DGX Spark vs. DGX Station: Which One Fits Your Work?

EPRI used two kinds of deskside NVIDIA system for two very different jobs. Its researchers ran a private, 120-billion-parameter technical-document assistant on a 128 GB Dell Pro Max with GB10, a DGX Spark-class machine. For a 50-member GenCast weather ensemble used in electric-grid research, they used DGX Station. In NVIDIA's account of EPRI's test, Station completed one timestep of that ensemble in about 150 seconds.

That is a more useful comparison than asking which box has the bigger headline number. Start with the model or experiment you need to run, how many people will use it, and the time they can wait for a result.

These are starting points for a comparison, not a claim that any published parameter ceiling determines your model's speed.

For local model development, the first question is whether the checkpoint, context cache and working data fit on one machine at a useful speed. EPRI's Power Chat shows that a carefully configured 120B-class model can work on a 128 GB GB10 system. It used gpt-oss-120b, vLLM and prefix caching to keep a large document cache on the local machine. That is evidence for EPRI's implementation, not a promise that every 120B checkpoint will fit a Spark.

Smaller labs are using Spark for work beyond chat. NVIDIA reports that NYU's Global AI Frontier Lab evaluates AI-generated radiology reports locally on DGX Spark, keeping imaging data inside the lab. Stanford researchers use Spark to prototype biological-agent workflows before moving larger runs to shared clusters. These are sensible starting points when a researcher needs a dedicated Linux AI environment without reserving a large system for every experiment.

Station comes into view when one job needs substantially more memory or when the team needs heavier shared capacity. At Carnegie Mellon University's Robotics Institute, researchers report fine-tuning a drone-detection model on Station and deploying the result to a Jetson Orin NX. At EPRI, the move from a document assistant on GB10 to a weather ensemble on Station followed the change in workload, not a simple rule about industry or team size.

DGX Spark uses a GB10 Grace Blackwell Superchip, a 20-core Arm CPU and up to 128 GB of unified memory. NVIDIA now lists 64 GB options through participating OEMs as well. The 64 GB and 128 GB systems are different buying choices; check the tier before using any model-size example. NVIDIA's published ceilings are up to 100B parameters on one 64 GB unit and up to 200B on one 128 GB unit, under suitable model and software conditions.

DGX Station uses a GB300 Grace Blackwell Ultra Desktop Superchip with a 72-core Grace CPU, one Blackwell Ultra GPU and 748 GB of coherent memory. That total combines 252 GB of GPU HBM3e and 496 GB of CPU memory. Data in those two memory types has different speed. NVIDIA quotes up to 20 PFLOPS of FP4 compute for Station and up to one PFLOP for Spark; those peaks do not predict a 20-fold gain for your application.

Ask your team for the checkpoint and version, precision or quantization, prompt and output length, concurrent requests and a response-time target. A test with those settings gives a far better buying answer than the parameter count on a product page.

Two 128 GB Sparks offer 256 GB in aggregate and can connect directly through a supported QSFP cable. Four provide 512 GB in aggregate; NVIDIA's guided four-node setup uses four links to a 200 GbE switch, while its manual playbook also documents a direct ring. The machines do not become one computer with 512 GB of flat memory. A distributed inference or training stack must divide the work and communicate across the network; an ordinary program running on one Spark still sees that one node's memory.

A pair or small cluster makes sense when the model has a tested multi-node recipe, several researchers need independent machines, or one workload can use the extra aggregate capacity. Station makes sense when a single large workload benefits from its coherent memory and GPU HBM, or the lab wants one managed shared node. NVIDIA also describes manual eight-Spark clusters, while its published NVIDIA Sync guided setup currently covers up to four. Treat a larger build as its own engineering project.

Compare the complete cost and measured performance of the proposed options. Include systems, cable or switch, installation, software work and support. Four Spark boxes can be useful; their combined memory figure alone does not establish that they can replace Station for a particular model.

Both platforms use Arm CPUs. DGX Spark ships with DGX OS; the current DGX Station platform uses Ubuntu with NVIDIA AI Developer Tools. Confirm that the exact containers, Python packages, licensed applications and model runtimes your team uses support the target platform. NVIDIA says Windows-based DGX Station systems are expected in December 2026, so teams tied to Windows software should request a Windows-specific OEM configuration and timeline.

Power can settle the placement question early. NVIDIA specifies a 1,600 W rating and a 20 A circuit for Station. Check the OEM plug, heat, clearance and service access with facilities. Spark is much smaller, but a cluster still needs power for every unit, the correct high-speed links, management access and a place for the switch if four nodes are planned.

The following are actual Uniqcli catalog choices, not one fixed bundle. The MSI Station listing is a barebones, request-pricing option; check the completed configuration before comparing its offer with ASUS or a Spark system.