Specifications
Product
NVIDIA ConnectX-9 SuperNIC
Throughput
Up to 1,600 Gb/s (1.6 Tb/s) per GPU
Host Interface
PCI Express Gen 6 x16 (backward compatible to Gen 5/4/3/2/1.1)
InfiniBand Rates
EDR, HDR100, HDR, NDR200, NDR, XDR (1600G)
Ethernet Rates
25 / 50 / 100 / 200 / 400 / 800 / 1600 GbE
Modulation
NRZ and PAM4 (force mode supported)
Cage / Form Factor
OSFP224 host cages; RHS and IHS cable options
Protocol Acceleration
RoCE, RDMA, GPUDirect RDMA, ASAP2, programmable congestion control
Fabric
NVIDIA Quantum-X800 InfiniBand and Spectrum-X Ethernet
Platform
Vera Rubin NVL144 / Kyber-class AI factories
Matched Cable
980-9IAM1-00X001 — 1,600 Gb/s OSFP active copper cable, 1.1 m, RHS–RHS
Matched Splitter
980-9IAO5-00X001 — 1,600 Gb/s to 2 × 800 Gb/s OSFP active copper splitter
Firmware
Release train v82.48.1000 (February 2026 GA)
Security
Secure boot, hardware root of trust, line-rate IPsec/MACsec
Telemetry
In-band and out-of-band telemetry, per-queue congestion counters
Target Workload
Front-end and east-west traffic for gigascale AI training and inference
Ecosystem
DOCA, NVIDIA Networking stack, Spectrum-X Ethernet
Availability
Volume / P-Rel — quote for allocation
Overview
The NVIDIA ConnectX-9 SuperNIC is the networking endpoint of the Vera Rubin generation. It lifts the per-GPU network interface to 1.6 Tb/s and moves the host link to PCIe Gen 6, so a single accelerator no longer has to share a 800G fabric path. That matters because training and large-scale inference are both limited by how fast gradients, KV cache and weights move between GPUs — not by local FLOPS.
ConnectX-9 is delivered as part of a co-designed stack: the Quantum-X800 InfiniBand fabric, Spectrum-X Ethernet, BlueField-4 DPU for infrastructure offload, and OSFP224 optical or copper media. NVIDIA pairs multiple ConnectX-9 devices per GPU on the Rubin platform to reach the 1.6 Tb/s figure, so system design starts with the cage layout and the cable variant (RHS vs IHS, straight vs splitter).
QS Compute supplies ConnectX-9 SuperNICs and the matched 980-9IAxx cable portfolio for AI factory builds, together with the Quantum-X800 and Spectrum-6 switches and the 1.6T OSFP224 optics they terminate into.
Key Benefits
1.6 Tb/s per GPU removes the network as the bottleneck for frontier model training and long-context inference. PCIe Gen 6 doubles host bandwidth so the NIC is not starved by the bus. Dual-protocol operation lets one card run InfiniBand XDR or Ethernet up to 1600GbE, so you can standardise on a single NIC SKU across training and storage fabrics. RoCE + GPUDirect RDMA keep traffic out of host memory, and the matched 980-9IAxx active copper cables cap the total cost per 1.6T link.
Applications
Gigascale AI training clusters and Vera Rubin NVL144 deployments; inference fleets running disaggregated prefill/decode with KV-cache transfer over the fabric; AI factory storage front-ends; and research HPC clusters upgrading from 400G NDR to 1.6T XDR.
Request a Quote — NVIDIA CONNECTX-9 SUPERNIC — 1.6 TB/S AI FACTORY NETWORK INTERFACE
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →