Can You Upgrade NVIDIA DGX Spark? Memory, Storage, Clustering and RTX Spark Paths

What is fixed and what isn't

Like laptops with unified memory, DGX Spark's CPU, GPU and memory are one package.

Upgrade 1: more storage for models

Quantized models are tens of gigabytes each, so storage fills quickly. The simplest upgrade is an external NVMe drive holding your model library, with the most-used models kept on the internal drive for fastest load times.

Upgrade 2: a second unit (clustering)

NVIDIA supports connecting two DGX Sparks directly via ConnectX-7 so they act as one system with 256 GB total. This lets you run models up to roughly 405B parameters at 4-bit, or split agents across two machines. Use a supported direct cable and follow NVIDIA's clustering playbook; see our Clusters page for topologies.

Upgrade 3: software

Keep DGX OS, drivers, CUDA and your runtime updated. New quantization formats such as NVFP4 and runtime improvements regularly add meaningful speed.

• Update DGX OS through its built-in updater.
• Re-pull optimised model builds after runtime updates.
• Re-run a quick benchmark after each update to catch regressions.

When to move on instead

If you need Windows or a portable machine, an RTX Spark laptop is the natural sideways move. If you need much faster tokens on big models, a discrete-GPU workstation or DGX Station class system is the step up. Wait for the next Spark generation (Vera Rubin) if your current unit still meets your needs.

Upgrade options ranked by value

From cheapest and most useful to most expensive.

How to cluster two DGX Spark units

Two units link directly over their ConnectX-7 ports. Once connected, inference software that supports multi-node runs (such as vLLM or TensorRT-LLM) can split one model across both machines.

• Use the high-speed cable NVIDIA recommends for the ConnectX-7 link.
• Install the same DGX OS and driver versions on both units.
• Set up passwordless SSH between the two machines.
• Start with a model that clearly doesn't fit on one unit to confirm the split works.

When to stop upgrading and switch

If you regularly need more than 256 GB of memory, more bandwidth for fast generation, or many users at once, a larger workstation or cloud GPUs will serve you better than a stack of DGX Spark units. For portability, an RTX Spark laptop complements a DGX Spark rather than replacing it.

Keeping DGX Spark up to date

Install DGX OS and driver updates as they are released and re-check speeds afterwards. Our TLS score tracks how often NVIDIA and partners ship updates, which is a good guide to how long a machine stays current.