AI calls for a elementary rethinking of how we design, construct, and safe knowledge facilities. Organizations are shifting previous the experimentation section and are deploying huge AI clusters that require unprecedented community bandwidth, energy, and safety controls. Constructing these giga-scale environments isn’t nearly including extra GPUs. It requires a holistic, deeply built-in structure that ensures each element—from the silicon to techniques to software program and the working mannequin—works in excellent concord.
Cisco is delivering the foundational infrastructure wanted to make this a actuality. By combining our networking experience with superior silicon, techniques, optics, software program, working fashions, and safety improvements, we’re offering enterprises, neoclouds, and sovereign cloud suppliers with the instruments they should deploy AI securely at scale. Cisco stands out as the one vendor able to delivering a real turnkey answer, seamlessly tailor-made to satisfy the calls for of shoppers at each scale.
By means of our partnership with NVIDIA, we’re pushing the boundaries of what AI networks can obtain, specializing in three essential pillars: infrastructure, networking, and safety.
Constructed for AI factories
The transition to giga-scale AI requires {hardware} able to dealing with huge knowledge throughput with minimal latency. Constructing on the growth of the lately introduced N9300 Silicon One G300 scale-out and 51.2T P200 scale-across techniques, Cisco is elevating the bar to satisfy these intense calls for.
We’re thrilled to introduce the brand new Cisco N9100 Sequence Change, N9164F-NS6, powered by NVIDIA Spectrum-6 silicon. This 102.4T swap delivers an enormous leap in knowledge middle capability to help next-generation safe AI factories. We’re additionally making the N9100 switches accessible with NVIDIA Spectrum-4 silicon, giving enterprises, neoclouds, and sovereign cloud suppliers versatile, high-performance deployment choices.
We imagine within the energy of silicon variety. This technique permits you to select the precise know-how that matches your particular efficiency and operational wants. To maintain implementation easy, we guarantee full reference structure compliance. This streamlines your deployment and ensures clean integration into complicated knowledge middle environments.
Cisco retains elevating the efficiency ceiling, making superior AI infrastructure sooner and simpler to deploy. We launched the N9100 switches, powered by NVIDIA Spectrum-6 Ethernet silicon, to handle the immense scale required by next-generation safe AI factories. This 100% liquid-cooled 102.4T swap represents a serious leap ahead in AI knowledge middle capability.
As well as, the N9100 swap, powered by NVIDIA Spectrum-4 silicon, can also be now accessible, additional increasing deployment choices for enterprises, neoclouds, and sovereign cloud suppliers.
To make deploying these complicated techniques simpler, we built-in Nexus One help with cloud-managed Cisco Nexus Hyperfabric. This creates a turnkey, full-stack AI answer. AI builders not need to piece collectively disparate elements and hope they perform effectively. Nexus Hyperfabric additionally now manages the newly accessible Cisco N9164E-NS4-O, powered by NVIDIA Spectrum-4 Ethernet silicon. It provides pods of plug-and-play knowledge middle materials, managed completely by means of the cloud, dramatically decreasing the time it takes to go from procurement to full operation. Nexus One features a strong, on-premises managed Nexus dashboard that could be a confirmed operations, automation knowledge middle, and AI networking answer.
Flexibility stays essential when designing these environments. Organizations have totally different wants based mostly on their measurement, safety necessities, and present infrastructure. To accommodate this, we provide strong reference architectures tailor-made to particular deployment varieties:
- Cisco Cloud Reference Structure (CRA): For enterprises and neoclouds deploying specialised AI servers with higher than 1000 GPUs and as much as 32,000 GPUs, the Cisco CRA supplies a extremely optimized, scalable path ahead utilizing industry-leading Cisco Silicon One and Cisco N9300 Sequence Switches. For deployments of fewer than 1000 GPUs, now we have ready-to-consume Cisco Enterprise Reference Structure (ERA).
