Nexus 9000 (AI Fabric)
The switch that builds lossless AI backend networks on standard Ethernet — no InfiniBand required.
A high-density 800G leaf/spine switch powered by the Cisco Silicon One G200 ASIC. RoCEv2 lossless transport (PFC · ECN) and dynamic load balancing reliably handle east-west traffic between GPU clusters.
What Is the Nexus 9000 AI Fabric
In AI training and inference clusters, GPU-to-GPU (east-west) communication determines total job time. The Nexus 9000 is a high-density 800G switch powered by the Cisco Silicon One G200 ASIC that implements RoCEv2 lossless transport on standard Ethernet instead of costly InfiniBand — building a backend network where GPUs never sit idle. You can construct an AI fabric while reusing your existing Ethernet operational knowledge and tooling.
Key Capabilities
800G high-density switching
64× 800GbE · 51.2 Tbps forwarding (2RU) connects large GPU clusters with just a few switches
RoCEv2 lossless transport
PFC and ECN congestion control guarantee packet-loss-free GPU-to-GPU communication
Dynamic load balancing (DLB)
Dynamically distributes elephant flows to prevent link skew and congestion
Deep buffers & telemetry
On-die packet buffers absorb bursts, with real-time visibility via Nexus Dashboard Insights
Key Specifications
| Item | Specification |
|---|---|
| ASIC | Cisco Silicon One G200 |
| Ports & capacity | 64× 800GbE · 51.2 Tbps (2RU) |
| Lossless transport | RoCEv2 (PFC · ECN) · dynamic load balancing (DLB) |
| Buffer | 256MB on-die packet buffer |
| Latency & operations | ~1.5μs low latency · Nexus Dashboard Insights |
What It's Used For
GPU cluster backend fabric
A leaf-spine AI backend network connecting many GPU servers losslessly
High-performance storage networks
A high-bandwidth storage fabric for training data and checkpoint I/O
InfiniBand replacement
For organizations that want to build and operate AI fabrics with standard Ethernet operations