NVIDIA InfiniBand Switches for AI GPU Clusters: QM9700, QM9790, Quantum-2 and Quantum-X800
- May 17
- 7 min read
AI GPU clusters are not only about GPUs. If you are building a cluster with NVIDIA H100, H200, B200, B300, HGX, DGX, GB200, or GB300 systems, the network fabric is a critical part of the setup.
NVIDIA InfiniBand Switches
QM9700, QM9790, Quantum-2 and Quantum-X800
For many AI and HPC projects, buyers need NVIDIA Quantum-2 QM9700, QM9790, NDR 400G InfiniBand switches, ConnectX adapters, and OSFP cables. For newer Blackwell-generation systems, Quantum-X800 XDR 800G InfiniBand becomes more relevant.
server-parts.eu can help source refurbished NVIDIA InfiniBand switches and complete GPU cluster networking parts for AI, HPC, and data center projects.
NVIDIA InfiniBand Switches for AI, HPC and GPU Cluster Networking
In a single GPU server, GPUs communicate through PCIe, NVLink, or NVSwitch, depending on the server design. But when an AI workload runs across multiple servers, the external network becomes critical.
This is where NVIDIA InfiniBand switches are used.
InfiniBand is popular in AI and HPC because it is built for:
low latency networking
high bandwidth
RDMA
GPU-to-GPU communication between servers
multi-node AI training
HPC cluster workloads
LLM training
AI factory networking
GPU cloud and GPU-as-a-Service platforms
For serious GPU clusters, normal Ethernet is often not the first choice. Many buyers search for NVIDIA InfiniBand switch, AI cluster switch, H100 cluster switch, H200 cluster networking, NDR InfiniBand switch, or GPU cluster networking hardware because they need the correct fabric for expensive GPU systems.
NVIDIA Quantum-2 QM9700 and QM9790 InfiniBand Switches
The NVIDIA Quantum-2 QM9700 and NVIDIA Quantum-2 QM9790 are among the most important switches for modern AI GPU clusters.
They are commonly used with:
NVIDIA H100 clusters
NVIDIA H200 clusters
HGX H100 / HGX H200 servers
DGX H100 / DGX H200 environments
Dell PowerEdge XE9680
Dell PowerEdge XE9640
Supermicro HGX GPU servers
Lenovo GPU servers
HPE GPU and HPC systems
custom AI training clusters
GPU-as-a-Service platforms
The NVIDIA Quantum-2 QM9700/QM9790 switch family supports 64 ports of 400Gb/s NDR InfiniBand in a 1U chassis, with 32 OSFP connectors and up to 51.2Tb/s aggregate bidirectional throughput.
NVIDIA QM9700 vs QM9790 InfiniBand Switches: What Is the Difference?
Both switches belong to the NVIDIA Quantum-2 NDR InfiniBand family. The main difference is usually the management style.
Feature | NVIDIA QM9700 | NVIDIA QM9790 |
Switch family | Quantum-2 | Quantum-2 |
Network type | InfiniBand | InfiniBand |
Speed class | NDR 400G | NDR 400G |
Effective ports | 64x 400G | 64x 400G |
Physical connectors | 32x OSFP | 32x OSFP |
Form factor | 1U | 1U |
Management | Managed | Externally managed / unmanaged |
Typical use | Standard AI/HPC clusters | Larger managed fabrics |
QM9700 is often easier for standard cluster builds. QM9790 is often used in larger fabrics where management is handled externally.
NVIDIA Quantum-X800 Switches for 800G InfiniBand AI Clusters
For newer AI infrastructure, especially Blackwell-generation GPU clusters, buyers may search for NVIDIA Quantum-X800, XDR InfiniBand, or 800G InfiniBand switch.
NVIDIA describes Quantum-X800 as an 800Gb/s InfiniBand platform. NVIDIA lists the Quantum-X800 switch with 144 ports of 800Gb/s connectivity per port.
NVIDIA’s Quantum-X800 cluster documentation also lists Q3400 systems with 144 XDR ports over 72 OSFP cages.
For many H100 and H200 projects, Quantum-2 NDR 400G is still the main target. For newer B200, B300, GB200, and GB300 projects, Quantum-X800 XDR 800G becomes more relevant.
Which NVIDIA InfiniBand Switch Fits Which GPU Cluster?
