NVIDIA ConnectX-8 SuperNIC 400GbE Adapter

The NVIDIA ConnectX-8 SuperNIC is a 400GbE intelligent network adapter engineered to address the extreme bandwidth and latency demands of modern AI training and inference infrastructure. Built on NVIDIA's eighth-generation ConnectX ASIC, it delivers up to 400 gigabits per second of full-duplex Ethernet connectivity while integrating programmable in-network compute engines that offload collective communication operations—such as AllReduce and AllGather—directly onto the adapter, dramatically reducing CPU and GPU involvement in MPI and NCCL workloads. The adapter natively supports RDMA over Converged Ethernet version 2 (RoCEv2), enabling microsecond-class latency for distributed deep learning frameworks including PyTorch, TensorFlow, and JAX running across multi-node GPU clusters.

The ConnectX-8 SuperNIC incorporates NVIDIA's SHARP (Scalable Hierarchical Aggregation and Reduction Protocol) in-network computing technology, allowing collective operations to execute within the network fabric itself rather than at endpoints. This architectural advantage reduces effective collective communication latency by up to 50% and alleviates GPU idle time during synchronization phases, directly improving GPU utilization in large-scale training jobs. The adapter supports PCIe Gen 5.0 x16 host connectivity for maximum host-to-NIC bandwidth, ensuring the interface does not become a bottleneck when feeding multiple high-speed 400GbE ports. Hardware-based congestion control, adaptive routing telemetry, and integrated precision time protocol (PTP) synchronization further enhance its suitability for converged AI and HPC fabrics.

Designed for GPU-dense rack deployments in hyperscale and enterprise AI data centers, the ConnectX-8 SuperNIC is compatible with NVIDIA Spectrum-X Ethernet networking platforms and integrates tightly with the NVIDIA AI Enterprise software stack. It supports advanced telemetry streaming for real-time fabric visibility, hardware-accelerated TLS 1.3 and IPsec for secure AI workload isolation, and multi-host virtualization capabilities that allow a single adapter to be partitioned across multiple virtual machines or containers. Its combination of extreme bandwidth, in-network intelligence, and hardware offload engines positions it as a foundational component for next-generation AI supercomputing clusters.

Ideal for

• Large-scale distributed AI model training across hundreds of GPU nodes requiring ultra-low-latency collective communication via RoCEv2 and SHARP in-network reduction
• Generative AI inference serving farms demanding high-throughput, low-latency east-west traffic between GPU inference nodes and KV-cache storage pools
• High-performance computing clusters running MPI-based scientific simulations that benefit from in-network AllReduce offload to reduce synchronization overhead
• Multi-tenant GPU cloud infrastructure requiring hardware-enforced network isolation through SR-IOV virtualization and hardware-accelerated IPsec encryption per tenant
• AI storage fabric connectivity linking NVMe-oF all-flash arrays directly to GPU compute nodes with RDMA for near-memory access latencies to training datasets
• Real-time financial analytics and algorithmic trading platforms demanding sub-microsecond fabric latency and precision hardware timestamping via PTP synchronization

Technical specifications

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