How to Cluster Two NVIDIA DGX Spark 64GB Systems for 128GB of Combined Memory

Two NVIDIA DGX Spark 64GB systems can work together as a distributed AI cluster, providing 128GB of combined unified memory capacity across two machines. To connect them, you need a compatible QSFP112 cable, NVIDIA Sync Cluster Assistant, and supported distributed workloads such as vLLM. The systems retain separate memory spaces rather than becoming one physical 128GB machine.

If you are planning a two DGX Spark 64GB cluster, the most important part is understanding how the hardware, high-speed networking, and distributed software work together. Connecting the cable is only the first step. You must also configure communication between the systems and deploy an AI workload that supports multiple nodes.

This guide follows NVIDIA’s official DGX Spark documentation and its multi-node setup playbooks, updated for the November 2026 research date. It explains the physical connection, NVIDIA Sync configuration, network verification, distributed inference, and common setup problems.

Why Connect Two NVIDIA DGX Spark 64GB Systems?

A single DGX Spark provides 64GB of unified memory. Connecting two systems gives supported distributed applications access to 128GB of combined memory capacity.

This configuration is useful when a model or workload needs more resources than one DGX Spark can provide. Instead of replacing the existing system, developers can connect another machine and distribute supported workloads across both nodes.

What Can a Two-System DGX Spark Cluster Do?

A two-node cluster can support several types of distributed AI workloads.

• Run supported large language models across multiple systems.
• Distribute AI inference between two nodes.
• Experiment with multi-node model execution.
• Perform supported distributed fine-tuning workloads.
• Build and test multi-node AI applications.
• Explore distributed computing and model parallelism.

NVIDIA provides separate playbooks for distributed inference with vLLM, multi-node NCCL communication, and PyTorch-based fine-tuning.

128GB Combined Memory vs. 128GB Unified Memory

This distinction is important before building a cluster.

Configuration	Memory capacity
One DGX Spark 64GB	64GB
Two DGX Spark 64GB systems	128GB combined
One DGX Spark 128GB	128GB unified
Two DGX Spark 128GB systems	256GB combined

Two 64GB systems do not automatically behave like one 128GB system. Each machine maintains its own memory space. Distributed software must divide or coordinate the model and workload between the nodes.

For example, a distributed inference framework may place different model components on different systems. This allows the workload to use resources across both machines without creating a single hardware-level memory pool.

If you are still deciding which DGX Spark configuration to purchase, read our DGX Spark 64GB vs. 128GB comparison.

Who Should Build a Two-Node Cluster?

This configuration is particularly useful for:

• AI developers experimenting with larger models.
• Machine learning engineers exploring distributed inference.
• Researchers working with multi-node computing.
• Developers building local AI infrastructure.
• Enthusiasts who want to experiment with distributed model execution.

However, if your applications run comfortably on one 64GB system, adding another machine may introduce unnecessary complexity and cost.

What You Need Before Building a DGX Spark Cluster

Before connecting the two systems, prepare the networking hardware, software, and administrative access.

The following requirements are based on NVIDIA’s current multi-node playbook and Sync documentation.

Hardware Requirements

Component	Requirement
AI computer 1	NVIDIA DGX Spark 64GB
AI computer 2	NVIDIA DGX Spark 64GB
Networking adapter	Integrated ConnectX-7
Connection cable	Compatible QSFP112 DAC
Cable quantity	One
Management network	Working LAN connection
Network switch	Not required for two directly connected systems

Both systems should be powered on and accessible through the same management network.

Software Requirements

Prepare the following:

• Current DGX Spark system software on both devices.
• NVIDIA Sync installed on a supported laptop or computer.
• SSH access to both DGX Spark systems.
• A user account with sudo privileges on each system.
• A compatible workload framework for the intended AI application.

NVIDIA’s current Cluster Assistant requirements specify the April 2026 DGX OS release or later. Check the latest release notes before beginning because the software requirements can change.

Which QSFP Cable Should You Buy?

