Large-Area Photonic Interconnects Enable Panel-Scale AI Systems on Glass Substrates

Artificial intelligence demands ever-increasing computational power, pushing the limits of traditional electronic interconnects, and researchers are now exploring the potential of light to overcome these challenges. Tzu-Chien Hsueh, Bill Lin, and Zijun Chen, along with colleagues at the University of California, San Diego, present a new approach to building large-scale, reconfigurable photonic interconnects on glass substrates, potentially revolutionising how AI systems are packaged and connected. Their work demonstrates a pathway to creating a panel-scale interconnect fabric that enables high-bandwidth, low-energy communication between processor chiplets and memory stacks without the need for signal repeaters, offering a compelling solution for building significantly more powerful and efficient AI hardware. This innovative system promises to dramatically increase data bandwidth density and enable the heterogeneous integration of complex AI systems on a single photonic interposer, exceeding thousands of square centimetres in size.

Silicon-Rich Nitride for Compact Photonics

Researchers are investigating silicon-rich nitride (SRN) as a promising material for building the next generation of integrated photonic devices, which manipulate light for computing and communication. SRN combines a high refractive index for compact designs, strong nonlinearity for efficient light modulation, compatibility with standard silicon manufacturing, and the ability to tune its optical properties. The team is focused on maximizing SRN’s nonlinear response and pioneering a technique called visible light trimming, which uses light to precisely control SRN’s refractive index across an entire wafer, offering unprecedented control over its optical behavior. This control extends to utilizing the DC-Kerr effect, a phenomenon within SRN that enables all-optical signal processing without converting light to electricity. Researchers are also minimizing optical losses within SRN waveguides, ensuring efficient light transmission. The results demonstrate that SRN exhibits a significantly larger change in refractive index compared to other materials, allowing for more compact and efficient devices.

Large Photonic Interposers with Racetrack Resonators

Researchers are developing large-scale photonic interposers, light-based circuits, to overcome the limitations of current electronic systems used in artificial intelligence and high-performance computing. These interposers, fabricated on glass substrates potentially exceeding 500mm x 500mm, aim to dramatically increase bandwidth and reduce energy consumption by directly manipulating light signals. A key innovation lies in the creation of miniature racetrack resonators, tiny circular pathways for light constructed from a silicon-rich material. These resonators exhibit anomalous dispersion, allowing for precise control over the wavelengths of light.

By carefully designing these resonators, researchers generate a “frequency comb”, a spectrum of evenly spaced wavelengths, which enables wavelength-division multiplexing, transmitting multiple data streams simultaneously on different colors of light. Researchers precisely tune the wavelength of light entering the resonator, triggering interactions that result in the generation of a stable, evenly spaced comb. This control is crucial for creating reliable, high-capacity data links. Furthermore, the team is developing crossbar switches based on these resonators, enabling reconfigurable connections and dynamic data routing, paving the way for more powerful and energy-efficient AI systems.

Photonic Interposer Doubles Data Transfer Speeds

Researchers have developed a new glass-based photonic interposer that dramatically increases data transfer speeds within computer systems, offering a significant advancement over traditional silicon-based interconnects. This innovative technology establishes panel-scale connections, reaching dimensions of 500mm x 500mm or larger, without the need for signal repeaters. The photonic interposer achieves a total data bandwidth of 26. 624 terabits per second per processor, more than double the 13. 312 terabits per second achieved by comparable silicon-based systems.

This improvement stems from the use of wavelength-division-multiplexing, sending multiple data streams on different colors of light through a single optical waveguide. The design minimizes physical interconnect routes and extends communication distances to over 500mm. A key innovation lies in the design of the optical transceivers, which connect processor chiplets to the interposer. These transceivers employ a novel silicon-rich-nitride micro-ring resonator that modulates light with exceptional efficiency, minimizing energy loss and maximizing signal clarity. The 3D integration of these components creates a remarkably short and efficient data path, reducing electrical losses and enabling near-zero DC power consumption.

Terabit-Scale Photonic Interconnects on Glass

This research demonstrates a novel panel-scale photonic interconnect capable of supporting high-bandwidth, low-energy data communication for advanced computing systems. The team successfully designed and tested a photonic switch fabric on a glass substrate, achieving a total data bandwidth of over 7 terabits per second with an average energy efficiency of 1. 1 picojoules per bit. This approach enables all-directional communication across the entire panel without requiring active repeaters. The developed interconnect offers substantial improvements over existing wafer-scale solutions, which are limited by manufacturing constraints and networking diversity.

While current silicon-based interposers are restricted by wafer size and primarily support communication between adjacent chips, this photonic approach allows for unlimited tiling and direct communication between any chiplet across the entire panel. The reconfigurable nature of the system maximizes networking diversity and enables dynamic optimization of computing resource utilization. Researchers are focused on improving component performance, such as detector responsivity and linearity, to further scale data rates. This research demonstrates the potential of glass-based photonics to overcome the limitations of traditional electronic interconnects and enable the next generation of high-performance computing systems.