Cisco and Aviz Networks outline Cisco 8000 AI fabric lossless design
A Cisco and Aviz Networks bootcamp focused on designing lossless, high-performance AI fabrics using Cisco 8000 switching with Silicon One and SONiC, alongside RoCE v2 with PFC and ECN. The content matters to enterprise infrastructure teams managing GPU cluster networking and congestion control.
Research Overview
The session framed data centers moving from CPU-centric to GPU-centric architectures and highlighted network design as a requirement for large AI and LLM workloads. It positioned lossless behavior, low latency, and congestion management as core elements for fabric operation under heavy traffic.
The bootcamp paired Cisco’s switching and software stack discussion with Aviz ONES telemetry and observability capabilities. It also described validated leaf-spine approaches and blueprints for scaling AI cluster topologies.
Key Findings
Cisco’s presentation connected AI fabric requirements to specific Ethernet-based mechanisms including RDMA over Converged Ethernet (RoCE v2), Priority Flow Control (PFC), and Explicit Congestion Notification (ECN). The goal presented was to keep latency low and support lossless traffic behavior for GPU workloads.
Aviz’s portion described ONES as providing multi-vendor observability and real-time visibility into congestion points and queue health. It also emphasized monitoring and validating QoS policies across environments that include multiple network platforms.
Technical Breakdown
The bootcamp outlined RDMA as a method for transferring data between server memory without CPU involvement, aimed at reducing latency and increasing throughput for GPU workloads. It stated that RoCE v2 enables RDMA over Ethernet networks.
For congestion behavior, the session described ECN as marking packets experiencing congestion instead of dropping them, with end devices adjusting rates based on feedback. It also described PFC as offering lossless service for selected traffic classes, with an emphasis on preserving RDMA traffic and supporting mixed traffic alongside ECN.
Product Update and Architecture Details
Satish Surani, Director of Product Management at Cisco, discussed Cisco 8000 switching with a networking stack built around Silicon One and SONiC. The session highlighted Silicon One G200 capabilities including ultra-low latency, a stated 51.2 Tbps capacity described as high radix, and fully shared packet buffers for burst absorption.
On software, it characterized SONiC as an open, modular NOS that customers can customize, with Cisco 8000 switches running SONiC with enterprise support. Aviz ONES was described as an agentless observability and orchestration layer that collects multi-vendor telemetry, monitors PFC, ECN, and RDMA health, and visualizes AI traffic versus non-AI traffic.
Operational Impact
The bootcamp covered leaf-spine topology practices aimed at keeping the network non-blocking and scalable for AI clusters. It stated that congestion management and open, programmable fabrics are part of achieving predictable performance as clusters grow.
It also described a factory topology demo using Cisco 8100 switches configured for PFC and ECN, paired with a non-blocking bandwidth setup. The demo was presented as showing real-time visibility into congestion points and queue health via Aviz ONES.
Overall, the bootcamp recap connected Cisco 8000 plus Silicon One and SONiC with RoCE v2, PFC, and ECN for lossless AI traffic, and paired that stack with Aviz ONES for telemetry-driven visibility and troubleshooting support. This “Blog Signals brief” is a fact-based summary of the vendor blog.