H3C CX6855 200G HDR InfiniBand and Ethernet Dual-Port NIC

The H3C CX6855 is a high-performance dual-port network interface card engineered for next-generation data center interconnect requirements. Each port supports 200G throughput across both HDR InfiniBand and native Ethernet protocols, enabling unified fabric deployment in GPU-accelerated computing clusters and HPC environments. The adapter implements full RDMA over Converged Ethernet (RoCE v2) support, eliminating TCP/IP stack latency to achieve microsecond-level message latencies critical for collective communication in distributed AI training and scientific computing workloads.

Architecturally, the CX6855 integrates advanced packet processing, flow steering, and congestion management hardware to maintain line-rate throughput while supporting thousands of concurrent connections. The dual-port design enables multi-path networking configurations, redundancy, and load balancing across cluster nodes. Native support for InfiniBand and Ethernet protocols on the same hardware platform simplifies network architecture decisions for heterogeneous HPC environments transitioning between technologies.

Launched in Q2 2024, the CX6855 addresses persistent capacity constraints in AI and HPC cluster buildout, where demand for high-bandwidth, low-latency interconnects significantly exceeds supply. The adapter integrates comprehensive monitoring, telemetry, and adaptive routing capabilities to maintain cluster stability at scale. Compatibility with PCIe 5.0 Gen 5 slot architecture enables maximum bandwidth utilization within modern server platforms.

Ideal for

• GPU cluster interconnect fabric for distributed large language model (LLM) training, enabling rapid all-reduce and collective operations across hundreds of compute nodes
• High-performance computing (HPC) simulation and modeling environments requiring sub-microsecond latency for MPI-based scientific applications
• Real-time financial data analytics and algorithmic trading infrastructure demanding deterministic low-latency network behavior
• Machine learning inference serving with in-memory feature stores and real-time data pipeline acceleration via RDMA
• Enterprise storage area networks (SANs) and NVMe-over-Fabrics (NVMe-oF) deployments for persistent memory pools and disaggregated storage architectures
• Research computing facilities requiring unified fabric for both compute interconnect and high-speed data transfer between nodes

Technical specifications

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