AI continues to come back at us from all instructions. Think about a world the place billions of individuals, AI brokers, and operators are in fixed communication, exchanging knowledge and fixing beforehand unthinkable issues at lightning velocity.
The change shouldn’t be restricted to the velocity of AI evolution; it’s additionally concerning the magnitude of complexity it’ll introduce. So, what do organizations must do otherwise to organize for this? AI transformation calls for high-speed interconnects each inside and between knowledge facilities—which solely fiber optics can present. Successfully navigating AI requires a basic architectural shift, and Cisco is right here for it.
Maximizing reliability and efficiency for AI knowledge middle connections
With AI investments anticipated to surge to $5.2 trillion by 2030 in response to McKinsey Quarterly, the calls for on networks have by no means been larger. Information middle visitors is exploding, with back-end visitors rising 10x each two years and the fast adoption of 800G and 1.6T speeds as reported by the Dell’Oro Group. AI clusters are scaling at an unprecedented tempo, with the most important fashions now exceeding 100,000 GPU nodes—over 20x bigger than just some years in the past in response to Epoch AI.
Amid such a dramatic shift in knowledge middle infrastructure, reliability isn’t simply vital, it’s important. An evaluation from Meta discovered {that a} single gradual GPU hyperlink or failed community connection can cut back cluster efficiency by as much as 40%, leaving costly GPUs idle and investments stranded. In contrast to conventional IP networking, AI workloads require the very best ranges of hyperlink integrity to make sure steady coaching and inference. This places immense strain on high-speed optics, which should function beneath the extreme circumstances present in in the present day’s AI knowledge facilities.
Cisco is uniquely positioned to satisfy these challenges. We’ve got what we prefer to name a “platform benefit” within the full stack of know-how—ASICs, optics, platforms, software program, safety, and visibility. This permits us to optimize the design and integration of each ingredient to make sure efficiency and reliability for AI environments.
Cisco’s technique for AI infrastructure depends on three core strengths:
Cisco silicon photonics: basis, efficiency, and reliability
- On the coronary heart of our innovation is Cisco silicon photonics, delivering increased efficiency, higher energy effectivity, and elevated reliability. Cisco has over 20 years of high-volume silicon photonics manufacturing and deployment expertise.
Business’s most sturdy stress testing for demanding AI environments
- Leveraging our techniques and optics design experience, we stress check our optics for AI networking parameters. Our testing methodology encompasses optical, electrical, and mechanical design verification; regulatory and requirements compliance; in addition to Cisco and third-party host platform and software program compatibility testing. Rigorous qualification is carried out on the transceiver module stage and on the platform stage of totally loaded switches and routers.
Architectural improvements to cut back energy consumption
- At Cisco Reside this previous June, we demonstrated a 51.2T swap populated with 64 x 800G linear pluggable optics. By shifting power-consuming DSP performance from pluggable optics to the Cisco Silicon One ASIC, we demonstrated a 30% discount in system energy in comparison with conventional retimed optics.
- We’re additionally dedicated to rising applied sciences equivalent to co-packaged optics (CPO) that may additional cut back total SerDes energy consumption required to generate the optical channel as a result of ultra-short electrical traces. Following functionality demonstrations in 2023, our method to this rising know-how/resolution contains leveraging our world provide chain for variety in parts from a number of foundries, optical engines, and geographies.
As well as, Cisco’s globally diversified provide chain leverages a number of sources for silicon photonics wafer fabs, DSPs, lasers, and manufacturing. This ensures well timed entry to vital parts, enhancing provide chain resilience and enabling fast deployments for essentially the most demanding AI knowledge middle infrastructure wants.
Connecting AI knowledge facilities with coherent transport
As AI knowledge middle infrastructure continues to develop, it has created demand for connecting knowledge facilities utilizing coherent transport. AI’s influence on coherent transport demand is undisputable, and Cisco can ship the bandwidth required for this new era of computing. Massive language mannequin (LLM) model updates and better performing GPUs have been launched, terminology equivalent to neocloud supplier and AI factories have develop into more and more mainstream, and new entities are being created worldwide to deal with the infrastructure to create huge AI coaching services.
