I had the good fortune to attend the 50th European Conference on Optical Communications, ECOC 2024, which was held in Frankfurt from September 22 to 26. The conference included 500+ presentations and over 1,800 conference delegates from 82 countries. Running alongside the conference from September 23 to 25 was the ECOC Exhibition, with 7,884 visitors and 330 exhibitors. As in the past few years, the trend of increased network operator presence continued. The Infinera booth hosted demonstrations of the latest Infinera innovations, including ICE-X 800G coherent pluggables and private line emulation, as well as an ICE-X demo that reflects worldwide trials and deployments, including coherent transmission over single-fiber working, metro aggregation with 4 x 100G coherent breakout, and streaming telemetry. Here are my key takeaways on the major innovation trends discussed at the technical conference:
1. Generative AI
As with all conferences these days, AI was a key theme. At the hotel we were staying, in there was a conference for a well-known doughnut brand, and I suspect even they managed to make AI a theme! Power consumption of AI was a key topic, with generative AI racks evolving to 400 kW and total power per data center evolving to up to 400 MW. One controversial topic was the impact of generative AI on network traffic outside the data center. The view from Cignal AI and others is that this impact is currently muted outside the data center, though this will likely evolve as power consumption challenges force AI data centers to be distributed across the metro.
On the other hand, a presentation from Meta gave a contrasting view, stating that AI is already having a dramatic impact on backbone traffic, driving growth from 30% per year to 100% per year. This traffic comes in part from data sent to AI training data centers and data transferred from training data centers to inference data centers. It also arises from user requests that can no longer be answered from caches at the internet service provider but instead have to traverse Meta’s backbone to an AI inference data center, typically in the U.S. NTT later talked about the need for AI-related traffic across the metro from enterprise and government sites to GPU clusters in their data centers when offering AI as a service.
2. “Coherent Lite”
Multiple presentations at ECOC 2024 looked at the related topics of reducing coherent power consumption and coherent lite for data center interconnect (DCI), intra-data center, and metro access applications such as coherent PON. Multiple presentations included a breakdown of power consumption inside coherent pluggables, including the functions inside the DSP. The functions outside the DSP (photonics, analog electronics) and inside the DSP such as the DAC/ADC make up a large percentage of overall power consumption and do not benefit from reduced power consumption as we move to smaller CMOS process nodes (i.e., 3 nm, 2 nm), so we can no longer rely on Moore’s law to reduce the power for us. Coherent lite therefore requires novel designs and techniques to reduce power consumption and cost, with power consumption being the primary metric for data centers and cost being key for access network applications such as coherent PON. One interesting technique presented by my colleague Carlos Castro was to turn off digital subcarriers in a point-to-multipoint network when the data capacity is not required (i.e., in the middle of the night), which can reduce DSP power consumption by up to 40%.
3. Scaling Capacity: Multi-band and Multi-rail
Multiple presentations looked at how to scale capacity as we approach the Shannon limit. One option championed by Infinera and others was Super C and Super L. For more spectrum beyond this, you have to look at the S-band and the E-band. Challenges of these novel bands include high cost and power-inefficient amplification, as well as stimulated Raman scattering causing power to transfer between bands. The results of a survey held at the end of the session asking when these multi-band systems would be commercially available were: five years (29%), ten years (27%), when hollow-core fiber is deployed (21%), and never (17%). Another option for scaling capacity is multiple parallel fiber pairs, often referred to as “multi-rail.” Google’s presentation showed that this is already happening. Other presentations compared options for multi-rail including stacking ROADMs and combining wavelength switching (i.e., ROADMs) with fiber switching.
4. QKD/Security
Quantum key distribution (QKD) was prominent at the conference, with around 30 presentations/papers. There was also a multi-vendor QKD interoperability demonstration. Multiple presentations covered continuous-variable QKD (CV-QKD), which can coexist with data-carrying DWDM wavelengths on the same fiber and can leverage existing optical transport technology and components. Reference was made to an indium phosphide (InP) QKD photonic integrated circuit (PIC) for CV-QKD, yet another novel use for InP PICs. This contrasts with discrete-variable QKD (DV-QKD) which requires specialized technology/components and dedicated fiber. However, DV-QKD can currently extend to longer distances, approximately 100 km, compared to approximately 50 km for CV-QKD. Mixing CV-QKD and data wavelengths on the same fiber also involves a trade-off between the QKD data rate and the data wavelengths’ transmit power levels and therefore data rates.
On a related security topic, there was an interesting presentation on eavesdropping and how to detect it. Eavesdropping on optical communications is typically done by bending the fiber, which causes a small percentage of the light to leak out. However, the loss is very small (~0.2-0.3 dB) and hard to detect. The solution proposed was to monitor the state of polarization and leverage machine learning to identify eavesdropping signatures.
5. Hollow-core Fiber
Hollow-core fiber (HCF) has gained a lot of attention since Microsoft acquired Lumenisity back in December 2022. At this year’s ECOC, I counted 16 presentations on HCF. Many of these presentations were from Chinese presenters, indicating a lot of interest in HCF in China. China Mobile is in fact planning a 100-km field trail in 2025. I also learned about a novel application for HCF using very high-power lasers to distribute power to remote sites such as mobile base stations or to power home fixed-line communications in the event of a power outage. However, despite its many potentially game-changing benefits (low latency, low loss, low nonlinearities, wide spectrum, etc.), HCF challenges still exist. One of these is coupling HCF with regular single-mode fiber (SMF). Changing HCF’s design for better coupling (narrower core) has a negative impact on its performance, including higher loss. Manufacturing single stands of HCF beyond ~10-15km is also a challenge – changing the design (e.g., thicker cladding tubes) to enable longer fibers can also have a negative impact on the performance of the HCF. A potential killer application discussed for HCF is leveraging its low latency for distributed metro AI data centers. HCF’s low latency can also be very valuable inside generative AI data centers.
Summary
Overall, ECOC 2024 was a great event that also covered a wide range of additional interesting topics including network automation, telemetry, novel materials, multi-core fibers, submarine networks, metro access convergence, next-generation PON, free-space optics, environmental and network monitoring, and optical circuit switching (OCS) inside the data center.