- Compliant with NVIDIA Cloud Companion Reference Design: For giant-scale AI infrastructures starting from 1000 to 32,000 GPUs, the Cisco N9100 Sequence Switches totally adjust to the NVIDIA Cloud Companion (NCP) Reference Structure. This ensures final efficiency for sovereign and neocloud deployments, using scale-out materials powered by Spectrum-X Ethernet swap silicon.
By providing a selection between NCP Reference Design and Cisco CRA architectures, we be certain that each buyer has a confirmed, validated blueprint for achievement, whether or not they’re constructing an enormous sovereign cloud or a extremely focused enterprise AI cluster.
Securing the AI workload on the edge
As AI clusters develop in energy and complexity, they develop into extremely enticing targets for malicious actors. Conventional perimeter safety fashions fail in these environments. Unprotected east-west visitors inside the AI material permits lateral threats to unfold quickly. A compromised AI workload may result in GPU useful resource hijacking or huge knowledge exfiltration, leading to a full lateral blast radius. Nevertheless, operating conventional safety brokers instantly on the AI servers taxes the CPU and GPU, draining the very compute assets you’re making an attempt to maximise.
To unravel this, we’re basically shifting the place and the way safety is enforced. We prolonged Cisco Hybrid Mesh Firewall help on to NVIDIA BlueField DPUs.
This innovation brings safety as near the workload as doable with out compromising efficiency. By embedding the firewall on the DPU, we ship built-in inline safety with high-performance scalability. Directors can outline safety insurance policies as soon as and implement them in every single place, isolating digital public clouds (VPCs) and blocking lateral assaults in actual time. This supplies choke-point-free enforcement, defending front-end material servers whereas leaving the host CPU and GPU completely devoted to processing AI workloads.
To study extra about how we’re defending AI environments from lateral threats with out sacrificing efficiency, learn our detailed technical breakdown: Cisco secures AI infrastructure with NVIDIA BlueField DPUs.
Advancing AI networking for peak efficiency
Even essentially the most highly effective GPUs will sit idle if the community can’t feed them knowledge quick sufficient. Maximizing GPU utilization and optimizing the key-value (KV) cache efficiency requires clever, high-speed connectivity throughout your entire material. That is the place Cisco Nexus One basically adjustments the sport for AI networking.
Nexus One supplies a unified administration aircraft throughout each NX-OS and SONiC environments with an on-premises managed Nexus Dashboard and cloud-managed Nexus Hyperfabric as operational fashions.
Nexus Dashboard delivers unprecedented visibility into the full-stack AI setting with Cisco N9000 Sequence Switches. Directors acquire entry to deep AI job-monitoring capabilities with Nexus Dashboard 4.2. Monitor precisely how knowledge strikes by means of the material, figuring out bottlenecks earlier than they influence coaching occasions. The system options clever, auto-adjusting load balancing and telemetry-based congestion management. As an alternative of counting on static routing protocols that break down underneath the distinctive, elephant-flow visitors patterns of AI workloads, the community dynamically adjusts to make sure optimum knowledge supply.
Moreover, Nexus Dashboard allows real-time GPU well being monitoring. Community operators can see instantly into the efficiency metrics of the compute layer, making certain that each costly GPU useful resource operates at most effectivity. Whether or not you’re constructing a scale-out community inside a single knowledge middle or a scale-across community connecting a number of amenities with high-performance, low-latency hyperlinks, Nexus Dashboard ensures the community acts as a strong accelerator somewhat than a bottleneck.
Empowering the following technology of innovation
The AI revolution requires extra than simply uncooked processing energy. It calls for a completely built-in ecosystem the place infrastructure, networking, and safety are designed to function as a single, cohesive unit.
By means of our continued improvements with the Spectrum-X-powered Cisco N9100 Sequence Switches, the clever administration of Nexus One, and the edge-enforced safety of the Cisco Hybrid Mesh Firewall on BlueField DPUs, Cisco supplies the inspiration that enterprises, neoclouds, and sovereign clouds must scale their AI ambitions. We’re constructing the networks that may energy the following decade of discovery, making certain they’re quick, dependable, and basically safe.
