DGX Station vs DGX Spark: GB300 or GB10, and the RTX PRO 6000 workstation between them

DGX Station vs DGX Spark: what separates them

DGX Spark and the DGX Station are NVIDIA’s two desktop AI systems, built on different chips for different jobs. DGX Spark is a 1.2 kg box with the GB10 chip and 128 GB of unified memory at 273 GB/s, powered by a 240 W external supply, and NVIDIA builds it for developers who prototype and fine-tune models at their desk. The DGX Station is a deskside tower with the GB300 chip and 748 GB of coherent memory, of which 252 GB is HBM3e at 7.1 TB/s. NVIDIA gives it 1,600 W of total system power and positions it as a desktop for one user or a compute node shared by a team.

DGX Spark comes as NVIDIA’s Founders Edition and as GB10 systems from other manufacturers, while the DGX Station is sold only as manufacturers’ systems. Between them sits a workstation with one to four RTX PRO 6000 cards, which serves some teams better than either.

GB300 vs GB10 specifications side by side

The table compares both systems as NVIDIA and Dell specify them in October 2026.

NVIDIA DGX Spark product page, hardware overview and clustering guide (the two guides updated 10 September 2026); NVIDIA DGX Station product page and datasheet (August 2026); Dell Pro Max with GB300 product page. Read on 9 October 2026.

On DGX Spark the CPU, DGX OS and the GPU draw on one 128 GB pool, and our article on what fits in 128 GB on a DGX Spark works out how much of it a model can use. The DGX Station has two tiers of memory. Its GPU has 252 GB of HBM3e at 7.1 TB/s, about 26 times the bandwidth of DGX Spark, and the Grace CPU has 496 GB of LPDDR5X at 396 GB/s. NVIDIA’s announcement of 18 March 2025 gave 784 GB of coherent memory; the product page and the August 2026 datasheet give 748 GB, so comparisons that quote 784 GB rest on the 2025 figure.

FP4 petaFLOPS, sparsity and token generation speed

Both FP4 headlines are sparsity figures. NVIDIA’s DGX Spark hardware overview gives “up to 1 PFLOP (petaFLOP) at FP4 precision with sparsity”, and we found no dense FP4 figure for GB10 from NVIDIA. For the DGX Station NVIDIA quotes 20 PFLOPS of FP4, and its datasheet states that “All Tensor Core specifications are with sparsity unless otherwise noted.” On the sparse figures the DGX Station has 20 times the FP4 compute of DGX Spark, though standard LLM serving does not use the 2:4 structured sparsity behind them.

Token generation depends on memory bandwidth rather than compute, because each new token reads the weights it uses from memory. A dense 70B model in FP8 reads about 70.6 GB per token. Dividing bandwidth by that figure gives an upper bound for one user of about 3.9 tokens per second on DGX Spark. The same arithmetic gives about 100 on the DGX Station, provided the weights sit in HBM3e. Measured rates stay below such bounds, and our DGX Spark benchmark article collects the published measurements.

NVIDIA says the DGX Station supports models of up to 1 trillion parameters, against up to 200 billion on one 128 GB DGX Spark, and by our arithmetic both figures assume weights of about 4 bits. At 4.5 bits per weight, the size of NVFP4 with its block scales, 1 trillion parameters take about 560 GB by our arithmetic. That is more than twice the HBM3e, so most of such a model sits in the LPDDR5X tier, whose bandwidth NVIDIA gives as 396 GB/s. Our article on DGX Station GB300 alternatives covers what the two tiers mean for inference.

Power, size, networking and remote management

DGX Spark runs from a 240 W external supply. NVIDIA rates the GB10 chip at 140 W TDP and leaves 100 W for the other components, so one Spark plugs into an ordinary office socket. The DGX Station needs a circuit planned for it. NVIDIA gives 1,600 W of total system power, and its US product page adds that a 20 A circuit is required. HP requires a dedicated 200 to 240 V AC circuit for its version, the ZGX Fury AI Station, and describes liquid cooling and a chassis that stands as a tower or mounts in a 5U rack.

DGX Spark has a ConnectX-7 controller with two QSFP ports at up to 200 Gb/s each. NVIDIA’s clustering guide supports up to three DGX Spark connected directly through cables and up to four through a switch, and our two-node DGX Spark article covers the cabling and what a second Spark adds. The DGX Station carries a ConnectX-8 SuperNIC rated at up to 800 Gb/s, and NVIDIA supports linking up to two DGX Stations.

The DGX Station has a BMC for system management on its own 1 GbE port, as a server does. NVIDIA lists no BMC for DGX Spark, which is managed through its operating system; when it hangs, someone next to it has to reset it, or a remotely switchable power outlet has to power-cycle it.