GPU cluster type | Common InfiniBand switch class | Notes |
A100 clusters | HDR 200G or NDR 400G | Depends on age, budget, and topology |
H100 clusters | Quantum-2 NDR 400G | QM9700 / QM9790 are common choices |
H200 clusters | Quantum-2 NDR 400G | Similar network class to many H100 builds |
B200 clusters | Quantum-2 400G or Quantum-X800 800G | Depends on server platform and scale |
B300 clusters | Quantum-X800 800G direction | More relevant for newer Blackwell Ultra systems |
GB200 / GB300 NVL72 | Rack-scale NVIDIA networking | Not a normal small cluster design |
The important point: the right switch depends on the GPU generation, server model, number of GPU nodes, ConnectX adapter type, topology, and cable plan.
ConnectX Adapters (NVIDIA InfiniBand Switch): The Server-Side Part of the Fabric
The switch is only one side of the network. The GPU servers also need the correct network adapters.
For many H100 and H200 clusters, the common adapter is:
NVIDIA ConnectX-7 NDR 400G InfiniBand
For newer 800G designs, buyers may look for:
NVIDIA ConnectX-8 / 800G-class networking
If an offer only says “network card included,” that is not enough. For serious AI cluster networking, buyers should check the exact adapter model, port speed, cable type, and firmware compatibility.
Cables for NVIDIA InfiniBand Switches
Cables are a major part of an AI cluster networking package. A cluster offer is not complete if it only lists switches.
For Quantum-2 QM9700 / QM9790 systems, buyers often need:
OSFP cables
OSFP112 AOC cables
OSFP DAC cables
400G NDR InfiniBand cables
800G OSFP to 2x400G breakout cables
ConnectX-7 compatible cables
rack-level short cables
inter-rack active optical cables
Before buying refurbished NVIDIA InfiniBand switches, always check whether the cable set is included. Cables can be expensive, hard to source, and project-critical.
NVIDIA InfiniBand Switch Price Range
Many buyers search for NVIDIA QM9700 price, NVIDIA QM9790 price, Quantum-X800 price, used InfiniBand switch price, or refurbished NVIDIA InfiniBand switch price.
Prices change quickly depending on condition, warranty, airflow direction, accessories, cables, region, and real stock availability.
Market prices are approximately:
Switch / platform | Public online price examples | Notes |
NVIDIA Quantum-2 QM9700 / MQM9700 | €25,000–€30,000 | Depends on new/refurbished condition, airflow, seller, and warranty |
NVIDIA Quantum-2 QM9790 / MQM9790 | €25,000–€30,000 | Availability can be more limited |
NVIDIA Quantum-X800 Q3400 / 800G XDR | €60,000–€80,000 | Newer 800G-class systems are expensive and less common on the refurbished market |
Older HDR 200G InfiniBand switches | €5,000–€10,000 | More relevant for A100 or older HPC clusters |
These prices should be treated as market indicators, not fixed project prices.
The final project cost depends on:
exact switch model
airflow direction
new, used, refurbished, or recertified condition
warranty length
rail kits and power supplies
cables and transceivers
delivery country
quantity
topology support
current AI hardware availability
For a serious AI cluster, the switch price alone is not enough. Buyers should calculate the full fabric cost, including NVIDIA InfiniBand switches, ConnectX adapters, OSFP cables, breakout cables, port mapping, and warranty.
Refurbished NVIDIA InfiniBand Switches
Refurbished NVIDIA InfiniBand switches can be a smart option for AI, HPC, and data center projects. But buyers should check the details carefully.
Check | Why it matters |
Exact model | QM9700, QM9790, Q3400 and Q3200 are different |
Speed | 400G NDR and 800G XDR are not the same |
Connector type | OSFP / OSFP112 compatibility matters |
Airflow direction | Must match the rack cooling layout |
Power supplies | Production clusters should use redundant PSUs |
Rail kit | Often missing in refurbished stock |
Firmware | Must fit the cluster environment |
Port testing | OSFP ports should be tested |
Cable compatibility | Wrong cables can delay the project |
Warranty | Important for AI production systems |
Topology | Ask for a cabling plan or port map |
Do not buy only by price. For AI clusters, the safer buying logic is:
correct model + tested ports + correct airflow + correct cables + warranty + topology support.
Used and Refurbished NVIDIA InfiniBand Switches for Sale
server-parts.eu supplies refurbished enterprise hardware for AI, HPC, and data center projects.
