ANDREAS MATISS CO-PACKAGED OPTICS
CPO has the potential to revolutionise the structure of interconnections for AI in multiple directions.
Traditionally, organic substrates are used for semiconductor packaging. These materials have larger thermal expansion coefficients than silicon which limits the maximum size a semiconductor package can have. As the industry is pushing for larger packaging substrates on the incumbent organic technology platforms, problems related to reliability, such as integrity of solder joints and an increased risk of delamination, as well as manufacturing challenges like the creation of high quality fine-pitch interconnects and high-density wiring are emerging and drive up packaging and testing costs. Glass can be engineered to have an optimised thermal expansion coefficient that is better matched to the silicon chips compared to the organic substrates. As a result, engineered glass substrates offer superior thermal stability, reducing the risk of mechanical stress and damage during temperature changes. They provide excellent mechanical strength and flatness, enhancing the reliability of chip packages. Additionally, glass supports higher interconnect densities and finer pitch interconnects, improving electrical performance and reducing parasitic effects. These properties make glass a highly reliable and precise option for advanced semiconductor packaging. Therefore, there are significant efforts in the semiconductor packaging industry to leverage advanced glass substrate as the next generation substrate technology. GLASS WAVEGUIDES SUBSTRATES In addition to its excellent thermal and mechanical properties, glass can be treated in a way that allows the implementation of optical waveguides. These waveguides in glass are typically created using a process called ion exchange. This involves replacing ions in the glass with different ions from a salt solution, altering the glass’ refractive index. The modified regions guide light by confining it within the higher refractive index area. This technique allows precise control over the waveguide’s properties, making it suitable for various optical applications. Therefore, in an optical waveguide - similar to an optical fibre - the light can propagate along the integrated glass waveguides and efficiently couple either into a fibre or into a silicon photonics
chip. This makes glass a very attractive material for advanced co-packaged optics applications. Combining electrical and optical interconnects on the same substrate also resolves the interconnect density problem that companies are facing when building large AI clusters. The number of optical lanes is currently constrained by the geometry of the optical fibre, which typically has a cladding diametre of 127 micrometres, roughly the thickness of a human hair. However, waveguides on glass can be made significantly denser, allowing for much higher input/output (I/O) density compared to what is achievable with direct fibre-to-chip connections. This integration of electrical and optical interconnects not only resolves the density issue but also enhances the overall performance and scalability of AI clusters. By leveraging the compact nature of glass waveguides, it becomes possible to fit more optical lanes within the same physical space, thereby increasing the data transfer capacity and efficiency of the system. This advancement is crucial for the development of next-generation AI infrastructure, where high-density interconnects are essential for managing the vast amounts of data processed by these systems. Integrated glass waveguides would also allow the building of full optical systems on the same substrate, with photonic integrated circuits communicating with each other directly through the optical waveguides. This could be achieved without the need for fibre interconnects, significantly enhancing the chip-to- chip communication bandwidth and reach. Using glass optical waveguides in densely packed systems with many interconnects offers lower signal loss, higher bandwidth density, and improved durability compared to discrete optical fibres. These advantages make glass waveguides ideal for high-performance
to rethink and redesign their networks. With the increased bandwidth and simplified network structures, they can achieve better network performance properties, leading to more efficient and streamlined operations. CONCLUSION CPO has the potential to revolutionise the structure of interconnections for AI in multiple directions. It can significantly improve power consumption and sustainability, making AI systems more environmentally friendly and cost- effective. Additionally, CPO facilitates better efficiency and scalability, allowing AI systems to handle larger and more complex tasks with ease. By addressing density, CPO can enhance data transfer rates, ensuring faster and more reliable communication between AI components. This also helps in decreasing bottlenecks in upcoming AI systems, leading to smoother and more efficient operations. The future of AI interconnections is expected to introduce direct optical links, which will remove the need for compute switches. This innovation will expand bandwidth for AI tasks, enabling faster and more efficient processing of large datasets. The optimal vehicle for achieving these advancements is glass, which offers superior properties for data transmission and scalability. Glass-based optical links will play a crucial role in the next generation of AI systems, providing the necessary infrastructure for high- performance computing and advanced AI applications.
optical interconnect systems. Using co-packaged optics in
next-generation data centres and AI supercomputer networks increases the chip escape bandwidth, opening up numerous possibilities for creating high-speed and high-radix switches for 102T and beyond. Network architects are presented with a unique opportunity
Senior Manager Optical Components and Integration, Corning Optical
www.opticalconnectionsnews.com
15
ISSUE 41 | Q2 2025
Made with FlippingBook interactive PDF creator