Anyone in networking understands that bandwidth growth is a constant. We’ve had to adapt to meet the increasing demand from consumers for data; from the now-minor transmission of email to the streaming of videos in 4k to the access and analysis of big data, the key to a seamless digital experience has always been bandwidth.
Data centres have mostly been built in close proximity to users (i.e. close to major urban areas). This has helped to reduce latency and the mean time to cloud, better providing that seamless experience that is critical to uptake.
More recently, the cost of power, proximity to renewables and technology infrastructure (5G and fiber to the home) are now more pressing concerns, in some cases pushing data centers further away from the user and near unpopulated areas, particularly as large campuses of data centres can place a strain on local power grids. With high-speed connections, it’s still feasible to maintain fast connectivity over great distances, and advances in technology continue apace to ensure the experience remains relatively seamless.
But Artificial Intelligence (AI) is posing a new challenge and opportunity for service providers. The challenge is not so much in just the extra bandwidth demands, but also determining a new strategy for data center builds. For AI to be truly effective, inference — the ability of AI to process commands based on requests — needs to happen instantly. This will require more, albeit smaller, data centers at the network edge so inferencing is situated as close as possible to end users, both humans and machines, to minimize the incurred latency.
Take a smart city leveraging AI and 4K surveillance cameras. These cameras need to monitor many things and detect everything from traffic patterns to suspicious activity, such as the real-time detection of a crime taking place and advances in motion image detection can create much more detailed representations of traffic flow and also detect such instances as a crime being committed or a weapon being drawn. These use cases create a greater demand on bandwidth given the higher-resolution video-centric data, and once we bring AI into the equation, which can create scenarios and selects how to best control the traffic or respond to threats, we’re talking about a requirement for large amounts of complex inferencing at that edge.
Furthermore, this edge needs to work with the other “edges” to create a bigger picture of traffic flow across a city and not merely at that localised level. Should data have to traverse continents to be inferred and sent back, the time to do so could be seconds to minutes — this clearly can’t be the way forward for this kind of system.
So, we’re looking at two factors at play: data centres being built in regions far, far away from users, sometimes overseas, and concurrently, a requirement to infer at the edge faster than ever.
Enter: submarine cables. We forget about this infrastructure but it’s perhaps the most important part of the global internet, the veins pumping the data lifeblood as fast as possible across vast distances and doing so invisibly. These submarine networks carry over 99% of the world’s intercontinental electronic communications traffic over a web of well over 500 submerged cables. The challenge has always been to keep data flowing reliably and cost-effectively, while continuing to maximize optical spectrum efficiency.
But these informational arteries are going to be even more critical in this new, dispersed environment, carrying AI-originated traffic from far-off locations as fast as possible, and connecting with metro and terrestrial networks to enable the inference required at the edge. And growing disaster recovery and redundancy requirements require massive submarine connectivity between data centres — the distance between Australia and the U.S. mainland might be vast, for instance, but to the average user this can’t be felt, particularly during downtime if we’re to increasingly rely on AI for our day-to-day needs.
Submarine cables, therefore, need to land at places where they previously haven’t — away from populated areas and closer to power sources. And there are now additional challenges in that traffic patterns will look different, particularly with AI; training loads have different network requirements from inference traffic, which is also different from “regular” traffic. The issue is that it’s not easy to simply build a new submarine cable — these are projects that cost in the hundreds of millions of dollars and take years to plan and deploy.
So, what do we do? Yes, we’ll need more submarine cables connecting to new, far-flung locations… but given the expense and time required to lay a new cable, we must optimise and enhance what we already have installed and deployed. Ciena recently released findings for a survey of more than 1,500 telecom and IT engineers and managers at CSPs in 17 countries across the globe. Most respondents — 99% — believe service providers will need to upgrade fiber-optic networks to support more AI traffic, including submarine cable networks.
We’re now at the stage in which 400Gb/s is the new 100Gb/s — basically, it’s the baseline submarine cables need to operate at to be successful at transmitting AI-generated data back and forth between remote data centers and terrestrial networks.
Coherent optical solutions are evolving, with the technology now able to cater to 800Gb/s comfortably and up to 1Tb/s over certain submarine distances — this will enable more traffic to traverse greater distances faster than ever before.
But crucially, submarine cables need to address the flexing needs of data centre interconnect requirements, such as the alternative traffic patterns traversing the network to far-flung places and edge data centres. Large AI clusters and idle compute challenges will drive increased need for load balancing and workload distribution across large distances. The global mesh, enabled by submarine cables, must re-optimize traffic patterns quickly, particularly as we’re expecting traffic to more than double each year thanks to AI.
Ultimately, the key to the AI-driving network of tomorrow will be to ensure that submarine, terrestrial and cloud networks continue to seamlessly blend together as one consumable network asset so AI can be delivered at the speed end users demand.