The cable is an important part of the setup. A connector that physically fits does not automatically mean the cable is compatible.

NVIDIA identifies these approved cable models:

Manufacturer	Model	Specification
Amphenol	NJAAKK-N911	QSFP112, 400mm, 32AWG
Amphenol	NJAAKK0006	0.5m version
Luxshare	LMTQF022-SD-R	QSFP112 400G DAC, 400mm, 30AWG

The DGX Spark QSFP ports support Ethernet configuration and provide up to 200Gbps per port. A cable rated for 400Gbps does not increase the port’s maximum speed.

My recommendation: Choose one of NVIDIA’s documented cable models rather than buying a generic QSFP cable based only on its advertised speed.

You can check the manufacturer’s part numbers through the NVIDIA DGX Spark networking documentation.

I have not verified an eligible Amazon listing for these exact approved models, so I am not inserting an unverified Amazon affiliate link.

How ConnectX-7 Networking Works on DGX Spark

NVIDIA uses ConnectX-7 networking to provide high-speed communication between connected DGX Spark systems.

Each system has two QSFP ports on its rear panel. These ports connect to the integrated ConnectX-7 network interface controller.

Each port provides up to 200Gbps, subject to the cable and connection configuration.

Understanding RoCE and RDMA

RoCE stands for RDMA over Converged Ethernet. It enables remote direct memory access over Ethernet networks.

RDMA allows supported applications to exchange data between systems with less CPU involvement than conventional networking approaches.

This matters for distributed AI workloads because model execution can require frequent communication between nodes.

The ConnectX-7 networking architecture also has a specific PCIe layout. Each physical QSFP port exposes two Linux Ethernet interfaces, with corresponding RoCE interfaces. This is one reason manual configuration can become complicated.

Do You Need a Network Switch?

No. Two DGX Spark systems can connect directly using one compatible QSFP cable.

Cluster size	Connection method
Two systems	One direct QSFP cable
Three systems	Three-cable direct ring or supported switch topology
Four systems	External switch for NVIDIA Sync

For a two-node cluster, direct connection is the simplest option. You do not need to purchase an expensive 200GbE switch.

How to Physically Connect Two DGX Spark Systems

Now you can connect the two computers.

Follow these steps using NVIDIA’s documented direct-link configuration.

Step 1: Prepare Both Systems

Power on both DGX Spark systems.

Make sure:

1. Both systems complete their startup process.
2. Both have working management network connections.
3. Both run the required DGX OS version.
4. You can access each system using an account with sudo privileges.

Use DGX Dashboard to check for and install available system updates.

Step 2: Identify the QSFP Ports

Turn the rear side of each DGX Spark toward you.

Each system has two QSFP ports:

• Port 0 is closest to the ordinary Ethernet port.
• Port 1 is farther away.

For the direct connection, NVIDIA’s playbook supports connecting port 0 to port 0 or port 1 to port 1. Choose the same port on both systems.

Step 3: Connect the QSFP Cable

Follow this procedure:

1. Hold the QSFP cable with its pull tab facing upward.
2. Align the connector with the selected port on the first DGX Spark.
3. Push it into the port until fully seated.
4. Connect the other end to the corresponding port on the second system.
5. Check that both connectors are properly seated.

Never force the connector into the port. If it does not slide in smoothly, check its orientation and alignment.

Important: One cable is sufficient for a two-system direct connection. Adding a second cable between the same systems does not increase performance in NVIDIA’s documented topology.

Step 4: Keep the Management Network Connected

The QSFP cable handles high-speed communication between the two systems.

Keep their ordinary Ethernet or Wi-Fi management connections available. NVIDIA Sync needs to reach the devices through the management network or a supported remote connection.

The management network and ConnectX-7 cluster network serve different purposes.

How to Configure NVIDIA Sync Cluster Assistant

NVIDIA Sync Cluster Assistant is the recommended method for configuring a supported DGX Spark cluster.

It automatically configures the ConnectX-7 network, checks the hardware and software requirements, and establishes inter-device SSH communication.

