DFT is proud to introduce the NVIDIA® ConnectX - 8 900 - 9X81Q - 00CN - ST0, a next - generation 400G SuperNIC that integrates a PCIe Gen6 - capable switch with ultra - high - speed networking in a single device.
Designed for modern AI, HPC, and cloud - scale infrastructures, CX8 delivers exceptional throughput, streamlines server architecture, and enhances both power and cost efficiency.
As an authorized NVIDIA and Mellanox partner since August 2022, DFT provides comprehensive networking solutions and expert technical support to assist enterprises in fully leveraging the performance of their AI and HPC clusters.
Overview of ConnectX - 8 SuperNIC 900 - 9X81Q - 00CN - ST0
The NVIDIA ConnectX - 8 SuperNIC900-9X81Q-00CN-ST0 brings next - generation AI networking performance to modern accelerated compute infrastructures. As part of the ConnectX - 8 portfolio, this model features dual QSFP112 ports, offering up to 400Gb/s per port and a total of 800Gb/s network capacity. This enables high - bandwidth connectivity for large - scale GPU clusters and latency - sensitive workloads. Designed for massive AI fabrics and next - generation data centers, it seamlessly integrates with NVIDIA’s advanced networking platforms, supporting the scalability and robustness required for trillion - parameter model training and high - density GPU deployments.
Specifications
- Form factor: PCIe Half Height, Half Length (HHHL)
- Networking Port: Dual - port QSFP112 InfiniBand and Ethernet
- InfiniBand: NDR/HDR/HDR100/EDR/SDR
- Ethernet (Default): 400/200/100Gb/s Ethernet
- Host interface: PCIe Gen6 ×16 lanes (Gen5 compatible)
- Network interface: Dual - port 400G QSFP112
With improved power efficiency and NVLink - optimized data paths, the ConnectX - 8 900 - 9X81Q - 00CN - ST0 helps build more sustainable, future - ready AI infrastructures operating at a large scale. It also enables advanced routing and telemetry - driven congestion control to maintain predictable performance during peak AI workloads. For InfiniBand environments, ConnectX - 8 SuperNICs extend the benefits of NVIDIA® In - Network Computing, accelerating collective operations and enhancing overall system efficiency in HPC and AI clusters.
Why is the ConnectX - 8 SuperNIC Solution Revolutionary for Server Architecture?
In traditional PCIe switch server designs, CPUs, GPUs, NICs, and DPUs are interconnected through multiple discrete PCIe switches, creating a complex, multi - hop topology. A typical configuration, for example, a dual - CPU server with eight GPUs (such as NVIDIA L40) and five NIC - class devices (four NVIDIA ConnectX - 7 NICs and one NVIDIA BlueField - 3 DPU), requires two to four standalone PCIe switches just to provide basic GPU - to - GPU and GPU - to - NIC connectivity. This increases the component count, board area, and cooling requirements, driving up both system complexity and cost. Architecturally, GPU - to - GPU traffic often traverses two CPU sockets, where host CPU resources and inter - socket links can become bottlenecks, limiting effective bandwidth to around 25 GB/s or less depending on link utilization. Similarly, in common 2:1 GPU - to - NIC topologies, GPU - to - NIC communication is constrained by the available PCIe bandwidth and tightly coupled to the underlying PCIe generation, making it more difficult to scale IO performance in line with rapidly growing GPU compute capabilities.
Advantages of NVIDIA ConnectX - 8 SuperNIC in Server Architecture
1. Integrated PCIe Gen6 Switch
The ConnectX - 8 SuperNIC integrates a 48 - lane PCIe Gen6 switch directly into the network card, eliminating the need for multiple discrete PCIe switches. This consolidation simplifies the motherboard layout, reduces the component count, and lowers system - level costs, creating a cleaner, more scalable architecture for GPU - dense servers.
2. Double Effective Network Bandwidth per GPU
With up to 800 Gb/s networking throughput, ConnectX - 8 provides each GPU with up to 400 Gb/s effective bandwidth in a 2:1 GPU - to - NIC configuration. This eliminates I/O bottlenecks, accelerates data movement between GPUs, NICs, and storage, and enables up to twice the collective communication performance for large - scale AI training workloads.
3. Improves GPU Communication Efficiency
ConnectX - 8 optimizes critical GPU communication paths, including GPU - to - GPU across CPU sockets, GPU - to - NIC communication, and GPU - to - GPU via the same PCIe switch. Compared to traditional architectures, these improvements deliver higher bandwidth, lower latency, and more consistent GPU utilization, accelerating distributed AI and HPC workloads.
4. Streamlined System Design, Improved Airflow and Serviceability
By combining networking and PCIe switching into a single SuperNIC, ConnectX - 8 simplifies PCB layout, improves airflow, and reduces power consumption. This results in a more compact, service - friendly, and energy - efficient server platform, helping system builders deploy high - density AI and HPC infrastructures more effectively.
Bottom Line
Whether building today’s high - speed GPU - to - NIC connections or planning next - generation AI and HPC data centers, the DFT NVIDIA® ConnectX - 8 SuperNIC 900 - 9X81Q - 00CN - ST0 delivers the performance, scalability, and efficiency needed for large - scale, latency - sensitive workloads. DFT also offers a full range of complementary 400G, 800G, and 1.6T optical transceivers to support diverse data center architectures.
