English 
| Model Number: | |
|---|---|
| Product Type: | |
| Port Configuration: | |
| Network Speed: | |
| Switching Capacity: | |
| Latency: | |
| Architecture: | |
| Quantity: | |
MQM8790-HS2F
Mellanox
The MQM8790-HS2F is a 1U rack-mount fixed-configuration InfiniBand switch based on the NVIDIA Quantum ASIC, featuring 40 ports of HDR 200Gb/s connectivity in a non-blocking architecture. Designed for AI training clusters, high-performance computing fabrics, and large-scale storage networks, this externally managed switch supports in-network computing through SHARP technology and delivers sub-130ns port-to-port latency.
Category | Parameter | Value |
General | Product Model | MQM8790-HS2F |
Product Series | NVIDIA Quantum HDR InfiniBand | |
Form Factor | 1U rack-mount, standard depth | |
Physical | Dimensions | 43.6mm × 433.2mm × 590.6mm (H×W×D) |
Airflow Direction | Port-to-Cooling (P2C) | |
Ports | Port Configuration | 40 × QSFP56 ports |
Port Speed | 200Gb/s HDR per port | |
Port Splitting | Up to 80 × 100Gb/s HDR100 via breakout cables | |
Performance | Switching Capacity | 16Tb/s non-blocking aggregate bandwidth |
Forwarding Rate | 7.2 billion packets per second | |
Port-to-Port Latency | < 130ns | |
MTU Range | 256 to 4000 bytes | |
Virtual Lanes | 9 (8 data + 1 management) | |
Hardware | CPU | Broadwell ComEx D-1508 2.2GHz |
System Memory | 8GB DDR4 | |
Management | Externally managed via NVIDIA UFM | |
Power | Power Supplies | 1+1 redundant, hot-swappable AC PSUs |
Input Voltage | 100-127VAC 4.5A, 200-240VAC 2.9A | |
Typical Power Consumption | 224W | |
Maximum Power Consumption | 740W | |
Cooling | Fan Configuration | N+1 redundant, hot-swappable fans |
Environmental | Operating Temperature | 0°C to 40°C |
Storage Temperature | -40°C to 70°C | |
Operating Humidity | 10% to 85% non-condensing | |
Operating Altitude | Up to 3200m | |
Regulatory | Safety Certifications | CB, cTUVus, CE, CU |
EMC Certifications | CE, FCC, VCCI, ICES, RCM | |
Environmental Compliance | RoHS-6 compliant |
The switch includes NVIDIA SHARP (Scalable Hierarchical Aggregation and Reduction Protocol) technology, which offloads collective communication operations from host CPUs to the switch silicon. This enables in-network processing of MPI and NCCL collective operations used in distributed AI training and HPC applications, reducing the number of data transfers between endpoints and improving overall fabric utilization.
The ASIC implements adaptive routing algorithms that dynamically select network paths to avoid congestion hotspots. Combined with hardware-based congestion control mechanisms, this feature maintains effective bandwidth across large-scale fabrics and supports modern topologies including Fat Tree, Dragonfly+, SlimFly, and 6-dimensional Torus configurations.
This switch is suitable for top-of-rack leaf deployments as well as spine layers in medium-sized clusters. The 40-port configuration supports 40 nodes at 200Gb/s or 80 nodes at 100Gb/s using breakout cables, enabling dense rack configurations with up to 80 servers per rack when using HDR100 connectivity.
QSFP56 ports support passive copper DAC cables, active optical cables (AOC), and optical transceivers for different reach requirements. The switch supports port splitting using QSFP56 to 2×QSFP28 breakout cables for migration from 100Gb/s to 200Gb/s fabrics. Backward compatibility includes EDR and FDR InfiniBand speeds for mixed-generation environments.
The MQM8790-HS2F is an externally managed switch designed to operate with NVIDIA UFM (Unified Fabric Manager) for end-to-end fabric provisioning, monitoring, and maintenance. Management interfaces include an RJ45 Ethernet port, RS232 console port, and USB management port running MLNX-OS for local configuration.
The switch includes 1+1 redundant hot-swappable power supplies and N+1 redundant hot-swappable fan modules, enabling component replacement without system downtime. All FRUs (field-replaceable units) can be serviced without taking the switch offline, supporting continuous operation in production data center environments.
The MQM8790-HS2F uses Port-to-Cooling (P2C) airflow, where air enters through the port side and exhausts through the power supply side. The MQM8790-HS2R uses Cooling-to-Port (C2P) airflow with the reverse direction. Airflow direction must match server airflow in hot-aisle/cold-aisle data center configurations.
No, the MQM8790-HS2F is externally managed and does not include an on-board subnet manager. It requires NVIDIA UFM or another external subnet manager for fabric operation. The MQM8700 series includes embedded subnet manager capability for smaller deployments.
No, the MQM8790-HS2F is an InfiniBand-only switch. It does not support Ethernet protocols. For mixed Ethernet and InfiniBand environments, separate switching infrastructure is required.
The unit includes a 1-year standard hardware warranty with advanced replacement options. Extended warranty and support contracts are available for enterprise deployments, including firmware updates and technical assistance.