You do not need to manually configure the complicated PCIe network topology for the standard two-system setup.

Step 1: Install NVIDIA Sync

Install NVIDIA Sync on a supported laptop or computer.

NVIDIA currently provides installation options for:

• Windows
• macOS
• Ubuntu

Download the appropriate installer from the official NVIDIA Sync documentation.

Complete the installation and open the application.

Step 2: Add Both DGX Spark Systems

Your laptop must be able to reach both systems through the management network.

For each DGX Spark:

1. Open NVIDIA Sync.
2. Select Add New.
3. Look for the device name in the available device list.
4. If it does not appear, select Add device manually.
5. Enter the device name or management IP address.
6. Enter the username and password.
7. Select Add.

Repeat this process for the second DGX Spark.

Both devices should now appear in NVIDIA Sync.

Step 3: Open Cluster Assistant

After adding both devices:

1. Open NVIDIA Sync Settings.
2. Select Cluster Assistant.
3. Choose Add New Cluster.
4. Enter a name for the cluster.
5. Select both physically connected DGX Spark systems.

NVIDIA Sync will now begin checking the devices and their connection.

Step 4: Complete the Device Checks

Cluster Assistant checks several requirements before changing the network configuration.

Check	What it verifies
SSH access	Whether NVIDIA Sync can access both devices
Hardware	Whether the systems are supported GB10 devices
Software	Whether the required system version is installed
Sudo access	Whether network configuration permissions are available
User information	Whether account differences could affect workloads
Physical link	Whether the expected network connection exists

Resolve any failed check before proceeding.

If a system requires a sudo password, NVIDIA Sync may ask you to enter it. Matching usernames and user IDs are optional, although consistent accounts can simplify file and workload management.

Step 5: Confirm the Network Configuration

After the checks pass, NVIDIA Sync presents the proposed network configuration.

Review the detected layout and confirm that it matches the physical connection.

The application then configures the ConnectX-7 network.

Do not interrupt this process while the network configuration is being applied.

Step 6: Establish Inter-Device SSH

NVIDIA Sync also configures key-based SSH communication between the connected systems.

This is different from the SSH connection used to access the devices through your management LAN.

The cluster uses this communication path for managing processes across nodes.

NVIDIA notes that the SSH configuration step can take several minutes. If it times out after five minutes for a device, retry the step.

Step 7: Save the Cluster Information

When NVIDIA Sync displays the successful setup screen:

1. Copy the displayed network information.
2. Save the details in a text file.
3. Keep the cluster information available for workload configuration.
4. Continue to the relevant NVIDIA workload playbook.

Network configuration complete

At this point, the ConnectX-7 network and inter-device SSH are ready. Your distributed AI workload still needs separate configuration.

How to Verify Your DGX Spark Cluster

A successful Cluster Assistant setup is a good starting point, but it is worth checking the actual network configuration before deploying a model.

NVIDIA provides a dedicated network inspection guide for this purpose. NVIDIA Sync User Guide

Step 1: Inspect the Network Configuration

On either DGX Spark, open a terminal and run:

sudo ls /etc/netplan/99-nvidia-sync-cluster.yaml

This checks whether NVIDIA Sync created the expected cluster configuration file.

Display its contents:

sudo cat /etc/netplan/99-nvidia-sync-cluster.yaml

Look for the configured Ethernet interfaces and their assigned addresses.

Do not manually edit this file unless you understand the network configuration and have a recovery plan.

Step 2: Check the Network Interfaces

Run:

ip -br link

This displays the network interfaces and their current states.

Look for the ConnectX-7 interfaces identified in the Netplan configuration.

The important states are:

• LOWER_UP: The physical network link is active.
• NO-CARRIER: No physical link is detected on that interface.

Not every interface must be active. The expected state depends on the selected cluster topology. NVIDIA Sync User Guide

Step 3: Check Assigned IP Addresses

Run:

ip -br addr

Compare the displayed addresses with the ConnectX-7 network configuration.

