NVIDIA InfiniBand Switches

Standard Ethernet works well for general business traffic, but AI and HPC workloads create far heavier communication between systems during training, simulation and parallel processing. As clusters grow, latency and congestion can start affecting overall workload efficiency.

NVIDIA InfiniBand Switches are designed for high-performance cluster fabrics, helping maintain predictable throughput across AI training environments, HPC platforms and high-performance storage deployments where the network directly affects compute performance.

Not every clustered environment needs 400Gb/s fabric capacity immediately. Departmental AI clusters, research environments and mid-sized HPC deployments often still need low-latency performance without the scale and infrastructure overhead associated with NDR.

The QM8700 and QM8790 provide 40 HDR 200Gb/s ports in a compact 1U design, making them well suited to clustered workloads that need high-performance InfiniBand without moving into larger NDR fabrics.

Internally managed switches can work well in smaller or isolated environments. The challenge appears when multiple AI, HPC or research clusters need consistent operational control across a wider infrastructure estate.

The QM8790 is externally managed for use with centralised fabric management platforms, while the QM8700 includes onboard management. Both provide the same 40 HDR 200Gb/s connectivity, so the decision is mainly operational rather than performance-based.

HDR fabrics remain effective while cluster sizes and traffic demands stay moderate. As environments scale, larger datasets and distributed workloads increase the need for higher bandwidth and flatter fabric designs.

The QM9700 provides 64 NDR 400Gb/s ports in 1U, helping support large AI training clusters, simulation platforms and research computing environments with fewer switching layers and lower infrastructure complexity.

Onboard switch management can suit smaller deployments, but larger clustered environments often need centralised operational control across the wider fabric.

The QM9790 delivers the same Quantum-2 NDR switching class as the QM9700, but is designed for externally managed environments using Nvidia Unified Fabric Manager tools for provisioning, monitoring and fabric maintenance across shared AI and HPC estates.

Smaller AI and HPC environments may continue operating effectively on HDR infrastructure while cluster traffic remains manageable. The move to NDR becomes more relevant as workloads, datasets and inter-node communication demands increase.

The QM9700 and QM9790 provide 64 NDR 400Gb/s ports in 1U, supporting large AI training environments, simulation platforms and research computing estates where consistent node-to-node performance matters at scale.