FS Launches NVIDIA ConnectX-8 800G SuperNIC for Massive-Scale AI

As AI workloads grow exponentially—from large language models (LLMs) to autonomous systems and real-world robotics—the need for faster, more scalable, and more efficient computing infrastructure has never been greater. To meet these demands, NVIDIA is advancing platform architecture with the introduction of the NVIDIA® ConnectX®-8 SuperNIC, the industry’s first to combine 800G networking and a PCIe Gen6 switch in a single device. This innovation delivers unprecedented throughput, simplifies system design, and enhances both power and cost efficiency.

As an authorized NVIDIA and Mellanox partner since August 2022, FS now offers the NVIDIA® ConnectX®-8 SuperNIC 900-9X81E-00EX-ST0, and complete 800G networking solutions and expert technical support to help enterprises accelerate AI, HPC, and cloud-scale workloads.

ConnectX®-8 SuperNIC Product Overview

The NVIDIA® ConnectX®-8 SuperNIC 900-9X81E-00EX-ST0 is built to accelerate hyperscale AI workloads with up to 800G aggregate bandwidth and support for both InfiniBand (800/400/200/100 Gb/s) and Ethernet (400/200/100/50/25 Gb/s). Featuring up to 16 lanes of PCIe Gen6, it delivers ultra-fast host communication for AI, HPC, and cloud data centers. At the core of NVIDIA’s AI networking portfolio, the ConnectX®-8 integrates seamlessly with next-generation NVIDIA platforms to support trillion-parameter models, advanced AI data systems, and agentic AI applications—delivering high performance, flexibility, and improved power efficiency for future-ready, sustainable GPU-scale infrastructure.

Hardware Features of ConnectX®‑8 SuperNIC

PCIe Gen6 Interface:

PCIe Gen6 x16 interface delivering up to 64GT/s data transfer rate.

Protocol Support:

Compatible with both InfiniBand and Ethernet, offering flexibility across diverse network architectures.

Maximum total bandwidth:

Supporting InfiniBand speed of 800/400/200/100 Gb/s and Ethernet speed of 400/200/100/50/25 Gb/s.

Networking Port:

Single cage OSFP.

Secure Boot:

Supports secure boot and hardware-based encryption to ensure firmware and software integrity.

Software Protocols of ConnectX®‑8 SuperNIC

RDMA / RoCEv2

: Low‑latency, high‑throughput remote direct memory access over InfiniBand and Ethernet.

NVGRE / VXLAN Offload

: Hardware offload for overlay network encapsulation and decapsulation, reducing CPU load in virtualized environments.

SR‑IOV Virtualization

: Isolation and dedicated resources for virtual machines within the server.

CPU Offload

: Reduces CPU overhead by processing network functions directly in hardware.

Advanced Routing & Congestion Control

: Telemetry‑based routing and congestion control to optimize network throughput for AI fabrics.

NVIDIA SHARP™

: In‑network aggregation and reduction acceleration for HPC and large‑scale AI training.

ConnectX®-8 SuperNIC Connectivity Solutions

1.6T Switch-to-Dual 800G ConnectX-8

The Twin port OSFP in the QM3400 and QM3200 switches use two, 4-channel MPO-12/APC optical connectors with two, straight 4-channel fiber cables. Each fiber cable can link to single port 800Gb/ s OSFP RHS form-factor transceiver in two ConnectX-8 adapters.

1.6T Switch-to-Quad 400G ConnectX-8

The Twin port OSFP using two, 1:2 splitter fiber cables create four, 2x200G-PAM4 links or 400G a.k.a XDR400. Each 2-fiber cable connector can link to a single port 800Gb/s OSFP RHS form-factor transceiver in four ConnectX-8 adapters. The 400G rate is achieved by using only 2-fibers at 200G-PAM4 in the 1:2 fiber cable ends into an 800G transceiver and are considered XDR400 rates - not 400G NDR which is 4x100G-PAM4. Power consumption is also reduced in 800G transceivers.

Bottom Line

NVIDIA ConnectX®-8 is redefining what’s possible for PCIe-based systems. By integrating a PCIe Gen6 switch with a high-performance SuperNIC into a single device, ConnectX®-8 streamlines server design, reduces component count, and unlocks the high-bandwidth communication pathways essential for modern AI workloads. This results in a simpler, more power-efficient platform with exceptional scalability and lower total cost of ownership (TCO). Additionally, ConnectX®-8 SuperNICs enable enhanced confidential computing capabilities in multi-GPU platforms, helping build more secure AI infrastructures.

FS provides comprehensive support and supply for these cutting-edge networking solutions, helping enterprises seamlessly build and expand GPU-scale AI infrastructures that meet evolving workload demands with reliability and cost-effectiveness.