As well as, a number of service suppliers which have historically supported wholesale bandwidth and fiber connectivity companies have publicly introduced offers to help AI connectivity progress. These connections would particularly handle connectivity between knowledge middle websites—with the AI connectivity alternatives not restricted to hyperscaler websites, but in addition together with enterprise buyer connectivity.
The ramp-up of AI networks and infrastructure is enjoying a component in present 400G coherent pluggable module adoption, in addition to anticipated 800G coherent module adoption. Cignal AI not too long ago acknowledged that progress in latest quarters of coherent pluggable 400G shipments is being pushed by knowledge middle interconnects (DCIs) for AI knowledge facilities, with AI software program being designed to help a number of bodily areas. As hyperscalers and repair suppliers construct out to help AI, Cisco coherent modules are already being deployed in these DCI builds, and clients are projecting growing bandwidth necessities to help these architectures.
Cisco Routed Optical Networking continues to develop and turns into much more vital within the AI period. In line with Heavy Studying, a good portion (58%) of communication service suppliers plan to deploy 800G coherent pluggables between 2026 and 2028. New Cisco coherent modules, together with 100ZR, 400G extremely long-haul (ULH), and 800G ZR/ZR+, are enabling 100% protection of community functions from shorter attain campus and entry functions to long-haul and even subsea, as evidenced by groundbreaking trials.
AARNet (Australia’s Tutorial and Analysis Community) achieved a outstanding milestone with the profitable completion of a area trial showcasing the capabilities of cutting-edge 400G ULH pluggable coherent optics know-how. Carried out over an unbelievable distance of 4,600 km on the Indigo Central subsea cable system, which connects Sydney and Perth, this achievement heralds a brand new period in long-haul optical networking.
GÉANT efficiently accomplished a area trial that achieved 3,403 km of terrestrial transmission, representing one of many longest identified reaches ever demonstrated on a terrestrial community utilizing a 400G QSFP-DD coherent pluggable transceiver with a 24 W energy envelope, making it suitable with all current 400G-capable routers. Beforehand, Arelion held an 800G ZR+ area trial that demonstrated decreased CapEx by 66% and OpEx by 95%. For the reason that integration of Cisco Routed Optical Networking, COLT has reported a shocking 97% power financial savings and 67% financial savings in CapEx.
Cisco is ideally positioned to assist hyperscaler, neocloud, service supplier, and enterprise clients evolve their networks for AI. Our Cisco Silicon One chips are powering the switching/routing platforms for AI coaching and inferencing functions. We’ve got the optics portfolio for 400G and 800G, and evolving to 1.6T. We are able to leverage our platform benefit to cut back energy and enhance latency for AI knowledge facilities with applied sciences like LPO and CPO. And we’re persevering with to advance the function of pluggables to succeed in into extra elements of the community, changing capabilities historically delivered from a devoted chassis—from DWDM transponders with ZR/ZR+ coherent optics to amplifiers, line techniques, conventional PON OLTs, and extra.
Our hyperscale, neocloud, service supplier, and enterprise clients have expressed their confidence in Cisco. That is evidenced by the AI momentum we not too long ago reported in our FY25 earnings, the place we greater than doubled our preliminary goal of AI orders, surpassing $2.1B. We stay up for strengthening our partnership with our clients.
Hear from our Cisco consultants and clients
Register for the Cisco Packet Optical Networking Convention (PONC)—September 17–18 in Dallas, TX (digital or in individual). We’ll discover how operator networks are evolving to satisfy the calls for of AI by benefiting from Routed Optical Networking. You’ll additionally be taught the newest in quantum networking; what’s new in routing, optical techniques, and optics; easy methods to higher handle your community with automation and assurance; and rather more.
You can even go to us at ECOC September 29–October 1 in Copenhagen at sales space #C4136.
Share: