NICs

The Last Mile Between Your GPU and the Network

A network interface card is a small piece of hardware with an outsized impact on AI cluster performance. It's the component that determines whether your GPU can communicate directly with the network fabric — bypassing the CPU entirely — or whether every packet has to take a detour through system memory and add latency that compounds across thousands of collective operations per training step.

GPUDirect RDMA. RoCEv2. SR-IOV. These aren't marketing terms — they're the technical capabilities that separate a NIC that enables high-performance distributed AI training from one that becomes a hidden bottleneck. DVUN's NIC collection is curated specifically for AI infrastructure use cases, with every card evaluated for its RDMA performance, driver maturity, and compatibility with the GPU platforms and switch fabrics it will be paired with.

Six Capabilities That Define a High-Performance AI NIC

  • GPUDirect RDMA: Enables direct memory access between the NIC and GPU memory, eliminating CPU involvement in data transfers and dramatically reducing latency for distributed training collectives.
  • RoCEv2 Protocol Support: RDMA over Converged Ethernet v2 delivers InfiniBand-class latency over standard Ethernet infrastructure — the foundation of modern AI cluster networking.
  • SR-IOV Virtualization: Single Root I/O Virtualization allows a single physical NIC to present multiple virtual functions, enabling efficient multi-tenant inference serving environments.
  • Hardware Offload Engines: Checksum, segmentation, and encryption offload reduce CPU overhead and free up cores for AI workload processing.
  • High Port Density Options: Dual-port and quad-port configurations for servers that need multiple high-speed network connections without consuming additional PCIe slots.
  • Mature Linux Driver Support: Every NIC in our catalog has verified, stable driver support for Ubuntu, RHEL, and Rocky Linux — the distributions AI teams actually run.

Matching the Right NIC to Your Workload

For Distributed Training: You need GPUDirect RDMA and RoCEv2 support, full stop. Our 100GbE and 400GbE RDMA NICs are validated for use with major GPU platforms and pair directly with our RoCEv2-capable switches for a complete lossless fabric.

For Inference Serving: High connection counts and low per-connection overhead matter more than raw bandwidth. Our 25GbE and 100GbE NICs with SR-IOV support are optimized for multi-tenant inference environments where you're handling thousands of concurrent API requests. See our Ready Systems for pre-configured inference node builds.

NIC Specifications Overview

  • Port speeds: 10GbE, 25GbE, 100GbE, 200GbE, 400GbE
  • Interface: PCIe Gen3, Gen4, Gen5 x8 and x16
  • RDMA protocols: RoCEv2, iWARP on select models
  • GPUDirect RDMA: supported on 100GbE+ models
  • SR-IOV: up to 256 virtual functions on select models
  • Form factors: standard height, low profile, OCP 3.0 mezzanine
  • OS support: Ubuntu 20.04/22.04, RHEL 8/9, Rocky Linux, Windows Server

Connect Your Cluster at the Speed It Deserves

The right NIC doesn't just connect your server to the network — it connects your GPU to every other GPU in the cluster, at the speed and latency that makes distributed training actually efficient. Request a quote for volume NIC orders, or reach out to our team for a compatibility check against your specific GPU server and switch platform.

No products found
Use fewer filters or remove all