Operating system, team use and how each is sold

DGX Spark ships with NVIDIA DGX OS, NVIDIA’s Ubuntu-based operating system. For the DGX Station, NVIDIA lists Ubuntu with NVIDIA AI Developer Tools, which Dell and HP ship on Ubuntu 24.04 LTS. Both systems have Arm processors, so check any binary-only dependency before you move a toolchain. In March 2025 NVIDIA said its software lets DGX Spark users move models to DGX Cloud or data-centre infrastructure “with virtually no code changes”, and its DGX Station datasheet speaks of deploying to the cloud or data centre “using the same tools, libraries, frameworks, and pretrained models”.

NVIDIA describes the DGX Station as “an individual desktop for one user” or “a centralized compute node for multiple team members”, and MIG splits its GPU into up to seven instances. NVIDIA’s DGX Spark specification lists no MIG. One Spark can still serve a team through a serving engine that batches requests, and the benchmark article works out which team sizes fit.

DGX Spark is sold as NVIDIA’s Founders Edition, with 128 GB and a 4 TB self-encrypting drive, and as GB10 systems from seven manufacturers, among them Acer, Dell, HP, Lenovo and MSI. NVIDIA also lists a 64 GB configuration, sold only through participating manufacturers. The DGX Station comes only from manufacturers: NVIDIA’s page lists seven, among them Dell, Exxact, HP, MSI and Supermicro, and directs buyers to them.

We supply the DGX Spark Founders Edition on one EU contract and invoice. Tell us how many developers will use it and which models they work with, and we reply whether one Spark, two DGX Spark or four fit the work.

DGX Spark alternative for teams: 1 to 4 RTX PRO 6000 in a workstation

The step between the two is a workstation with one to four RTX PRO 6000 Blackwell cards. Each card has 96 GB of GDDR7 with ECC at 1,792 GB/s, 6.6 times the bandwidth of DGX Spark. By the arithmetic above, that allows about 25 tokens per second for one user of a dense 70B model in FP8. The Workstation Edition draws up to 600 W. The Max-Q keeps the same memory and bandwidth at 300 W, and NVIDIA says it “enables up to four GPUs in a single system”, 384 GB in total, with up to four isolated MIG instances per card. The platform is x86, so it runs the operating systems and tools a company already manages.

It serves a team better than DGX Spark when several people use a model every day and generation speed is the limit. One card holds Llama 3.3 70B in FP8 with an FP8 cache for about twelve 8K conversations at once, and four cards run four independent copies of gpt-oss-120b. Against the DGX Station, the workstation is the better fit when every model copy fits on one 96 GB card. Two to four cards then serve copies side by side, a failed card stops one copy rather than the whole service, and sixteen MIG instances against seven give more developers a GPU of their own.

The workstation is the weaker choice when one model outgrows a card. The RTX PRO 6000 has no NVLink, so a larger model is split over PCIe, and a model larger than 384 GB does not fit at all, which makes it a case for the DGX Station or a GPU server with more cards. For one developer, DGX Spark remains the smaller step in power and space. A tower with four Max-Q cards draws about 1.8 to 2 kW at the wall by our estimate, more than the 1,600 W NVIDIA gives for the DGX Station. Our article on the four-card RTX PRO 6000 Max-Q workstation covers the towers, the PCIe lanes and the power supplies.

The choice also depends on where the models will run in production. On x86 servers with RTX PRO 6000 Server Edition cards, a workstation with the same GPU keeps development close to production; on NVIDIA’s GB200 or GB300 systems, DGX Spark or the DGX Station keeps it on NVIDIA’s Arm-based platform.

We build GPU workstations to order on Threadripper PRO or Xeon W platforms with one to four RTX PRO 6000 cards. Send us the models and the number of users through the form below, and the configuration and quote follow within one business day.

DGX Spark, RTX PRO 6000 workstation or DGX Station: which fits

The table sorts the choice by need; where several rows apply to one team, the largest model decides.

Our reading of NVIDIA’s DGX Spark, DGX Station and RTX PRO 6000 pages as of 9 October 2026; the 2 kW figure is the estimate from our four-card workstation article.

Where the model is still changing, one Spark is enough to start, and an RTX PRO 6000 workstation follows once a team uses the result every day.

What we supply

We supply the NVIDIA DGX Spark Founders Edition (128 GB, 4 TB) across the EU on one contract and invoice, with manufacturer warranty, for example two DGX Spark for two developers or four for a team. We also supply the RTX PRO 6000 Blackwell in its Workstation, Max-Q and Server editions, as cards or in workstations and AI servers built to order, assembled and burn-in tested. With the models and the number of concurrent users, we propose a configuration and quote within one business day. The DGX Station is sold by the manufacturers NVIDIA lists, and we do not supply it.

FAQ