MPN: MCX9-OSFP
NVIDIA ConnectX-9 SuperNIC
800 Gb/s · 1.6 Tb/s per GPU · PCIe Gen6

InfiniBand XDR · 800GbE · Spectrum-X · OSFP · Ultralow-Latency

Specifications

Model

ConnectX-9 SuperNIC

Throughput

800 Gb/s per port
up to 1.6 Tb/s per GPU

Ethernet

25/50/100/200/400/800 GbE

InfiniBand

EDR · HDR · NDR · XDR

Host Interface

PCIe Gen6 x16
(Gen5/4/3/2/1.1 compat)

Protocols

RoCE · RDMA
GPUDirect · NVMe-oF

Fabric

NVIDIA Spectrum-X Ethernet

Form Factor

OSFP

Latency

Ultralow-latency

PAM4 / NRZ

Both supported

Availability

GA (firmware Feb 2026)
Vera Rubin platform

Overview

The NVIDIA ConnectX-9 SuperNIC is the scale-out network adapter of the Rubin generation, delivering up to 1.6 Tb/s of throughput per GPU — double the ConnectX-8 SuperNIC — to keep pace with gigascale AI factories. It supports 800 Gb/s Ethernet and InfiniBand XDR with sub-microsecond latency.

ConnectX-9 is engineered for NVIDIA Spectrum-X Ethernet, accelerating RoCE performance, GPUDirect RDMA, and NVMe-oF for converged AI data center networking. Its PCIe Gen6 x16 host interface is backward-compatible with Gen5/4/3/2/1.1 slots.

As one of the six new chips in the Vera Rubin platform, ConnectX-9 SuperNICs are the fabric that links Rubin GPUs across the NVL72 rack and out to scale-out clusters, and are also deployed with the BlueField-4 DPU for the Inference Context Memory Storage Platform.

Request a Quote — NVIDIA ConnectX-9 SuperNIC

QS Compute — global B2B supply for NVIDIA data center and edge AI hardware. Enterprise procurement, direct manufacturer channels.

Request Quote