What Is ConnectX?_
ConnectX is NVIDIA's family of network interface cards that connect GPU servers to the cluster fabric, supporting both InfiniBand and Ethernet at speeds up to 400 Gb/s per port (ConnectX-7) and 800 Gb/s in newer generations. ConnectX adapters provide the RDMA and GPUDirect capabilities that distributed AI training relies on, and they terminate the fiber or copper links that carry the back-end network.
Technical Details
ConnectX adapters implement RDMA in hardware and support GPUDirect RDMA for direct GPU-to-network transfers. They operate in either InfiniBand or Ethernet (including RoCE) mode, giving operators flexibility in fabric choice. Recent generations — ConnectX-7 at 400 Gb/s and ConnectX-8 at 800 Gb/s — use QSFP-DD or OSFP cages and pair with DAC, AEC, AOC, or transceiver-and-fiber connections depending on distance. In a GPU node, the number and placement of ConnectX adapters define the per-node network bandwidth and the PCIe topology relative to the GPUs, both of which affect collective performance. As the termination point for the back-end fabric, the adapter's link quality is verified during network certification.
How Leviathan Systems Works with ConnectX
Leviathan Systems connects and certifies the links into ConnectX adapters during cluster build-out, selecting the right DAC, AEC, AOC, or transceiver option for each run and validating every port.
Related service
Network Testing & Commissioning →