NVIDIA Quantum MQM8790-HS2F 200G InfiniBand Switch 40-Port Unmanaged 200Gb/s

NVIDIA Quantum MQM8790-HS2F 200Gb/s InfiniBand Smart Switch

Accelerate HPC, AI, and cloud infrastructures with a 40-port 200G non-blocking switch that integrates in-network computing and double-density 100G splitting—engineered for maximum throughput and minimal latency.

• 40 QSFP56 ports operating at 200Gb/s or bifurcated into 80 x 100Gb/s ports
• Aggregate switching capacity up to 16 Tb/s (non-blocking)
• Integrated SHARP technology for collective communication acceleration
• Unmanaged SKU (MQM8790-HS2F) with x86 dual-core Broadwell CPU & onboard subnet manager support up to 2000 nodes
• P2C airflow, 1+1 redundant power supplies, 1U standard depth, backwards compatible with ConnectX-6 and previous-gen adapters

Characteristics

• High-Density 200G Fabric: 40 QSFP56 ports each delivering 200Gb/s with full bidirectional bandwidth.
• Port Splitting Flexibility: Convert any 200G port into two independent 100G ports—ideal for double-density ToR deployments.
• NVIDIA SHARP Technology: Hardware-based in-network computing that offloads and accelerates MPI, NCCL, and SHMEM collectives.
• Advanced Architecture Support: Fat Tree, DragonFly+, multi-dimensional Torus, and adaptive routing for congestion-free fabrics.
• Energy-Efficient Design: 80 PLUS Gold certified, power scales with port utilization, RoHS compliant.

Technology & Standards

Built on NVIDIA Quantum InfiniBand architecture, the MQM8790-HS2F leverages Scalable Hierarchical Aggregation and Reduction Protocol (SHARP) to perform data reductions inside the switch fabric. It supports static and adaptive routing, QoS, and congestion control. The switch is fully compatible with 200G/100G HDR/HDR100 InfiniBand and backward compatible with FDR/EDR, ensuring smooth integration into existing data centers. The x86 management plane runs MLNX-OS (unmanaged version relies on external UFM or OpenSM).

Working Principle: In-Network Computing

Unlike traditional switches that simply forward packets, the MQM8790-HS2F processes data as it traverses the network. When collective operations (e.g., all-reduce, barrier) are issued by servers, the switch hardware aggregates or reduces data at wire speed. This eliminates multiple endpoint transmissions, dramatically lowering latency for AI training and HPC simulations. The embedded x86 processor manages fabric initialization and topology discovery, while data acceleration occurs in dedicated silicon.

Applications & Deployment

• Hyperscale AI Clusters: Parallel training of large models requiring low-latency all-reduce.
• High-Performance Computing (HPC): MPI-based simulations in weather, genomics, and physics.
• Double-Density Top-of-Rack (ToR): Splitting ports to 100G for dense server connectivity.
• Cloud & Storage: NVMe-oF over InfiniBand, high-throughput storage backends.
• Research & Education: Backbone for campus supercomputing centers.

Technical Specifications & Ordering Options

Note: All models include x86 dual-core Broadwell CPU, 8GB memory, 1x RJ45, 1x console, 1x micro USB. Weight with 2 PSUs: ~12.48kg. Cables: QSFP56 passive/copper/active optical.

Advantages & Differentiators

• vs. Traditional 100G Switches: Native 200G ports plus port splitting delivers double the radix, reducing network tiers and lowering CAPEX.
• Integrated SHARP: Unlike standard switches, the MQM8790-HS2F accelerates collectives in hardware, reducing job completion times by up to 10x.
• Energy & Space Efficient: 1U form factor with up to 16Tb/s throughput consumes less power per Gb than competing solutions; ENERGY STAR certified.
• Ecosystem Backward Compatibility: Seamless operation with ConnectX-4/5/6 network cards and existing cabling.

Service & Support

We provide comprehensive pre- and post-sales support: 24/7 technical consultation, RMA warranty handling, and on-site troubleshooting coordination. Our team assists with topology design, firmware compatibility checks, and integration with existing InfiniBand fabrics. All units carry a minimum 1-year warranty (extendable). Orders ship within 24-48 hours from our 10M+ inventory.

Frequently Asked Questions (FAQ)

Q: Is the MQM8790-HS2F compatible with standard network cards like Mellanox ConnectX-6?

A: Yes, it is fully compatible with ConnectX-6 network cards (200G/100G) and previous generations (ConnectX-5, ConnectX-4) in backward-compatible mode. For optimal performance, use with network card that supports HDR or HDR100.

Q: Can I manage this switch without an external controller?

A: The MQM8790-HS2F is an unmanaged switch; however, it includes an onboard subnet manager capable of bringing up a fabric of up to 2000 nodes out-of-the-box. For advanced monitoring and management, NVIDIA UFM or OpenSM is recommended.

Q: How does port splitting to 100G work technically?

A: Each QSFP56 port carries four SerDes lanes. By configuring the port as two independent 2-lane interfaces, the switch can drive two separate 100G connections (using a splitter cable). This effectively doubles the port count for 100G endpoints.

Q: What is the typical power consumption at half load?

A: The switch dynamically reduces power when ports are unused. At 50% utilization, typical power draw is around 150-180W, depending on cable types and traffic.

Q: Which airflow direction does the MQM8790-HS2F use?

A: The -HS2F variant uses P2C (port side to cooling fan) airflow—air enters from the port side and exits at the fan side. For reverse airflow (C2P), order MQM8790-HS2R.

Precautions & Compatibility Notes

• Airflow Direction: Ensure your rack cooling matches the switch model (P2C for MQM8790-HS2F, C2P for -HS2R). Improper airflow may cause thermal throttling.
• Power Supply: Input range 100-127V or 200-240V AC. Use only supplied power cords. Dual feeds are recommended for redundancy.
• Cabling: For 200G operation, use QSFP56 passive copper (up to 3m) or active optical modules. 100G splitting requires QSFP56 to 2xQSFP28 breakout cables.
• Operating Environment: Temperature 0°C–40°C, humidity 10%–85% non-condensing. Altitude ≤3200m. EMC compliant with CE, FCC, VCCI, ICES, RCMS.
• Software: While unmanaged, firmware updates are performed via external host. Always verify firmware compatibility with your network card and subnet manager version.

Company Introduction