A survey from Ciena printed earlier this yr revealed the twin nature of the potential influence of synthetic intelligence on telecom networks.
On one hand, greater than half of the telecom and IT engineers surveyed reported that they thought the usage of AI would enhance community operational effectivity by 40% or extra—let’s name that the “AI for the community” side. However when requested in regards to the wants of the opposite side, “the community for AI”, almost all the respondents—99%—stated that they believed fiber community upgrades might be required to be able to help extra AI site visitors.
“The survey highlights the optimistic long-term outlook of CSPs concerning AI’s means to boost the community in addition to the necessity for strategic planning and investments in infrastructure and experience to completely understand the advantages,” stated Ciena CTO Jürgen Hatheier. In a current interview with RCR Wi-fi Information, two different consultants from Ciena mentioned the dueling elements of AI in telecom and the community particularly.
AI for the community
“AI, and issues like ML and information analytics, have been embedded in assurance for a lot of, a few years, for each on a regular basis necessities like sooner trouble-shooting and fault isolation, and extra lately, newer use circumstances round proactive fault identification, isolation and prevention,” mirrored Kevin Wade, senior director of product advertising at Ciena’s Blue Planet division, which focuses on community automation and orchestration. He sees AI’s general function inside community operations as an extension of automation: One other solution to leverage information to optimize operational processes. By way of the place he sees operators being occupied with AI use circumstances, the bulk at this second are targeted on assurance and optimizing community planning. Operators have a number of information, he identified, and as they construct extra hyperlinks throughout that distributed information, the extra insights they get and the higher they’ll plan the evolution of their providers, their networks and their enterprise. It is a related, however maybe extra refined, evolution of conventional telecom purposes of AI and ML.
However over the past couple of years, Wade famous, service suppliers have additionally develop into extremely occupied with generative AI specifically, and its implications for his or her companies and their networks. “That’s a essentially totally different method,” Wade stated. “Sure, it’s all AI, nevertheless it’s not essentially an evolution of ML.”
Gen AI has been essentially constructed round pure language and enormous language fashions (LLMs)—not as an extension of AI/ML that lives on the earth of knowledge largely generated by community tools and software program. So the gen AI use circumstances that Wade sees the business working in the direction of that instantly influence the community itself are basically an extension of, or maybe an intersection of, coding and orchestration and intent-based networking. “The concept is perhaps, let’s use pure language for an end-customer to specific their intent of what they need for a service, a connection from this level so far, for this period of time, this quantity of bandwidth, with this sort of safety privilege connected—if you happen to can simply say that or write that down in easy language, and it automagically occurs,” Wade explains.
That’s the objective of AI in telecom that service suppliers are trying towards, from Blue Planet’s view, “Nevertheless it’s actually nonetheless very a lot within the formative stage,” Wade provides. “It would take a few years, in all probability, to get there, as a result of there are not any requirements for gluing this all collectively—these are additionally simply being formulated.” The Extremely Accelerator Hyperlink (UALink) group, which is targeted on standardizing interconnect interfaces for AI accelerators inside information facilities, was simply established earlier this yr—and likewise is seen as an effort to set up a substitute for Nvidia’s NVLink.
That need to keep away from proprietary know-how can also be carrying over to the fashions for gen AI, Wade stated, and it could result in some community operators treading frivolously on gen AI till extra interoperable or standardized frameworks for making use of AI emerge. “Some service suppliers, proper now, are at all times involved with lock-in … round software program distributors specifically,” he stated. “They don’t essentially wish to be locked into one LLM both. So … there’ll be a mixture of some ready till some requirements, guardrails, are in place for interoperability and so forth. However others have initiated, and among the bigger European operation specifically initiated their very own, telco-specific LLM actions.” (SK Telecom, Deutsche Telekom, e&, Singtel and SoftBank, after making a dedication at MWC Barcelona 2024, introduced a three way partnership in June of this yr to collectively develop and launch a multi-lingual LLM particularly for telcos, with an preliminary focus that features the usage of gen AI in digital assistants for customer support.)
The community for AI
“AI infrastructure challenges lie in cost-effectively scaling storage, compute, and community infrastructure, whereas additionally addressing large will increase in power consumption and long-term sustainability,” wrote Brian Lavallée, senior director of market and aggressive intelligence at Ciena, in a current weblog put up. He identified that “Conventional cloud infrastructure success is pushed by being cost-effective, versatile, and scalable, that are additionally important attributes for AI infrastructure. Nevertheless, a brand new and extra in depth vary of community efficiency necessities are wanted for AI.”
That features each inside the information middle and out of doors it. Lavallée cited numbers from Omdia on anticipated site visitors progress for AI, with month-to-month “AI-enriched” community site visitors anticipated to see a 120% compound annual progress charge by way of 2030. He additionally touched upon the wants of generative AI to maneuver large quantities of knowledge inside an information middle, over hyperlinks working at 400G, 800G and 1.6 Tb/s or extra. Ciena has had two current trials of 1.6 Tb/s capabilities, one with Telstra and Ericsson and one other with international fiber spine supplier Aurelion.
“We all know inside the information middle, site visitors is exploding already at this time,” Lavallée stated. “It’s going to spill out in a short time into campus networks.” He expects to see information facilities start to be construct in that unfastened campus type, with a number of buildings inside 10 kilometers of each other—resulting in the virtualization of knowledge facilities. “You’re going to have a number of information facilities appearing as one bigger, digital information middle. … for an entire bunch of causes,” Lavalee added—primarily, energy. “There’s not sufficient electrical energy in current information facilities to park all of the AI {hardware}, which is 10 occasions extra capability per rack,” he continued. “So you might have 10 occasions much less house consumed, however you’ve used 100% of the electrical energy coming into that constructing.”
Lavallée factors out in his weblog put up that the gen AI fashions are “notoriously power-hungry of their LLM coaching section” and eat “immense quantities of electrical energy.” Energy utilization in data-center scorching spots comparable to Ashburn in northern Virginia, is anticipated to double over the following 15 years, pushed largely by information middle progress. The GPU utilization depth wanes as soon as a gen AI mannequin is sufficiently educated and “pruned”, nonetheless, and affords a chance for algorithms to be moved to a extra distributed edge location nearer to end-users.
Whereas the ability wants and intensive high-speed hyperlinks inside the information facilities are already changing into obvious, it’s much less sure what the wants of the remainder of the community might be. As Lavallée informed RCR Wi-fi Information, whereas there are some estimates on the general site visitors influence that AI could have, that also leaves some vital uncertainty about precisely the place within the community, and by how a lot, information transmission capability will must be bolstered. In metro rings? In long-haul hyperlinks? Throughout submarine cables? That breakdown isn’t recognized but. And Lavallée makes the purpose in his put up that “AI will solely scale efficiently if information can transfer securely, sustainably, and cost-effectively” from core information facilities to edge information facilities.
He additionally thinks that the community efficiency calls for of AI in telecom could in the end imply that it is rather well-suited to being supported by 5G—which, in any case, is meant to be a extremely distributed, cloud-native transmission community that may help excessive information charges and low latency.
“I believe 5G was a community improve seeking a use case. AI is the use case,” stated Lavallée. “If we will marry the 2 collectively, I believe among the promise and alternative of 5G will be enabled with synthetic intelligence.”
