the entire contents of a massive library in a fraction of a second.
The Lanes: They didn't just use one lane; they opened up 14 different lanes (channels) simultaneously, each carrying data at different speeds using complex coding (like QAM-64, which is a very efficient way to pack information into a signal).
The Turns: Real chips need to turn corners. The researchers tested this highway with sharp 90-degree turns. Even with the turns, the data flow remained incredibly fast (0.895 Terabits per second), proving the road is flexible enough for complex chip layouts.

4. Why It Matters for Your Future

The paper claims this technology solves a major bottleneck for Artificial Intelligence (AI) and 6G networks.

AI: As AI gets smarter, it needs to move data between different parts of a computer chip faster than ever before. This new "silicon highway" allows for that speed without burning too much energy.
Simplicity: Unlike fiber optics, which require complex lasers and lenses, this system is made entirely of silicon. It's like building a highway out of the same material as the cars, making it easier and cheaper to mass-produce using standard factory techniques (CMOS-compatible).

In a nutshell: The researchers built a smooth, pothole-free, silicon-based road for invisible waves. This road can handle a massive amount of traffic (data) at incredible speeds, even when it has to make sharp turns, paving the way for faster, more efficient computers and AI systems.

Technical Summary: Dispersion-Engineered Terahertz Silicon Interconnects Enabling Terabit-Scale Data Links

Problem Statement
The exponential growth of artificial intelligence (AI) and data-centric computing is driving data traffic toward exabyte scales, pushing data center and high-performance computing (HPC) infrastructures to their physical limits. While optical interconnects offer high capacity for long-reach communication, their reliance on complex electro-optic transceivers introduces significant energy overheads and scalability challenges for short-reach, chip-scale links. Conversely, conventional metallic Terahertz (THz) interconnects are constrained by low bandwidth-distance products. Although THz silicon waveguides offer a promising CMOS-compatible alternative with low-loss propagation and high mode confinement, existing platforms face critical limitations: they are often restricted to single THz bands, suffer from Bragg-induced stopbands at high frequencies due to structural periodicity, and struggle to simultaneously support multi-band transmission, dual-polarization, low group-velocity dispersion (GVD), and terabit-per-second throughput.

Methodology
To address these constraints, the authors developed a CMOS-compatible, centimetre-scale, unclad silicon waveguide platform fabricated on high-resistivity float-zone intrinsic silicon (HR-Si). The core innovation lies in a dispersion-engineered, effective-medium-supported design that suppresses Bragg scattering.

Structure: The waveguide features a central unclad silicon core (width $w = 0.26$ mm) flanked by effective-medium claddings formed by periodic arrays of circular and semi-circular through-holes. The semi-circular geometry adjacent to the core is specifically engineered to suppress Bragg scattering, flattening transmission spectra and minimizing GVD.
Fabrication: Devices were fabricated using deep reactive ion etching (DRIE) on 500 $\mu$ m HR-Si wafers, followed by back-grinding to 220 $\mu$ m. The process utilizes standard single-mask etching, ensuring CMOS compatibility.
Characterization: The team systematically evaluated straight waveguides and waveguides with 90° and 180° bends (with subwavelength radii $R \approx 0.83\lambda$ at 500 GHz). They measured transmission ( $S_{21}$ ), group delay, insertion loss, path loss, and bending loss across the WR-3.4 (220–330 GHz) and WR-2.2 (330–500 GHz) bands for both transverse-electric (TE) and transverse-magnetic (TM) polarizations.
Data Transmission: A hybrid photonic transmitter (Tx) and electronic receiver (Rx) architecture was employed to demonstrate on-chip data links. The system utilized high-order complex modulation formats (QAM-16, 32, 64, and APSK-16) across multiple non-simultaneously excited channels.

Key Results
The study demonstrates the first on-chip THz communication link achieving terabit-scale aggregate throughput with the following specific metrics:

Bandwidth and Transmission: The platform supports a broadband 3-dB bandwidth of 280 GHz (220–500 GHz), the largest reported for THz silicon waveguides. The design successfully suppresses Bragg-induced stopbands, yielding flat transmission and low-ripple group delay across both WR-3.4 and WR-2.2 bands.
Loss Characteristics:
- Insertion Loss: Average coupling losses of 0.21 dB (TE) and 0.48 dB (TM).
- Path Loss: Average values of 0.22 dB/cm (TE) and 0.19 dB/cm (TM), with minimums as low as 0.016 dB/cm.
- Bending Loss: Average losses of 0.35–0.37 dB per turn for 90° bends and 0.39–0.56 dB per turn for 180° bends, confirming that tight subwavelength bends impose only a modest penalty.
Dispersion: The group-velocity dispersion (GVD) was minimized to 0.18 ps²/mm for straight waveguides and 0.15 ps²/mm for 90° bent waveguides across the full band, effectively reducing pulse broadening and intersymbol interference.
Data Rates:
- Straight Waveguide (WB1): Achieved an aggregate data rate of 1.004 Tbps using 14 channels.
- 90° Bent Waveguide (90B1): Achieved an aggregate data rate of 0.895 Tbps using 12 channels.
- Modulation: Successful transmission was verified using QAM-16, 32, 64, and APSK-16, with Error Vector Magnitude (EVM) measurements confirming minimal signal impairment even at frequencies above 400 GHz.

Significance and Claims
The authors claim that this work establishes a scalable, energy-efficient platform for multi-band THz interconnects that directly addresses the bandwidth and energy bottlenecks of next-generation AI systems and emerging 6G technologies. By suppressing Bragg stopbands and engineering dispersion, the platform enables high-density on-chip and chip-to-chip communication fabrics that bridge the performance gap between on-chip electronics and off-chip optics. The demonstration of terabit-scale throughput in both straight and bent configurations, while maintaining CMOS compatibility and low loss, represents a significant step toward integrated electronic-photonic THz systems with enhanced spectral efficiency and channel counts. The paper asserts that no prior work has simultaneously demonstrated THz waveguides that suppress Bragg stopbands, support multiple THz bands with low loss and low GVD, and achieve near-terabit data transmission using multiple carriers in both straight and bent geometries.

1. The Problem: The "Traffic Jam" of Light

2. The Solution: A Smooth, Custom-Paved Road

3. The Results: A Terabit-Per-Second Express Lane

4. Why It Matters for Your Future

More like this