产品详情
| Application | AI, HPC, Hyperscale Computing, Scientific Research |
| Condition | [Manual Fill] |
| Model Number | QM9700 Series (see comparison table for OPN details) |
| Port Configuration | 64 ports 400Gb/s InfiniBand (32 OSFP) or 128 ports 200Gb/s via port split |
| Management Type | Managed (MQM9700) / Unmanaged with UFM (MQM9790) |
| Airflow Direction | P2C (forward) or C2P (reverse) per model |
| Highlight | 51.2Tb/s aggregated throughput · 66.5Bpps · SHARPv3 · 1U |
NVIDIA Quantum‑2 QM9700 series 400Gb/s InfiniBand smart switches: 64 ports, 51.2Tb/s in 1U with third‑gen SHARP in‑network computing.
Product Overview
The NVIDIA Quantum‑2 QM9700 family redefines scalable networking for the most demanding AI and HPC environments. Integrating 64 ports of 400Gb/s NDR InfiniBand in a compact 1U chassis, the switch delivers an unprecedented aggregated bidirectional throughput of 51.2Tb/s and over 66.5 billion packets per second. Built on the NVIDIA high‑speed interconnect architecture, it combines RDMA, adaptive routing, and the third generation of NVIDIA Scalable Hierarchical Aggregation and Reduction Protocol (SHARPv3) to dramatically accelerate distributed applications. Whether deployed as a managed Top‑of‑Rack switch with on‑board subnet manager (MQM9700) or as part of a large fabric controlled by NVIDIA Unified Fabric Manager (MQM9790), the QM9700 series offers unmatched flexibility, density, and efficiency for next‑generation data centers.
Key Characteristics
- Extreme Throughput & Density: 64 ports of 400Gb/s (51.2Tb/s aggregate) and 66.5Bpps in a 1U form factor – the industry’s highest radix.
- In‑Network Computing with SHARPv3: Third‑generation SHARP delivers 32X higher AI acceleration by offloading and reducing data movement.
- Port Split Technology: Double the port count to 128 ports of 200Gb/s, enabling cost‑effective Fat Tree and other topologies.
- Advanced Management Options: On‑board subnet manager for up to 2,000 nodes (MQM9700) or UFM‑based fabric management (MQM9790) with CLI, WebUI, JSON, SNMP.
- Low‑Latency & Self‑Healing Fabric: RDMA, adaptive routing, QoS, enhanced virtual lane mapping, and congestion control ensure optimal application performance.
- Backward Compatible & Flexible Topologies: Supports Fat Tree, SlimFly, DragonFly+, multi‑dimensional Torus; seamless integration with previous InfiniBand generations.
Core Technologies
- NDR 400Gb/s InfiniBand: 400G per port with 51.2Tb/s non‑blocking switching fabric.
- SHARPv3 (Scalable Hierarchical Aggregation and Reduction Protocol): In‑network computing that accelerates AI and collective operations by 32X vs. previous gen.
- RDMA (Remote Direct Memory Access): Low‑latency, kernel‑bypass data transfer for MPI and storage traffic.
- Adaptive Routing & Self‑Healing: Dynamically avoids congestion and reroutes around link failures for maximum fabric utilization.
- Port Splitting: Converts 32 OSFP slots into 128 x 200Gb/s ports, slashing network design costs.
- Comprehensive Telemetry & QoS: Enhanced virtual lanes, congestion control, and fine‑grained quality of service.
Operational Principle
The QM9700 series acts as a smart spine‑leaf element within an InfiniBand fabric. By embedding SHARPv3 technology, the switch aggregates and reduces data directly in the network as messages traverse towards the root, dramatically cutting inter‑node communication overhead. Adaptive routing dynamically balances traffic across available paths, while the on‑board subnet manager (in managed models) automatically configures routing tables and discovers up to 2,000 nodes. When deployed with UFM (MQM9790), the fabric gains AI‑driven predictive maintenance and real‑time optimisation without burdening the control plane.
Application Scenarios
- AI Training & Deep Learning: Massive parallelism with SHARPv3 for gradient aggregation and all‑reduce acceleration.
- High‑Performance Computing (HPC): Large‑scale simulations, weather modeling, and quantum chemistry with ultra‑low latency.
- Hyperscale Data Analytics: Real‑time processing of extreme‑size datasets across thousands of nodes.
- Cloud & Enterprise Data Centers: Dense 400G fabric for multi‑tenant workloads, supporting RDMA and virtualized environments.
