Few transformational technologies have generated the kind of buzz as Artificial Intelligence. Organizations around the world are finding exciting new ways to collaborate, manage applications, enhance their digital security, and reimagine customer experiences. Cisco already harnesses AI-powered capabilities across our product and customer service portfolio. We are also leading innovation that enables AI infrastructure. Our vision for AI Network fabrics offers unparalleled programmability, performance, flexibility, and efficiency. And we enable these fabrics with a familiar technology that has broad ecosystem support — Ethernet.
At the foundation of every AI generated product capability will be ground-breaking infrastructure. GPUs and massively scalable data stores and compute will be deployed in AI/ML platforms of unprecedented scale and performance. This massive movement and processing of data will need a network fabric that meets the considerable demands of these platforms. These networks need to enable AI with economical, flexible, and innovative connectivity. An economical and flexible platform enables you to invest early and evolve in a way that fits your unique business model. Innovation offers you a fabric that is high throughput, low-latency and lossless — without sacrificing flexibility. As a result of Cisco’s considerable investments, that AI fabric is already here. And we believe that this fabric needs to be built on Ethernet.
AI Fabrics will be Ethernet-based
Ethernet offers advantages of economy and scale. As transformational technologies approach maturity, customers tend to move away from sole-source components towards multi-source components, to improve purchasing power. In the case of AI/ML infrastructure, this means a move away from single-source connectors, fabric, and GPUs. Moving towards multi-source components typically means using a broadly adopted standard, and it is difficult to find a more broadly deployed data center network technology than Ethernet. This increases the ability to find the right ROI with interoperability within component types, as well as between component types. Additionally, familiar and consistent technology for all data center fabrics produces efficiencies in training, procurement, and support. Lastly, Ethernet is well-placed to support the massive scale that the world’s largest AI network fabrics require, with a broad base of suppliers.
Every organization’s AI journey is unique and requires agile technology and the ability to scale on your terms. Whether you are retraining models, deploying model inference, or building your own large language model, you need infrastructure that keeps pace with your demands. Moving quickly means leveraging the skills that you have and deploying familiar technology. Ethernet is the most deployed infrastructure and has the broadest architectural support. You can find your choice of partners, suppliers, and architects to help your unique organizational goals.
AI/ML infrastructure will need to be set up for innovation that drives greater capabilities and performance. Ethernet has a proven history of disruptive innovation. In the approximately 20-year period following 1998, Ethernet standard speeds grew from 1Gbps to 400Gbps, with 800G options available in systems such as the Cisco 8111-EH and the Cisco Nexus 9232E switch. Ethernet’s versatility has been proven in data center networks through protocols such as iSCSI and RoCE; along with several innovations that reduce the need for separate types of network connectivity.
This impressive rate of growth, combined with flexibility, has made it possible to connect several generations of front-end and back-end components. Storage networks that once required specialized back-end fabrics can now connect resources with Ethernet through advances in IP based storage solutions. Similarly, resources requiring InfiniBand connectivity can be deployed with RoCE. This means that back-end networks for advanced compute or GPUs can now be consolidated with a single network fabric.
Cisco’s unique value for Ethernet fabrics
Cisco expands the value that you can achieve from Ethernet by advancing the economics, agility, choice and innovation even further. We make it possible to standardize switching and routing with a single silicon architecture — Cisco Silicon One. This gives you the option of reducing the complexity of your network architectures by managing one experience across the entire network, across all network functions. To maximize economic benefits, Cisco produces a full range of silicon, systems and optics in a way that transfers economies of scale to the end user. We build silicon with Customer Owned Tooling (COT), rather than using traditional ASIC models, or off shelf components, so that we can eliminate the premiums introduced when adding additional vendors to a value chain — and drive end to end innovations.
Cisco offers you the choice of procuring components, white boxes, or fully built systems in your AI Network fabric. Does your business model drive value by building systems and software on top of silicon components? Or do you want ease of deployment by having fully built systems arrive ready for use? Alternatively, you may be in the middle — needing a white box shipped to you with the option to customize software that is unique to your own platform. Whichever model you need, you can partner with Cisco.
Our development of silicon offers differentiated capabilities for Ethernet that truly unlock the power of AI Network fabrics. With one architecture, you can build three deployment models that offer increasing degrees of efficiency for precious system resources through multipathing. You may choose to use standard Ethernet for the widest interoperability solution, or an advanced scheduled Ethernet that halves the Job Completion Time (JCT) compared to standard Ethernet, or the middle ground with enhanced Ethernet taking the best from standard and scheduled Ethernet. Because of the innovation and power of Cisco Silicon One, you can confidently use the same network for high throughput, lossless training fabrics, and ultra-low latency inference fabrics.
Ethernet is ready for the next technology transformation. Whether you are fine-tuning your own models, running off-the-shelf models for inference, or building your own large language models, you can be assured that you have a fabric that meets your needs. Cisco will partner with you so that you can build your AI infrastructure in a way that matches your business model.
Learn more about building AI/ML networks and review
the latest innovations from Cisco Silicon One: