As AI workloads continue to scale, enterprises are quickly realizing that traditional networking isn’t built for the demands of modern, distributed GPU clusters. To address this, NVIDIA and Aviz Networks hosted a joint bootcamp showcasing the deep integration between NVIDIA Spectrum-X — the industry’s first Ethernet fabric optimized for AI — and the Aviz Open Networking Enterprise Suite (ONES), purpose-built for AI infrastructure orchestration and observability.
"AI had its iPhone moment with ChatGPT. Suddenly, enterprises everywhere wanted to deploy generative AI at scale — but Ethernet couldn’t keep up."
How is Ethernet evolving for AI clusters?
From InfiniBand to Ethernet: The AI Networking Shift
Enter Spectrum-X. NVIDIA took key capabilities from InfiniBand and extended them to Ethernet — enabling RDMA, adaptive routing, and congestion control, all with the governance and familiarity enterprises demand from Ethernet environments.
"The larger the AI cluster, the bigger the impact the network has on performance, Spectrum-X gives enterprises a purpose-built Ethernet fabric to unlock full GPU performance.
What is Spectrum-X RA 1.3.0 and how is it validated?
Spectrum-X RA 1.3.0: Validated at Supercomputer Scale
Aranga Madipuri, Product Manager at NVIDIA, detailed the Spectrum-X Reference Architecture (RA 1.3.0), tested on real-world supercomputers like Israel-1. The RA offers a prescriptive blueprint combining SONiC/Cumulus NOS, NetQ telemetry, NVIDIA AIR digital twin simulation, and BlueField-accelerated switching — ensuring performance, reliability, and reproducibility for massive AI clusters.
How does Aviz ONES support Spectrum-X deployments?
ONES by Aviz: Purpose-Built for Multi-Tenant AI Infrastructure
Aviz CTO Chit Perumal and Principal Engineer Kasi Nath demonstrated how ONES seamlessly integrates with Spectrum-X RA 1.3.0. ONES delivers:
- Day 0–2 automation: Declarative fabric design, NVIDIA AIR simulation, automated switch and host configuration using RA-aligned templates.
- Multi-tenancy orchestration: EVPN and VRF-based segmentation, GPU-aware resource provisioning, and policy-driven isolation across workloads.
- Telemetry and alerting: Agentless, real-time visibility from switches, hosts, and GPUs — plus built-in alerting integrated with Slack, ServiceNow, and Zabbix.
- Lifecycle operations: Config drift detection, backup/restore, structured RMA workflows, and topology-aware fabric comparison.
"We wanted customers to scale GPU clusters effortlessly while maintaining network visibility and operational simplicity — ONES makes that possible."
What was covered in the live demo?
Real-World Topologies, Live Demo
The session closed with a detailed demo covering:
- Automated orchestration of a two-SU Spectrum-X fabric
- Tenant creation and GPU assignment
- Policy-driven isolation validation
- Real-time monitoring dashboards and anomaly detection
- Full config comparison and RMA workflows
Explore More
- Watch the full bootcamp recording
- Learn about Aviz ONES
Whether you’re building a private AI cloud or launching GPU-as-a-Service, the NVIDIA + Aviz stack gives you the tools to scale with confidence — and visibility.
Frequently Asked Questions
1. What is NVIDIA Spectrum-X and how is it optimized for AI workloads?
Spectrum-X is the first Ethernet fabric designed specifically for AI clusters. It extends InfiniBand’s low-latency, lossless transport to Ethernet, delivering RDMA, adaptive routing, and congestion control — all with familiar enterprise Ethernet governance.
2. How does Spectrum-X improve performance compared to traditional Ethernet?
Traditional Ethernet often struggles with network congestion and packet loss during large-scale GPU training. Spectrum-X solves this by:
- Enabling RDMA over Converged Ethernet (RoCE) for direct GPU communication.
- Using adaptive routing to bypass congestion.
- Delivering consistent throughput across thousands of GPUs.
3. What is the Spectrum-X Reference Architecture (RA 1.3.0)?
It’s a validated deployment blueprint tested on real supercomputers like Israel-1. RA 1.3.0 combines:
- Open NOS (SONiC/Cumulus).
- NetQ telemetry.
- NVIDIA AIR for digital twin simulation.
- BlueField DPUs for hardware acceleration.
This ensures predictable performance and easier scaling of AI networks.
4. How does Aviz ONES integrate with Spectrum-X?
ONES is a software layer for orchestration and observability. It connects directly with Spectrum-X fabrics to automate deployment, manage multi-tenant AI workloads, and deliver agentless, real-time telemetry — reducing operational complexity.
5. What Day 0–2 automation capabilities does ONES offer?
- Declarative fabric design: Define your network layout upfront.
- NVIDIA AIR simulation: Test configurations in a digital twin before deployment.
- Zero-touch provisioning: Automatically configures switches and hosts using Spectrum-X RA templates.
6. How does ONES enable multi-tenant isolation for AI infrastructure?
ONES uses EVPN and VRF-based segmentation to isolate traffic between tenants. It also provisions GPU resources intelligently and applies policies to guarantee secure workload separation.
7. How does ONES deliver real-time visibility without extra agents?
ONES leverages built-in telemetry from the NOS, hosts, and GPUs. It integrates with tools like Slack, ServiceNow, and Zabbix for automated alerting — without installing third-party agents that consume resources.
8. What practical workflows did the bootcamp demo cover?
- Orchestration of a two-switch Spectrum-X fabric
- Tenant creation and GPU assignment
- Policy validation for isolation
- Real-time monitoring dashboards and anomaly detection
- Config comparison and structured RMA workflows.
9. Who should adopt Spectrum-X with Aviz ONES?
Organizations running large AI training clusters, GPU-as-a-Service providers, and any enterprise building private AI clouds that require high performance, robust isolation, and full-stack observability.
10. Where can I learn more or see this solution in action?
Watch the bootcamp recording for a full walkthrough and explore Aviz ONES for detailed docs, case studies, and a deeper look at deployment best practices.
