ANDREAS MATISS CO-PACKAGED OPTICS
CPO has been around for a while, however it is still a technology in development. Andreas Matiss , Senior Manager Optical Components and Integration, Corning Optical Communications explains how glass can play an important role in moving silicon-based electro-optical converters as close as possible to the silicon processor. UNLEASHING THE POTENTIAL OF CPO AND GLASS WAVEGUIDE SUBSTRATES A LEAP TOWARDS FIBRE-TO-THE-CHIP TECHNOLOGY
D ata centre networks are evolving rapidly. This has accelerated further with the emergence of artificial intelligence (AI) and large deployments of AI clusters. This trend has accelerated recently, with notable deployments such as NVIDIA’s DGX SuperPOD and Google’s TPU clusters. The shift is fueled by the need for high-performance computing to support AI training and inference tasks. It is expected that multiple million AI-optimised GPU units will be shipped annually in the next five years just by NVIDIA alone, reaching high volumes by 2028. The number of transceivers required to build these networks are in the range of tens of millions transceiver units per year operating at the highest speeds, like 1.6Tbps and 3.2Tbps. Industry analysts are expecting the number of transceivers per GPU to go up to more than 10 transceivers per accelerator (GPU) in the future, which would correspond to roughly a 10x increase in optical connectivity needs versus today’s deployments. In a typical data centre, a standard pluggable Ethernet transceiver consumes about 20W of electrical power. For a next generation transceiver, the power consumption is expected to almost double. With shipment volumes today, roughly 200 Mega-Watt (MW) of power to drive transceivers is deployed in 2024. Continuing on the transceiver trajectory and the expected 10x increase in optical connectivity needs, the annual deployed power for transceiver could increase to 2 Giga-Watt (GW) per year, about
the power generated by a large nuclear power plant. This does not include the power required to drive the host-side electronics and electrical re-timers to move the data from the integrated circuit to the transceiver at the front of the equipment. As an example, the introduction of CPO in an AI data centre with one million GPU’s would save about 150 MW of power generation capacity alone, for one single data centre. Besides the investment costs to build the required power plant capacity, there are very significant operating cost savings that, depending on the regional energy costs, can easily exceed 100M Euro per year in power bills. Addressing this unsustainable trend requires new innovations. INTRODUCING CO-PACKAGED OPTICS Co-packaged optics (CPO) is the most promising technology to overcome this power bottleneck in the near future. This technology describes moving the electro-optic conversion module from the transceiver at the front-plate deeper into the equipment. Ideally, directly on the CPU or GPU package substrate. This way, the electrical losses over copper lanes can be minimised. This accounts for more power efficient links and can reduce the power consumption compared to a pluggable transceiver by more than 50% - even 75% in some instances. The power saving is not only derived from removing lossy copper lanes, but also simplifying, or even eliminating, the digital signal processor (DSP) that is needed to compensate signal impairments over the electrical pathway.
A secondary effect is that the latency is improved, as a signal transmitted over a co-packaged optics module needs less pre- and postconditioning for error-free communication. In summary, co-packaged optics offer high-speed, low power and low latency optical connections. Everything that is key for advanced AI networks. Another notable alternative to reduce power is Linear-Drive Pluggable Optics (LPO), which eliminates DSP chips, thus reducing power consumption and latency while maintaining the front-plate pluggable transceiver form factor and ecosystem. While CPO offers better signal integrity and lower latency, LPO is more cost-effective and suitable for short distances. LPO’s cost-effectiveness and lower power consumption might delay CPO’s widespread adoption due to quicker market readiness. Still, as LPO consumes more power than co- packaged-optics and is also significantly more difficult to manage to maintain high signal quality as link speeds go up to 200G and beyond, CPO is expected to become the solution of choice when the technology becomes available over time. HOW GLASS IS ENABLING CO- PACKAGED OPTICS Glass is expected to play an important role in the next generation of co- packaged optics. In order to move the electro-optical converter, mostly a silicon photonics chip, as close as possible to the actual silicon processor (CPU/GPU), a new packaging technology is required that allows for large substrate size and also optical connectivity to the silicon photonics chips.
14
www.opticalconnectionsnews.com
| ISSUE 41 | Q2 2025
Made with FlippingBook interactive PDF creator