Specifications & Model Selection
| Feature | Specification |
|---|---|
| Performance | 400Gb/s per port; 51.2Tb/s aggregate throughput; 66.5Bpps |
| Port Configuration | 64 ports 400Gb/s (32 OSFP connectors) or 128 ports 200Gb/s with port split |
| CPU / Memory / Storage | x86 Coffee Lake i3; 8GB DDR4 SO‑DIMM (2666 MT/s); 16GB M.2 SATA SSD |
| Management Interfaces | 1x USB 3.0, 1x USB (I2C), 1x RJ45, 1x RJ45 (UART); CLI, WebUI, SNMP, JSON |
| Power Supply | 1+1 redundant & hot‑swappable; 200‑240Vac; 80 Gold+ / ENERGY STAR certified |
| Cooling | Front‑to‑rear or rear‑to‑front; hot‑swappable fan units |
| Software | MLNX‑OS / On‑board subnet manager (MQM9700) / UFM ready (MQM9790) |
| Dimensions (H×W×D) | 1.7 in (43.6 mm) x 17.0 in (438 mm) x 26.0 in (660.4 mm); 1U |
| Weight | 14.5 kg |
| Operating Conditions | 0 °C to 40 °C; 10%–85% non‑condensing; altitude up to 3050m |
| Compliance & Warranty | CE, FCC, VCCI, ICES, RCM; CB, cTUVus, CE, CU; 1‑year limited warranty |
Series Comparison
| Orderable Part Number (OPN) | Description | Key Differentiator |
|---|---|---|
| MQM9700-NS2F | 64‑port 400Gb/s, 32 OSFP, managed, P2C airflow (forward) | On‑board subnet manager, CLI/WebUI, front‑to‑rear cooling |
| MQM9700-NS2R | 64‑port 400Gb/s, 32 OSFP, managed, C2P airflow (reverse) | Same management, rear‑to‑front cooling for hot‑aisle containment |
| MQM9790-NS2F | 64‑port 400Gb/s, 32 OSFP, unmanaged, P2C airflow (forward) | UFM‑managed fabric, ideal for large‑scale dynamic networks |
| MQM9790-NS2R | 64‑port 400Gb/s, 32 OSFP, unmanaged, C2P airflow (reverse) | UFM ready, reverse airflow for flexible data center layouts |
Competitive Advantages
- Highest Port Density: 64 ports of 400Gb/s or 128 ports of 200Gb/s in 1U, reducing rack space and cabling.
- Embedded SHARPv3: Unique in‑network acceleration that boosts AI and HPC application performance while reducing network load.
- Dual Management Paradigm: Choose integrated subnet manager for small/medium clusters or UFM for hyperscale telemetry and preventative troubleshooting.
- Topology Flexibility: Native support for Fat Tree, SlimFly, DragonFly+, Torus; backward compatibility protects investment.
- Rich Software Ecosystem: Full MLNX‑OS support, JSON automation, and integration with NVIDIA collective acceleration libraries.
Service & Support
- Warranty: 1‑year limited hardware warranty (as specified).
- Software & Management: MLNX‑OS included; on‑board subnet manager for up to 2,000 nodes (MQM9700); UFM optional for MQM9790.
- Certifications: CE, FCC, VCCI, ICES, RCM (emissions); CB, cTUVus, CE, CU (safety); RoHS, ENERGY STAR compliant.
- Technical Resources: NVIDIA enterprise support portal, firmware updates, and documentation.
Frequently Asked Questions
What is the total aggregated throughput of the QM9700 series?
The switch delivers 51.2 Tb/s bidirectional aggregated throughput (64 ports × 400Gb/s full duplex).
How many nodes can the internal subnet manager handle?
The on‑board subnet manager (on MQM9700 models) supports up to 2,000 nodes out of the box.
Does the switch support previous generation InfiniBand?
Yes, the QM9700 series is backward compatible with earlier InfiniBand generations, enabling mixed‑speed fabrics.
What airflow directions are available?
Both forward (P2C, power‑to‑connector) and reverse (C2P, connector‑to‑power) versions are offered – see ordering information.
Important Precautions
- Operating Temperature: 0 °C to 40 °C; non‑operating -40 °C to 70 °C.
- Humidity: Operating 10% to 85% non‑condensing; storage 10% to 90% non‑condensing.
- Altitude: Do not exceed 3050 meters above sea level.
- EMC & Safety: Installation must comply with local regulations; use only specified power supply range (200‑240Vac).
- Cooling: Ensure proper airflow clearance (front‑to‑rear or rear‑to‑front) and hot‑swap fans are accessible.