DGX Spark Cluster Setup and Sizing
How to connect units with a QSFP cable and pool memory across them.
Two DGX Spark units connect directly with a QSFP cable, and NVIDIA says Cluster Assistant configures the ConnectX-7 network. At Layer3Labs, we build and operate AI systems inside other people's businesses. A cluster pools memory to run larger models.
Price and single-unit specifications are in our DGX Spark guide. NVIDIA lists the specifications on its product page.
How to Cluster Two DGX Sparks
You cluster two DGX Spark units by connecting their ConnectX-7 ports with a QSFP cable and running NVIDIA Sync Cluster Assistant. First, confirm that your target model needs more than one unit's memory. The ConnectX-7 NIC operates at 200 Gbps.
Cluster Assistant detects connected units, validates device configuration, and configures the ConnectX-7 network. NVIDIA says every node runs the same software stack.
Memory Pooling and Performance Across Connected Units
Clustering two 64 GB DGX Spark units pools memory to 128 GB across the 200 Gbps network. NVIDIA says this memory pooling expands model support to up to 200 billion parameters. A single DGX Spark lists 273 GB/s of unified memory bandwidth across a 256-bit interface. NVIDIA says two clustered units have twice the memory bandwidth.
In NVIDIA's Qwen 3.8 27B test, two clustered 64 GB systems delivered up to 1.7x the performance of a single system. The 1.7x figure comes from a two-unit test, and your results will depend on the model.
Clusters of Up to Four DGX Spark Units
ConnectX networking lets you connect up to four DGX Spark systems into a cluster. NVIDIA publishes specific model-size ceilings for each configuration. A single 64 GB DGX Spark supports models up to 100 billion parameters. A single 128 GB DGX Spark supports inference for models up to 200 billion parameters.
NVIDIA lists up to 400 billion parameters for two 128 GB units, which provide 256 GB of memory. Four 128 GB units create a 512 GB pool supporting models up to 700 billion parameters.
NVIDIA lists fine-tuning for models up to 70 billion parameters with 128 GB of memory.
What NVIDIA Has Not Published About Clusters
NVIDIA has not published cluster pricing or multi-node power draw figures. Each DGX Spark carries a 240-watt power supply. The GB10 superchip has a 140-watt TDP rating.
NVIDIA says the playbook Connect Multiple DGX Sparks for Distributed Workloads is coming soon to 64 GB devices. NVIDIA also lists Serve LLMs With vLLM and Run OpenClaw With a Local LLM as coming soon.
NVIDIA also names NemoClaw, OpenClaw, Hermes Agent and OpenShell playbooks on build.nvidia.com. NVIDIA Sync Model Launcher is scheduled for release at the end of October 2026. NVIDIA says it downloads and launches Qwen3.8 27B on one unit or a cluster, makes it accessible from laptops, and sets up OpenCode.
A Reddit post titled "From 1x3090 to 20 DGX Sparks: my house fuses were the first bottleneck" and another titled "For dual DGX spark users; GLM 5.3 flash got a 50%+ performance boost" exist. We did not verify either.
Cases Where DGX Spark Is the Wrong Answer
Small operations without technical staff are a poor fit for a DGX Spark cluster. In our opinion, local hardware fits when query volume is high or a policy rules out sending documents to a cloud API. Our answer would change if hosted prices dropped sharply.
Our local AI hardware calculator estimates the GPU, VRAM and RAM a model needs for a model size and use case. Layer3Labs also runs a free AI workflow audit.
Business Workloads for a DGX Spark Cluster
NVIDIA lists agentic AI, prototyping, fine-tuning, inference, data science, and edge applications as target workloads for DGX Spark. These are NVIDIA's listed workloads, not our test results. Edge applications listed by NVIDIA include Isaac, Metropolis, and Holoscan.
NVIDIA's blog lists three use cases: run an AI agent around the clock, run inference on a single DGX Spark while laptops do other work, and scale as work grows.
Preinstalled software includes DGX OS, NVIDIA Agent Toolkit, CUDA-X AI libraries, and Nemotron open models. Supported runtimes include Ollama, vLLM, and PyTorch with CUDA. The blog's get-started steps also name llama.cpp and LM Studio.
Our guide on how to use Ollama covers running models with Ollama.
Frequently Asked Questions
- You cluster two DGX Spark units by connecting their ConnectX-7 ports with a QSFP cable and launching NVIDIA Sync Cluster Assistant. The assistant detects connected units, validates device configuration, and configures the ConnectX-7 network. NVIDIA says every node runs the same software stack.
- Two clustered 64 GB DGX Spark units pool unified memory to 128 GB. Two 128 GB DGX Spark units pool memory to 256 GB. NVIDIA lists 273 GB/s of memory bandwidth for a single DGX Spark and says two clustered units have twice the memory bandwidth.
- Yes, ConnectX networking lets you connect up to four DGX Spark systems. Four 128 GB units pool to 512 GB of memory and support models up to 700 billion parameters. NVIDIA's 1.7x figure comes from a two-unit test.
- NVIDIA says two clustered units delivered up to 1.7x the performance of one in its Qwen 3.8 27B test with two 64 GB systems. Your results will depend on the model.
Deciding Whether to Cluster DGX Spark Units?
Layer3Labs runs a free AI workflow audit.
Book the Free AI Workflow Audit