NVIDIA ConnectX-7 400Gb/s Smart Network Adapter
The NVIDIA ConnectX-7 is a high-performance Smart Network Adapter designed for modern data centers, artificial intelligence, high-performance computing (HPC), hyperscale cloud, storage, and enterprise networking environments.
Supporting up to 400Gb/s total network bandwidth, ConnectX-7 provides both InfiniBand and Ethernet connectivity, depending on the specific adapter configuration. It supports advanced networking technologies including RDMA, RoCE, GPUDirect RDMA, GPUDirect Storage, SR-IOV, hardware offloads, NVIDIA In-Network Computing, and accelerated security functions.
The adapter uses a PCI Express Gen 5 host interface and is available in several physical configurations, including PCIe HHHL/FHHL and OCP 3.0 form factors. Depending on the exact model, ConnectX-7 cards can provide single-, dual-, or multi-port connectivity and support speeds from 10GbE up to 400GbE.
Description
The NVIDIA ConnectX-7 family represents a generation of high-speed network adapters engineered to remove networking bottlenecks in GPU-accelerated computing and data-intensive infrastructure.
With support for 400Gb/s networking, ConnectX-7 is particularly suited to AI training clusters, distributed HPC systems, large-scale storage environments, cloud infrastructure, and high-performance enterprise networks where extremely high throughput and low latency are required.
High-Speed Ethernet and InfiniBand
ConnectX-7 supports both major high-performance networking ecosystems:
- InfiniBand NDR up to 400Gb/s
- InfiniBand HDR up to 200Gb/s
- Ethernet up to 400GbE
- 200GbE
- 100GbE
- 50GbE
- 40GbE
- 25GbE
- 10GbE
Specific supported speeds depend on the adapter OPN, port configuration, transceiver/cable, and network infrastructure.
PCIe Gen 5 Connectivity
The ConnectX-7 platform supports PCIe Gen 5, providing a high-bandwidth host connection for servers equipped with modern PCIe infrastructure. Current NVIDIA documentation specifies PCIe Gen 5/Gen 4 operation through an x16 interface on supported adapter configurations.
RDMA and Low-Latency Networking
Hardware-accelerated Remote Direct Memory Access (RDMA) allows data to be transferred between systems with reduced CPU involvement and lower latency. This is particularly valuable for distributed AI, HPC, storage, and clustered applications.
For Ethernet deployments, ConnectX-7 supports RoCE (RDMA over Converged Ethernet), including advanced RoCE acceleration features.
AI and GPU-Accelerated Computing
ConnectX-7 is designed to work closely with NVIDIA GPU infrastructure. Technologies such as GPUDirect RDMA and GPUDirect Storage can help accelerate data movement between networking, GPUs, and storage systems while reducing unnecessary CPU data copies.
This makes ConnectX-7 especially appropriate for distributed AI training and inference clusters, scientific computing, and GPU-accelerated data centers.
Smart Networking and Hardware Offload
The adapter provides hardware acceleration for networking, storage, virtualization, and security workloads. Features include SR-IOV, overlay networking acceleration, programmable packet processing, TCP/IP offloads, QoS, flow processing, and accelerated security functions.
Port Flexibility
Certain ConnectX-7 models support port splitting, allowing a high-speed physical interface to be divided into multiple logical network ports. For example, NVIDIA documents a ConnectX-7 configuration capable of operating as one 400GbE port or four 100GbE ports, depending on the specific OPN and configuration.
