Co-packaged optics for hyperscale data centres

The continuing push for greater efficiency, smaller form factors and faster throughput is leading to the development of next-generation co-packaged optics for datacentres and related applications. EPIC explores continuing research from major manufacturers working to keep the data moving while the world’s appetite from broadband grows with each passing day.

BY IVAN NIKITSKIY, PROGRAM MANAGER, PHOTONICS TECHNOLOGIES, EUROPEAN PHOTONICS INDUSTRY CONSORTIUM (EPIC)

With data centre traffic growing at an unprecedented pace, fuelled by advances in AI and Machine Learning, networking infrastructure must scale in capacity while maintaining or even reducing its total power consumption and footprint. How is the industry going to move forward to hyperscale data center operations with the introduction of next-generation / second- generation Co-Packaged Optics (CPO)? Here we overview how the large industry players, members of EPIC – the European Photonics Industry Consortium – address user requirements for CPO from different perspectives.

HiSilicon (a Huawei company), China
Huawei is a major player in co-package optics, and their Advanced Photonic section develops the photonic optical components that go into Huawei’s larger systems. They are currently addressing the challenges of developing 100 Tbit/s co-packaged optics. Eric Bernier, Leader of Advanced Photonics at HiSilicon Technologies in Canada, the former ASIC Design Center of Huawei explains: the general consensus within the industry is that at the bandwidth required by a 100 Tbit/s switch, it becomes impossible to move the data electronically without consuming the entire power budget for the switch chip. As a consequence, at 100 Tbit/s, co-packaged optics are essential. But from a module perspective, it’s not possible to double the number of modules because of reliability issues and also because it becomes harder to package. As a result, the solution for achieving the required higher density will be to increase the capacity of the co-packaged optics to around 200 Gbit/s per fiber coming out of each module with multiple wavelengths per fiber. This will require inputting more optical power into the system, and although a lot of progress has been made, Bernier believes that ultimately, they will need new technology.

For this reason, they are presently engaged on two research projects initiated by The International Photonics Advocacy Coalition (IPAC). One aims to develop a standard form factor for the external laser source, and the other is looking at the issues, the system architectures and the evolution of the electronics that are limiting 100 Tbit/s co-packaged optics. Currently, it is generally assumed that any future device will incorporate a connector because it will be easier for the supply and assembly chains. However, if the aim is to increase density while simultaneously reducing the power requirement on the staircase, the connector may have to be eliminated.

Senko Advanced Components Inc., Japan:
Tiger Ninomiya, Senior Technologist at Senko Advanced Components in the US, identifies four main challenges for CPO connectors in a data centre switch applications: 1) an increase in fiber count and how to arrange the fibers in and out; 2) a use of external laser sources; 3) a change in face plate density that requires reserved spaces for laser sources and TRx channels, and 4) the challenge of internal fiber routing as fibers are now inside the system.

As regards to fibre count increase, the 12.8 terabit switch typically has 32 ports with eight fibers per transfer module that adds up to a total of 256 fibers in the case of using parallel optics. Similarly, the 51.2 terabits switch with CPO has 16 modules embedded on the switch ASIC substrate. And with parallel optics, each one of CPO optical engines can have up to 64 fibers, which adds up to a total of 1,024 fibres just for TRx – 4 times what they are dealing with now.

The issue with faceplate density derives from the need to have more fibers and to find space for external laser sources. The MPO connector has better density over duplex types of connectors. However, there is a correlation between fiber count per connector and optical performance. With a larger fiber count, it becomes challenging to maintain the lower loss, especially having multiple rows of multi fibers such as MPO-24 and MPO-32.

Senko is addressing this issue with their SN-MT connector carrying 16 fibers per connector, which improves fiber density at the panel while maintaining lower loss. Compared with MPO connectors, the SN-MT is roughly half the size and has a 2.7x density increase compared with MPO-16F. SN-MT even provides a better density than MPO-24 and MPO-32, while using a 1-row type ferrule, the optical performance is comparable to 1-row MPO. Senko also uses other technologies to overcome face plate density issues. These include fiber routing options for mid-board connectors; a fibre routing shuffle box, and backplane connectors.

The Consortium of On-Board Optics (COBO) and Co-Packaged Optics Working Group, which Tiger is chairing, aims to provide technical guidance and standards for CPO implementations focusing on optical connectivity and remote laser sources. In July 2022 they released a white paper on optical connectivity that details how these technologies can be utilized. [1]