Placa de rede MELLANOX MCX516A-CCAT interna PCI Express, interface de fibra, 100000 Mbit/s

MELLANOX MCX516A-CCAT is a dual‑port 100Gb Ethernet host adapter with a PCI Express interface and fiber ports, engineered for high‑bandwidth servers, storage and HPC clusters. It implements RoCE (RDMA over Converged Ethernet) and offers sub‑600 ns latency, very high message rates, PCIe switch and NVMe over Fabrics offloads, and features to improve multipath reliability and scalability in demanding compute and data‑analytics environments.

Advantages

The adapter delivers low latency and high throughput for data‑intensive workloads, with hardware offloads for RDMA, MPI and SHMEM/PGAS rendezvous tag matching. It supports out‑of‑order RDMA write/read with ordered completion semantics, Network Atomic and PCIe Atomic operations, Burst Buffer offload for background checkpointing, and Dynamic Connected Transport (DCT) for extreme scalability. These capabilities reduce CPU overhead and improve overall system efficiency.

Suitable for

• High‑performance computing (HPC) clusters and supercomputing nodes
• Machine learning training and inference clusters
• Data analytics and large‑scale data processing
• Storage networks and NVMe over Fabrics deployments

Key specifications

• Network speed: up to 100000 Mbit/s (100 Gbps)
• Interface: fiber network ports
• Host interface: PCI Express
• Latency: sub‑600 ns
• RDMA over Converged Ethernet (RoCE) support
• Dual‑port 100Gb Ethernet
• Hardware offloads: RDMA, NVMe over Fabrics, MPI/SHMEM/PGAS tag matching

How to use

Install the MCX516A-CCAT into a free PCI Express slot on a compatible server or workstation, connect the fiber transceivers and cabling to the network fabric, and load the vendor drivers and firmware recommended by MELLANOX. Configure RoCE and RDMA settings according to your cluster or storage fabric requirements, enable NVMe over Fabrics offload if using NVMe targets, and verify multipath and adaptive routing settings for your topology.

Recommendations

For optimal performance in HPC, AI and storage environments, use certified firmware and drivers, ensure the network fabric supports RoCE and lossless operation, and configure DCT and Burst Buffer features according to your checkpointing and scaling needs.