We can help source:
NVIDIA Quantum-2 QM9700 switches
NVIDIA Quantum-2 QM9790 switches
NVIDIA NDR 400G InfiniBand switches
NVIDIA Quantum-X800 / 800G-class InfiniBand parts, depending on availability
NVIDIA ConnectX-7 adapters
NVIDIA ConnectX-8 parts, depending on availability
OSFP and OSFP112 cables
400G and 800G breakout cables
GPU cluster networking packages
refurbished AI server networking hardware
If you are building, expanding, or repairing an AI GPU cluster, send us the switch model, quantity, cable type, airflow direction, delivery country, and warranty requirement.
NVIDIA InfiniBand Switches
QM9700, QM9790, Quantum-2 and Quantum-X800
FAQ: NVIDIA InfiniBand Switches for AI GPU Clusters
What is the best NVIDIA InfiniBand switch for H100 clusters?
For many H100 clusters, the common choice is NVIDIA Quantum-2 QM9700 or QM9790 NDR 400G InfiniBand. These switches are designed for high-performance AI and HPC networking.
What is the best NVIDIA InfiniBand switch for H200 clusters?
H200 clusters often use the same class of networking as H100 clusters: NVIDIA Quantum-2 NDR 400G InfiniBand, usually with QM9700 or QM9790 switches.
What is the difference between QM9700 and QM9790?
Both are NVIDIA Quantum-2 NDR 400G InfiniBand switches. QM9700 is usually the managed version, while QM9790 is often used in externally managed fabrics.
What is NVIDIA Quantum-X800?
NVIDIA Quantum-X800 is NVIDIA’s newer XDR 800G InfiniBand platform for large AI clusters and newer Blackwell-generation infrastructure.
How much does an NVIDIA QM9700 switch cost?
Prices range from around €5,000–€10,000 for older HDR 200G InfiniBand switches, €25,000–€30,000 for NVIDIA Quantum-2 QM9700/QM9790 400G switches, and €60,000–€80,000 for newer NVIDIA Quantum-X800 800G XDR systems, depending on condition, availability, airflow, seller, and warranty.
Do B200 and B300 clusters need 800G InfiniBand?
Not always. It depends on the server platform, cluster size, and workload. Many current clusters still use 400G NDR, but newer Blackwell and Blackwell Ultra systems can move toward 800G XDR networking.
Can I use refurbished NVIDIA InfiniBand switches?
Yes. Refurbished NVIDIA InfiniBand switches can be a good option when the exact model, port condition, airflow, firmware, accessories, cables, and warranty are checked properly.
What does NDR mean?
NDR is an InfiniBand generation commonly associated with 400Gb/s networking. NVIDIA Quantum-2 is the main NVIDIA NDR InfiniBand switch platform for many H100 and H200 cluster designs.
What does XDR mean?
XDR is NVIDIA’s newer 800G InfiniBand generation used with Quantum-X800-class networking.
Are cables included when buying used NVIDIA InfiniBand switches?
Not always. Always ask clearly. For AI clusters, cables are critical and can be expensive. Ask for OSFP, OSFP112, AOC, DAC, breakout cables, and a full port map if needed.
What should I send when asking for an offer?
Send the required switch model, quantity, airflow direction, new or refurbished preference, cable type and length, server model, number of GPU servers, adapter type, delivery country, and warranty requirement.
Sources: NVIDIA InfiniBand Switches for AI GPU Clusters
NVIDIA Quantum-2 QM9700 Series datasheet: 64 ports of 400Gb/s NDR InfiniBand, 32 OSFP connectors, and 51.2Tb/s aggregate bidirectional throughput.
NVIDIA QM9700/QM9790 user documentation: Official NVIDIA user documentation for 1U NDR 400Gb/s InfiniBand switch systems.
NVIDIA QM9700/QM9790 specifications: Official specifications showing 40/56/100/200/400Gb/s per port and 32 OSFP connector cages.
NVIDIA Quantum-X800 product page: Official NVIDIA product page for 800Gb/s InfiniBand connectivity.
NVIDIA Quantum-X800 XDR cluster documentation: Official NVIDIA documentation for Quantum-X800 XDR cluster systems.
NVIDIA Q3400 / Q3200 XDR switch user manual: Official NVIDIA manual covering Q3400 systems, including 144 XDR ports over 72 OSFP cages.
Comments