Make sure the interfaces and assigned addresses correspond to the network plan.

Step 4: Test Communication Between Nodes

Identify the peer cluster IP address from the configuration on the other system.

Then run:

ping -c 3 <peer-cluster-ip>

Replace <peer-cluster-ip> with the actual IP address assigned to the connected peer interface.

Repeat the appropriate connectivity check from the other node.

Do not assume that every interface on one machine must communicate directly with every interface on the other. Check the expected peer connections for your topology. NVIDIA Sync User Guide

How to Run AI Inference Across Two DGX Spark Systems

Once networking works, you can configure a distributed AI workload.

One important option is vLLM. It supports model serving and distributed execution when configured with compatible hardware and software.

NVIDIA provides a dedicated playbook called Serve LLMs with vLLM Across DGX Sparks.

What You Need for Distributed vLLM

NVIDIA’s current two-node workflow identifies these requirements:

• Two compatible DGX Spark systems.
• Completed cluster networking and inter-device SSH.
• NVIDIA Sync installed on the management computer for the guided path.
• Docker installed and available to the user on both systems.
• Authenticated Hugging Face CLI on both systems for the intermediate and advanced paths.
• The uv package manager on the selected head node for those paths.

The command-line examples currently cover two systems. NVIDIA also specifies that systems in the same cluster should use the same memory configuration, such as two 64GB systems.

Choose a Model That Fits the Distributed Configuration

Model size alone does not determine whether a workload will run.

You must consider:

• Model architecture.
• Model precision.
• Memory requirements.
• Context length.
• KV cache requirements.
• Distributed execution support.
• Communication overhead.

A model that does not fit on one system may become usable when its execution is distributed across multiple systems, provided the software supports that configuration.

Follow NVIDIA’s Official vLLM Workflow

NVIDIA provides two implementation paths.

Intermediate path

• Use NVIDIA Sync to configure the cluster.
• Select a compatible model recipe.
• Use NVIDIA’s provided scripts.
• Configure the containerized vLLM deployment across two systems.

Advanced path

• Use the manual configuration workflow.
• Configure the cluster and SSH.
• Use NVIDIA’s provided scripts and model recipes.
• Configure the vLLM launch options, including tensor parallelism.

The official two-system example uses a Qwen model recipe from NVIDIA’s spark-vllm-docker project. The documented model is nvidia/Qwen3.8-27B-NVFP4.

I recommend following the exact model recipe and software versions in NVIDIA’s official vLLM cluster playbook rather than copying generic multi-node vLLM commands. The launcher, container settings, and model configuration must work together.

Common DGX Spark Cluster Problems and Solutions

Even with compatible hardware, a cluster can fail during discovery, network configuration, or workload deployment.

The following diagnostic table follows NVIDIA’s current troubleshooting guidance.

Problem	Possible cause	Recommended action
DGX Spark does not appear in NVIDIA Sync	Device is offline or inaccessible	Check power and management LAN access
SSH check fails	Incorrect access or SSH configuration	Verify direct SSH access
GB10 check fails	Unsupported device selected	Select only supported DGX Spark systems
Software check fails	Outdated DGX OS	Update through DGX Dashboard
Password check fails	Missing sudo access	Verify user permissions
Incorrect network layout	Loose cable or unsupported topology	Check physical connections
Link remains down	Cable or port problem	Reseat the cable and check port status
SSH setup times out	Configuration process exceeds timeout	Retry the SSH step
Slow data transfer	Network errors or poor switch configuration	Inspect link errors and hardware offload where applicable

Problem 1: NVIDIA Sync Cannot Detect a DGX Spark

Start by checking whether the system is powered on and connected to the management network.

Confirm that your laptop can reach the system through its management IP address or supported device name.

If necessary, test SSH access directly.

Problem 2: ConnectX-7 Interfaces Are Not Active

Check the cable and port connections first.

Then inspect the interfaces:

ip -br link

If the expected physical link remains down, reseat the cable and check the selected ports.