3. Complete Technical Specification
| Specification | Details |
|---|---|
| Product Family | NVIDIA ConnectX-7 |
| Product Type | Smart Network Adapter / Host Channel Adapter |
| Primary Applications | AI, HPC, cloud, storage, enterprise networking, hyperscale data centers |
| Maximum Total Bandwidth | Up to 400Gb/s |
| InfiniBand | Up to NDR 400Gb/s |
| InfiniBand Standards | IBTA Specification 1.5 |
| Ethernet | Up to 400GbE |
| Ethernet Speeds | 10/25/40/50/100/200/400GbE, depending on model |
| Host Interface | PCI Express Gen 5 |
| PCIe Interface | Up to x16 on standard PCIe adapter configurations; some Socket Direct configurations support additional PCIe connectivity |
| PCIe Compatibility | PCIe Gen 5 / Gen 4 / Gen 3 and other backward-compatible configurations depending on card |
| Network Connector | QSFP/QSFP112, OSFP, or other connector configuration depending on OPN |
| Port Configuration | 1, 2, or 4 ports depending on model |
| Port Splitting | Supported on selected models |
| RDMA | Supported |
| RoCE | Supported |
| GPUDirect RDMA | Supported |
| GPUDirect Storage | Supported |
| SR-IOV | Supported |
| NVIDIA In-Network Computing | Supported |
| NVIDIA Multi-Host | Supported on applicable configurations |
| Socket Direct | Supported on applicable configurations |
| Virtualization | Hardware-assisted virtualization and SR-IOV |
| Overlay Networking | VXLAN, GENEVE, NVGRE and related acceleration |
| Security Acceleration | Hardware-assisted TLS, IPsec and MACsec functions on applicable Ethernet configurations |
| Network Processing | Hardware packet processing and programmable acceleration |
| Form Factors | PCIe HHHL, FHHL; OCP 3.0 TSFF/SFF |
| Operating Systems | Linux, Windows, VMware and supported virtualization/container environments |
| Linux Software | NVIDIA OFED / OpenFabrics Enterprise Distribution |
| Windows Software | WinOF-2 |
| Management | NVIDIA management and firmware utilities |
| Typical Deployment | Data-center server / AI server / HPC node / storage server |
| Maximum Throughput | Up to 400Gb/s |
| Architecture | NVIDIA ConnectX-7 networking architecture |
Specifications vary between ConnectX-7 OPNs; NVIDIA’s current documentation lists multiple configurations with different port counts, connector types, speeds and physical designs.
4. Applications
Artificial Intelligence & Machine Learning
- Distributed AI training
- Large-scale GPU clusters
- AI inference infrastructure
- GPU-to-GPU data movement
- Multi-node deep-learning environments
- High-performance AI factories
High-Performance Computing
- Scientific simulations
- Engineering simulations
- Computational research
- Molecular modeling
- Weather and climate modeling
- Financial modeling
- Parallel computing clusters
Data Centers & Cloud
- Hyperscale cloud infrastructure
- High-speed server networking
- Private and public cloud platforms
- Data-center interconnects
- High-throughput application servers
Storage
- NVMe-oF environments
- High-performance storage networks
- Distributed storage
- Software-defined storage
- GPUDirect Storage architectures
Enterprise Networking
- High-speed Ethernet aggregation
- Virtualized data centers
- Network-intensive enterprise applications
- High-performance database infrastructure
- Telecommunications and service-provider infrastructure
5. Compatibility
The NVIDIA ConnectX-7 is designed for servers and workstations equipped with compatible PCI Express Gen 5/Gen 4 infrastructure, with exact requirements determined by the adapter model.
It can be deployed with:
- NVIDIA GPU-accelerated servers
- AI/HPC servers
- Enterprise rack servers
- High-performance storage servers
- NVIDIA networking infrastructure
- Ethernet switches supporting the selected link speed
- NVIDIA InfiniBand switches
- RoCE-enabled Ethernet environments
- VMware virtualization environments
- Linux and Windows server platforms
- Kubernetes and containerized environments
NVIDIA documentation lists interoperability with Ethernet switches from 1/10/25/40/50/100/200/400Gb/s and InfiniBand infrastructures supporting SDR through NDR200, depending on the adapter configuration.
Important: Because ConnectX-7 covers numerous OPNs, the exact compatibility of a particular card should be checked against its full part number. Connector type, port count, maximum speed, InfiniBand/Ethernet mode, bracket and power requirements can differ between variants.
6. Key Benefits
- Up to 400Gb/s networking for extremely high-throughput infrastructure
- InfiniBand and Ethernet support across different ConnectX-7 variants
- PCIe Gen 5 host connectivity
- Hardware-accelerated RDMA/RoCE
- Excellent fit for GPU-accelerated AI clusters
- Support for GPUDirect RDMA and GPUDirect Storage
- Advanced networking offloads reduce CPU processing overhead
- SR-IOV support for virtualized environments
- Advanced overlay networking acceleration
- Hardware-assisted security acceleration
- Low-latency communication for distributed applications
- Flexible single-, dual-, and multi-port configurations
- Available in multiple server-friendly form factors
- Designed for AI, HPC, cloud, storage and hyperscale deployments
- Supports advanced NVIDIA networking technologies and In-Network Computing
7. SEO Title
NVIDIA ConnectX-7 400Gb/s Smart Network Adapter – PCIe Gen5 InfiniBand & Ethernet RDMA NIC



