The escalating bandwidth requirements of AI scale-up architectures are pushing current electrical interconnects toward their physical limits, necessitating a shift toward optical solutions. To address this bottleneck, Lumentum, Qualcomm Technologies, and Corning Incorporated have announced a joint demonstration of a high-density die-to-die (D2D) optical interconnect scheduled for ECOC 2026 in Málaga, Spain. This collaboration aims to showcase a massively parallel architecture capable of extending high-bandwidth, low-power, and low-latency connectivity over multimode fiber for distances spanning tens of meters. By integrating Qualcomm’s D2D interface IP, Lumentum’s 1060 nm VCSEL-based optical platform, and Corning’s multimode fiber connectivity, the trio is testing a practical roadmap for future co-packaged optics (CPO) and near-packaged optics (NPO) implementations.
Qualcomm and Lumentum Architecting Optical D2D Interfaces
The proposed demonstration focuses on overcoming the density and distance limitations inherent in traditional electrical die-to-die connections. Qualcomm Technologies is contributing a D2D interface IP subsystem designed to optimize power and area while enabling efficient direct-drive optical connectivity. This technology is intended to support emerging interfaces, such as UCIe, within future AI system architectures. The current proof-of-concept configuration utilizes optical channels operating at 32 Gb/s NRZ, a baseline intended to demonstrate how the massively parallel architecture can scale through increased channel counts and higher per-lane data rates.
Lumentum is providing the optical engine, which utilizes advanced-packaged 1060 nm VCSEL technology. This platform features high-density two-dimensional VCSEL and photodetector arrays enhanced with backside lens technology, aiming for approximately 10 Tb/s of aggregate transmit and receive capacity. The architecture is positioned to support a shoreline bandwidth density of approximately 1 Tb/s/mm today, with a projected path toward 4 Tb/s/mm as lane rates and channel densities increase. This approach seeks to provide the scalable bandwidth density required as AI workloads drive interconnect demands beyond the capabilities of existing electrical infrastructure.
Corning Fiber Integration for Advanced Packaging Environments
As optical interconnects move closer to compute devices, the physical medium must adapt to higher density and precision requirements. Corning Incorporated is integrating its multimode fiber glass ferrules into the demonstration to provide high-density fiber connectivity. This component is critical for supporting the evolution of optical interconnects within advanced packaging environments, such as co-packaged optics (CPO) and near-packaged optics (NPO).
The collaboration highlights the necessity of integrating specialized fiber solutions directly into the compute ecosystem. Corning’s involvement suggests that as optical links transition from board-level to system-level AI scale-up architectures, the precision and scalability of the fiber connectivity will become a primary factor in maintaining signal integrity and energy efficiency. By combining these elements, the partners are demonstrating a complete link—from the silicon die through the optical engine to the fiber medium—designed to meet the rigorous demands of next-generation AI infrastructure.
Key Takeaways
- The joint demonstration at ECOC 2026 will showcase a massively parallel optical D2D architecture supporting emerging interfaces like UCIe.
- The architecture aims for a current shoreline bandwidth density of 1 Tb/s/mm, with a projected capability of 4 Tb/s/mm through higher lane rates.
- Lumentum’s optical engine utilizes 1060 nm VCSEL technology to provide approximately 10 Tb/s of aggregate transmit and receive capacity.
TechInsyte's Take
In our view, this collaboration signals a critical industry pivot from electrical to optical die-to-die interconnects as the primary method for scaling AI compute clusters. The move toward 1060 nm VCSEL technology and the focus on shoreline bandwidth density (targeting 4 Tb/s/mm) suggests that the industry is no longer just looking for faster links, but for a fundamental redesign of how silicon components communicate. By aligning Qualcomm’s IP, Lumentum’s photonics, and Corning’s physical fiber, these companies are attempting to standardize a vertically integrated optical stack. For enterprise architects, this indicates that the next generation of AI hardware will likely move away from traditional pluggable optics toward highly integrated, near-packaged optical solutions to solve the looming "interconnect wall."
Questions & Answers
How does this architecture address the limitations of current AI interconnects?
The architecture replaces or augments electrical connectivity with a massively parallel optical D2D design. This allows for high-bandwidth, low-latency communication over distances of tens of meters, which helps bypass the physical density and power constraints that currently limit electrical links in AI scale-up environments.
What are the specific bandwidth density targets for this technology?
The architecture is designed to support approximately 1 Tb/s/mm of shoreline bandwidth density in its current state. The partners have indicated a roadmap to reach approximately 4 Tb/s/mm by increasing lane rates and channel density.
Which emerging industry standards is this technology intended to support?
The demonstration is designed to support emerging die-to-die interfaces, specifically mentioning the Universal Chiplet Interconnect Express (UCIe) standard, within future co-packaged optics (CPO) and near-packaged optics (NPO) implementations.
What role does the 1060 nm VCSEL technology play in the system?
Lumentum’s 1060 nm VCSEL-based optical engine provides the high-density optical connectivity required for the link. It is designed to deliver approximately 10 Tb/s of aggregate transmit and receive capacity using two-dimensional VCSEL and photodetector arrays with backside lens technology.
Source: Lumentum
