NVIDIA ConnectX-7 400GbE Dual-Port OSFP InfiniBand/Ethernet Adapter

The NVIDIA ConnectX-7 MCX755106AC-HEAT is a dual-port 400GbE/HDR InfiniBand adapter built on the ConnectX-7 ASIC, delivering up to 400 Gb/s per port over OSFP connectors. Engineered for the demands of next-generation AI supercomputing and high-performance computing (HPC) clusters, the adapter supports both InfiniBand HDR and Ethernet protocols natively, enabling seamless fabric convergence across heterogeneous data center environments. Hardware-based crypto acceleration is integrated directly into the silicon, supporting IPsec and TLS offload without CPU overhead.

The ConnectX-7 architecture introduces NVIDIA's SHARP (Scalable Hierarchical Aggregation and Reduction Protocol) in-network computing engine, which offloads collective operations—such as all-reduce—directly onto the network fabric, substantially reducing latency and freeing host compute resources during large-scale distributed training workloads. The adapter supports PCIe Gen 5.0 x16 host connectivity, providing the bandwidth headroom necessary to keep pace with modern GPU accelerators and multi-socket server platforms. Advanced features include hardware-based flow steering, RDMA over Converged Ethernet (RoCE v2), and GPUDirect RDMA support for direct peer-to-peer data transfers between network and GPU memory.

Targeted at AI training clusters, deep learning inference farms, financial analytics platforms, and large-scale storage fabrics, the MCX755106AC-HEAT is designed to operate reliably in dense, thermally demanding environments. The adapter ships in a standard low-profile, full-height form factor and is compatible with leading server platforms. Its dual-port OSFP configuration supports 2×400GbE or 2×HDR 200Gb/s InfiniBand, providing flexible fabric connectivity for evolving infrastructure architectures.

Ideal for

• Large-scale distributed AI and deep learning model training across GPU clusters requiring ultra-low latency, high-bandwidth fabric interconnects with in-network collective offload via NVIDIA SHARP
• High-performance computing (HPC) simulation workloads—such as computational fluid dynamics, molecular dynamics, and climate modeling—demanding deterministic, low-latency InfiniBand HDR fabric connectivity
• NVMe-oF (NVMe over Fabrics) all-flash storage arrays and disaggregated storage architectures requiring high-throughput, low-latency RDMA transport between compute and storage tiers
• Secure multi-tenant cloud and enterprise data center environments leveraging hardware-accelerated IPsec and TLS encryption to protect east-west traffic without imposing CPU overhead
• Financial services high-frequency trading and quantitative analytics platforms requiring sub-microsecond MPI latency and deterministic network performance across tightly coupled compute nodes
• AI inference serving infrastructure and real-time data analytics pipelines benefiting from GPUDirect RDMA to eliminate host CPU bottlenecks in GPU-to-network data paths

Technical specifications

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