Collection: Networking

The Interconnect Layer Is Where AI Clusters Win or Lose

Here's a number that changes how you think about networking: in a distributed AI training job, up to 40% of total training time can be spent on gradient synchronization across nodes. Not compute. Not storage I/O. Network. The bandwidth, latency, and topology of your interconnect fabric directly determines how efficiently your GPU cluster trains — and how much of your hardware investment you're actually using.

This is why DVUN treats networking as a first-class infrastructure discipline, not an afterthought. Our Networking collection covers the full physical layer of AI cluster interconnect: high-radix switches for spine-leaf topologies, high-speed NICs for GPU-to-GPU communication, optical transceivers for long-reach and high-density cabling, DAC and AOC cables for short-reach high-bandwidth links, and the accessories that hold it all together reliably.

Whether you're building a 4-node training cluster or a 200-node private AI cloud, the networking decisions you make here will define your system's scalability ceiling and your operational complexity for years to come.

Core Capabilities Across the DVUN Networking Collection

• 100G / 200G / 400G / 800G Switch Fabrics: Spine and leaf switching platforms for AI cluster topologies, with RDMA and RoCEv2 support for GPU-direct communication.
• High-Speed NICs: 25GbE to 400GbE network interface cards with RDMA offload, SR-IOV, and GPUDirect RDMA support for low-latency GPU cluster networking.
• Optical Transceivers: QSFP28, QSFP-DD, and OSFP modules for 100G to 800G links, with verified compatibility across major switch and NIC platforms.
• DAC & AOC Cables: Direct attach copper and active optical cables for short-reach, high-density rack and inter-rack connections.
• Network Accessories: SFP adapters, fiber patch panels, cable management, and diagnostic tools to complete your network build.

Two Architectures, One Collection

For Training Clusters: High-bandwidth, low-latency all-to-all communication is the priority. Our 400G switch platforms and RDMA-capable NICs are selected for exactly this topology. See our Switches and NICs for the core components of a training fabric.

For Inference Infrastructure: Throughput and connection density matter more than all-to-all latency. Our 100G and 25G switching and NIC options provide cost-effective, high-density connectivity for inference serving clusters. Pair with our AI Storage Nodes for a complete inference infrastructure build.

Specifications Overview

• Switch port speeds: 25GbE, 100GbE, 200GbE, 400GbE, 800GbE
• NIC speeds: 10GbE to 400GbE per port
• Optical module form factors: SFP+, QSFP28, QSFP-DD, OSFP
• Cable types: DAC (passive/active), AOC, MPO trunk cables
• Protocol support: Ethernet, InfiniBand, RoCEv2, RDMA
• Management: OpenConfig, SONIC, and vendor NOS options on switch platforms

Network Fast. Scale Further.

Your GPU cluster is only as fast as the network connecting it. DVUN's networking collection gives you the hardware to build an interconnect fabric that matches the ambition of your compute investment. Request a quote for full cluster networking builds, or browse individual subcategories to find the specific components you need.