NVIDIA MQM8790-HS2F MELLANOX Quantum HDR Infiniband Switch 40 QSFP56 Ports 2 Power Supplies (ac)

NVIDIA MQM8790-HS2F Mellanox Quantum HDR Infiniband Switch

Overview

The NVIDIA MQM8790-HS2F is a 40-port HDR Infiniband switch designed for high-performance computing and AI infrastructure clusters. Built on the Mellanox Quantum architecture, this switch delivers the fabric performance required for tightly coupled compute nodes, GPU clusters, and data-parallel workloads where network latency and throughput directly impact application performance. The MQM8790-HS2F (commonly searched as MQM8790 HS2F) integrates dual AC power supplies for redundancy, minimizing single points of failure in production deployments.

Key Features

• 40 QSFP56 Ports: Native HDR speed (200 Gbps per port) without oversubscription — every port connects at full bandwidth. This matters when scaling GPU training jobs across dozens of nodes; packet loss or contention becomes a measurable drag on training throughput.
• Quantum HDR Architecture: Sub-microsecond latency fabric optimized for RDMA (Remote Direct Memory Access) workloads. GPU-to-GPU communication skips the CPU entirely, leaving compute cycles for actual training rather than protocol overhead.
• Dual AC Power Supplies: Redundant power input eliminates downtime from a single PSU failure. In HPC clusters running 24/7 model training or batch inference, unexpected power loss stops all jobs and forces restart from the last checkpoint.
• Compact 1RU Form Factor: 29.0 x 23.0 x 6.0 inches and 33 lbs fits standard 19-inch datacenter racks. Dense packaging means more switching capacity per rack unit, critical when power and cooling budgets are shared across many systems.
• In-Network Computing (INC) Capable: Mellanox Quantum switches support hardware-offloaded collective operations (AllReduce, AllGather). Distributed training frameworks like PyTorch and TensorFlow offload synchronization to the switch, cutting gradient aggregation time by 30–50% depending on cluster size.
• Non-Blocking Fabric: 8 Tbps switch fabric ensures no port-to-port bottleneck. All 40 ports can simultaneously communicate at line rate, essential for all-to-all communication patterns in distributed deep learning.

Integration and Compatibility

The MQM8790-HS2F integrates into Infiniband-based HPC clusters and AI infrastructure. Compatible with NVIDIA CUDA-aware networking, OpenFabrics Enterprise Distribution (OFED), and GPU Direct RDMA. Works with NVIDIA DGX systems, mainstream Linux distributions with Mellanox drivers, and third-party HPC workload managers (Slurm, PBS, Kubernetes with Infiniband CNI plugins). Supports Infiniband Subnet Manager software for fabric discovery, routing, and QoS configuration.

What's in the Box

The MQM8790-HS2F ships with 2 AC power supplies integrated; no additional PSU purchase required. Rack-mount brackets and documentation are included.

Frequently Asked Questions

Q: What is the bandwidth per port on the MQM8790-HS2F?

A: Each of the 40 QSFP56 ports delivers 200 Gbps unidirectional bandwidth in HDR mode, totaling 8 Tbps non-blocking fabric throughput across all ports.

Q: Does the MQM8790-HS2F support redundancy?

A: Yes. The two integrated AC power supplies provide N+1 power redundancy. If one PSU fails, the switch continues operating on the remaining supply without interruption.

Q: Is the MQM8790-HS2F suitable for GPU clustering?

A: Yes. Quantum HDR switches are standard in NVIDIA DGX SuperPOD clusters and multi-node GPU systems. Sub-microsecond latency and RDMA support make it ideal for distributed AI training and inference.

Q: What management interfaces are available?

A: The MQM8790-HS2F includes Infiniband Subnet Manager (ISM) for fabric management, in-band and out-of-band Ethernet management ports, and support for Mellanox Unified Fabric Manager (UFM) for enterprise monitoring and analytics.

Q: What are the physical dimensions and power requirements?

A: The switch measures 29.0 x 23.0 x 6.0 inches (L x W x H) and weighs 33 lbs. Dual AC supplies handle standard 100–240V input; total system power draw is typically 1.2–1.5 kW under full load.

Q: Is the MQM8790-HS2F compatible with third-party Infiniband adapters?

A: Yes. The switch is interoperable with any standard Infiniband HDR adapter card (HCA) that complies with Infiniband specifications, including Mellanox ConnectX-6, ConnectX-7, and equivalents from other vendors.

Eden Phillips

I've deployed the MQM8790-HS2F into multi-node GPU clusters scaling from 8 to 64 compute nodes, and the switch fabric becomes non-negotiable once you move beyond basic throughput metrics. The MQM8790-HS2F's 8 Tbps non-blocking architecture and sub-microsecond latency directly cut distributed training iteration time — we've seen 30–40% reductions in collective operation overhead (AllReduce synchronization) on large clusters compared to older generation fabric.

Technical Highlights:

• 8 Tbps Non-Blocking Fabric: All 40 ports run at full 200 Gbps simultaneously with zero oversubscription. In all-reduce patterns (GPU gradient synchronization), this eliminates port contention that would otherwise serialize slow-draining nodes and stall the entire iteration.
• Dual AC Power Supplies: N+1 redundancy means a PSU failure doesn't interrupt training. In 24/7 batch jobs, the difference between a graceful failover and a restart from the last checkpoint is often 6+ hours of lost compute time.
• Quantum HDR 200 Gbps per Port: 2.5x bandwidth density compared to EDR. For workloads with high all-to-all communication (multi-GPU training, distributed inference), the extra bandwidth keeps gradient aggregation from becoming the iteration bottleneck.

Deployment Considerations:

• Subnet Manager configuration is essential — misconfigured QoS or routing can eliminate fabric gains. Plan for OFED driver alignment across all compute nodes; version mismatches occasionally cause silent performance cliffs, not errors.
• The switch operates at full 1.2–1.5 kW power draw under load; rack power budgeting and cooling provisioning must account for this upfront, especially in shared-rack HPC clusters where other systems share the same PDU.

The MQM8790-HS2F is the right choice for enterprise HPC clusters and large-scale AI training infrastructure where GPU-to-GPU synchronization latency directly impacts cluster efficiency. For smaller clusters (≤8 nodes) or bandwidth-light workloads, older Infiniband generations can suffice, but once you scale to 16+ GPUs with frequent collective operations, the fabric becomes a measurable cost factor.