NVIDIA’s troubleshooting guide also recommends checking the physical layout and rebooting the systems when no ConnectX-7 interface becomes active.

Problem 3: Network Configuration Does Not Match

If Cluster Assistant identifies an incorrect topology, verify:

1. Both ends of the cable are fully seated.
2. The same port selection is used on both systems.
3. You are not mixing direct and switch connections.
4. The systems are running supported software.
5. The expected physical links are active.

If the configuration remains incorrect, review NVIDIA’s network inspection documentation before changing Netplan settings manually.

Problem 4: SSH Setup Times Out

Check that both systems remain accessible.

Verify sudo permissions and retry the SSH setup stage through NVIDIA Sync.

Avoid manually deleting SSH keys or changing network files without understanding the effect on the cluster.

Problem 5: The Cluster Connects but Inference Does Not Work

A successful network connection does not guarantee a successful AI deployment.

Check the workload configuration separately:

• Are both nodes included in the model deployment?
• Are the model files accessible?
• Is the correct model recipe selected?
• Are Docker and other prerequisites installed?
• Are the distributed execution parameters correct?
• Are both systems using compatible software and hardware configurations?

Use the official NVIDIA vLLM troubleshooting instructions for deployment-specific problems.

Limitations of a Two-System DGX Spark Cluster

Connecting two systems provides additional resources, but it does not remove every hardware and software limitation.

Distributed Memory Is Not a Single Memory Pool

The two machines retain separate 64GB memory spaces.

A distributed framework must manage how model components, tensors, and other data are placed across the nodes.

An application designed for one GPU or one machine cannot automatically use the additional memory just because another DGX Spark is connected.

Network Communication Adds Overhead

Distributed inference requires communication between systems.

Depending on the model and workload, communication can become a performance bottleneck.

A workload that spends much of its time exchanging data may not benefit as much as one that can distribute computation efficiently.

Two Systems Do Not Guarantee Twice the Performance

Adding another DGX Spark does not automatically double inference speed.

Actual results depend on:

• Model architecture.
• Workload distribution.
• Network performance.
• Tensor parallelism.
• Batch size.
• Context length.
• Software implementation.

The benefit may be greater model capacity rather than higher token generation speed.

When Should You Choose One 128GB DGX Spark?

Consider a single 128GB system if your priority is a larger unified memory capacity within one machine and you want a simpler deployment.

Choose two 64GB systems if you specifically need to experiment with distributed computing, multi-node inference, or workloads that benefit from multiple connected machines.

The best choice depends on the workload, not simply the combined memory number.

FAQs: Two DGX Spark 64GB cluster

Build Your DGX Spark Cluster with the Right Configuration

Connecting two NVIDIA DGX Spark 64GB systems is a practical way to explore distributed AI computing and workloads that need more resources than one machine can provide.

Start with NVIDIA’s approved QSFP112 cable, prepare both systems, and use NVIDIA Sync Cluster Assistant to configure the network. After verifying communication, follow an official workload playbook to deploy distributed inference or fine-tuning.

The most important decision is choosing software that can actually use both systems. Combined memory capacity is useful only when the application knows how to distribute its workload.

For the next step, review the official NVIDIA multi-node setup guide and vLLM cluster playbook before deploying your first model.

Related MetaDictory guides:

• NVIDIA DGX Spark 64GB
• DGX Spark 64GB vs. 128GB
• Will RAM Prices Fall in 2027?

Sources and further reading

1. NVIDIA DGX Spark User Guide – ConnectX-7 Networking
2. NVIDIA Sync – Cluster Assistant
3. NVIDIA Build – Connect Multiple DGX Sparks
4. NVIDIA Build – Serve LLMs with vLLM Across DGX Sparks

Editorial note: This article is based on official NVIDIA documentation and published technical playbooks. It does not claim independent hands-on testing of a two-system cluster. Software instructions and compatibility should be rechecked against NVIDIA’s latest releases before publication.