NVIDIA ConnectX-7 NDR200 InfiniBand Single-Port OSFP Adapter

The NVIDIA ConnectX-7 NDR200 InfiniBand Single-Port OSFP Adapter (900-9X766-003N-SQ0) is a high-performance Host Channel Adapter (HCA) built on the ConnectX-7 ASIC, delivering 200 Gb/s NDR InfiniBand throughput through a single OSFP port. Operating over PCIe Gen 5.0 x16, the adapter provides the bandwidth and latency characteristics required by the most demanding AI training, high-performance computing, and data-intensive workloads. The ConnectX-7 architecture incorporates NVIDIA's SHARP (Scalable Hierarchical Aggregation and Reduction Protocol) in-network computing engine, enabling collective operations to be offloaded directly to the network fabric and significantly reducing the computational burden on host processors during distributed training runs.

The adapter supports GPUDirect RDMA, allowing GPU memory to be accessed directly by remote systems without CPU involvement, which is critical for tightly coupled GPU clusters executing large language model training and scientific simulation workloads. Hardware-based transport offloads including RDMA over InfiniBand (RoCE is not applicable here; native IB transport is used), zero-copy data movement, and end-to-end QoS mechanisms are implemented in silicon. The single OSFP port design consolidates high-bandwidth connectivity into a standard PCIe form factor, supporting integration into dense server platforms where slot availability is constrained.

Targeted at enterprise HPC centers, hyperscale AI infrastructure, and national research laboratories, the ConnectX-7 NDR200 adapter operates within NVIDIA Quantum-2 and compatible InfiniBand fabric ecosystems. It is compatible with the NVIDIA MLNX_OFED software stack and integrates with NVIDIA UFM (Unified Fabric Manager) for centralized fabric telemetry, congestion management, and routing. The adapter supports adaptive routing, multi-path communication, and Advanced Error Reporting (AER) for production-grade reliability in mission-critical compute environments.

Ideal for

• Large-scale distributed AI and deep learning model training across GPU clusters requiring low-latency, high-bandwidth inter-node communication
• High-performance computing (HPC) simulations in computational fluid dynamics, molecular dynamics, and climate modeling where MPI collective operations benefit from SHARP in-network offload
• GPU-to-GPU direct communication via GPUDirect RDMA in dense accelerated compute nodes, eliminating CPU bottlenecks in data movement
• Parallel file system access over InfiniBand in storage-intensive workloads such as genomics pipelines and seismic processing, leveraging RDMA for high-throughput I/O
• Hyperscale inference infrastructure requiring deterministic, ultra-low latency fabric connectivity between inference accelerator nodes and shared memory or storage resources
• Enterprise research and national laboratory environments deploying NVIDIA Quantum-2 InfiniBand fabrics where end-to-end NDR200 connectivity is required from host adapter to switch

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.