AI Networking
the fabric between GPU servers
Ethernet and InfiniBand switching for AI and HPC fabrics — spine, leaf, and storage back-end — for the GPU servers they connect.
800GbE
Spine port speeds
XDR 800G
InfiniBand per port
Leaf-spine
Rail-optimized designs
AI networking brands we source
Spectrum Ethernet and Quantum InfiniBand switching for AI and HPC fabrics, from 800GbE spine to 100GbE leaf.
Spectrum SN5600SN4600Quantum-X800Quantum-2 QM9700QM8700
- Ports from 200G to 800G for AI and HPC fabrics
- Quantum InfiniBand with in-network computing (SHARP)
Source every layer from one place
Ethernet
Ethernet AI Fabric
Spectrum-X, Nexus 9000, and Arista 7000 switching for GPU-to-GPU east-west traffic over RoCEv2.
InfiniBandNDR 400G · XDR 800G
InfiniBand Fabric
Quantum InfiniBand for multi-node model training and HPC.
Spine400/800GbE
Spine Switching
The backbone layer every leaf switch connects to.
Leaf
Leaf / Rail Switching
The layer that connects the GPUs in each server to the spine.
Capabilities
- Port speeds
- 100 / 200 / 400 / 800GbE
- InfiniBand
- XDR 800G · NDR 400G · HDR 200G
- Fabric
- RoCEv2 · adaptive routing · congestion control
- Topology
- Leaf-spine, rail-optimized GPU clusters
Common deployments
- GPU cluster east-west fabric
- Multi-node model training
- NVMe-oF storage back-end
- HPC & simulation interconnect