NVIDIA DGX Spark alternatives: workstations with RTX PRO Blackwell cards from 16 to 48 GB

DGX Spark alternatives built as workstations with one to four cards

The workstation alternatives to an NVIDIA DGX Spark in our range are builds with NVIDIA RTX PRO Blackwell cards: one or two NVIDIA RTX PRO 5000 Blackwell with 48 GB, two to four NVIDIA RTX PRO 4500 Blackwell Workstation Edition with 32 GB or NVIDIA RTX PRO 4000 Blackwell with 24 GB in a single slot, and one NVIDIA RTX PRO 2000 Blackwell with 16 GB for small models. The Spark holds one large model in 128 GB of unified memory at 273 GB/s. The cards have less memory each, 16 to 48 GB, at 288 to 1,344 GB/s, so a model within a card's budget is generated from memory with up to about five times the bandwidth, and a model above it has to be split or does not fit.

The DGX Spark Founders Edition is a 1.2 kg desktop system on the GB10 chip, in which 20 Arm cores and a Blackwell GPU share 128 GB of LPDDR5x under DGX OS. Its figures and the published measurements are in our articles on the DGX Spark specs, the DGX Spark benchmarks and what fits in its 128 GB; this page adds nothing to them and looks at the workstation side.

NVIDIA datasheets 5349550 (RTX PRO 5000, June 2026), 5108623 (RTX PRO 4500 Workstation Edition, April 2026), 5323450 (RTX PRO 4000, June 2026) and 5322351 (RTX PRO 2000, June 2026); DGX Spark figures as on our DGX Spark page; usable memory by our sizing rule, our estimate.

Usable memory is the budget for weights plus KV cache by our sizing rule: 0.9 times the memory the driver reports, less 3 GiB per card. The reported memory is scaled from the 95.6 GiB the driver shows on a 96 GB card; that scaling is an assumption, and the driver on the delivered card decides. For the Spark NVIDIA publishes no usable-memory figure. The memory fractions of 0.8 and 0.9 in NVIDIA's vLLM playbooks give roughly 102 to 115 GB of the pool for weights and cache, and the 0.7 of its launch script for one Spark about 90 GB.

128 GB at 273 GB/s or 16 to 48 GB per card at up to 1,344 GB/s

To generate one token for one user the GPU reads the active weights once, so the rate of a single conversation is bounded by the memory bandwidth. On a model that fits, the RTX PRO 5000 has a generation ceiling about five times that of the Spark, 1,344 GB/s against 273 GB/s. By our division of the datasheet figures, the 896 GB/s of the RTX PRO 4500 Workstation Edition is about 3.3 times and the 672 GB/s of the RTX PRO 4000 about 2.5 times the Spark's bandwidth, and the 288 GB/s of the RTX PRO 2000 is about level with it. These are ratios of bandwidth. No published LLM token rate that we can cite exists for these four cards, so we print none.

For the Spark, rates have been published. NVIDIA measured gpt-oss-120b at 55.4 tokens per second for one request, and LMSYS measured Llama 3.1 70B in FP8 at 2.7 tokens per second for one user in October 2025. Both show what the 128 GB are for: gpt-oss-120b has 60.8 GiB of weights and a 70B model in FP8, such as Llama 3.3 70B, has 67.7 GiB; neither fits a single 48 GB card. NVIDIA's statement of models up to 200 billion parameters is a 4-bit claim, 106 to 112 GB with the block scales.

A workstation reaches such models only by splitting them. Each card addresses its own memory, NVIDIA's datasheets for these cards list no NVLink, and the serving software divides the model over PCIe. By our memory arithmetic two RTX PRO 5000 48 GB cards hold Llama 3.3 70B in FP8 as one tensor-parallel copy with 9 conversations of 8,192 tokens and an FP8 KV cache, and gpt-oss-120b with 17 conversations of 32,768 tokens and a 16-bit cache; four RTX PRO 4500 Workstation Edition cards hold 28 and 37. We found no consistent published measurement of tensor-parallel scaling on these cards, so the counts say what fits and nothing about speed. Where one model of that size is the task, the Spark holds it in one pool, as RTX PRO 5000 72 GB vs DGX Spark works out for the 72 GB card, which we also supply.

Development and prototyping with open models

For this task the question is which files load on one card with room for context. The table uses publishers' own files with their sizes on Hugging Face, read 10 November 2026.

File sizes from Hugging Face, read 10 November 2026; classes by our sizing rule, not a measurement. Every file that fits a smaller card also fits the larger ones.

We do not plan on a file that is above our budget, even where it would still load. All four cards are Blackwell cards, so NVFP4 files compute natively on them: NVIDIA's NVFP4 checkpoint of Qwen3.8-27B, 20.4 GiB, fits one 32 GB card with about 5.3 GiB left. On the RTX PRO 5000, MIG gives two developers an isolated 24 GB instance each; it needs Linux, driver 575.57.08 or later, vBIOS 98.02.73.00.00 and the display mode set to compute. The smaller cards are covered in local LLMs on 16 GB cards and the RTX PRO 4000 Blackwell for local LLMs.

You can order the workstation alone, the setup of vLLM, Ollama or ComfyUI alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

A small team assistant on one workstation

A team assistant is sized by the conversations held at the same moment. The counts below are conversations of 8,192 tokens with an FP8 KV cache.

Our estimate, memory-bound, not a measurement: (usable memory of the copy − weights) ÷ KV cache per conversation (0.625 and 1.0 GiB), rounded down; weights 15.2 and 32.0 GiB. TP=2 is a tensor-parallel copy over two cards.

A conversation here is a request in flight, and a team produces far fewer of them than it has members. The example values of our guide to a private ChatGPT server by company size give about 4 requests in flight per 100 employees, so by memory every build in the table carries a small team with one 14B-class model. On one RTX PRO 4500 Workstation Edition, Qwen3-32B in NVFP4 (19.3 GiB) holds 6 such conversations by our estimate. For a dense 70B model in FP8 on the Spark, our DGX Spark page reads the 2.7 tokens per second as fine for one developer and not for twenty people on a chat assistant.

Fine-tuning small models

NVIDIA's fine-tuning playbook for the Spark shows a full fine-tune of a 3B model, LoRA on an 8B model and QLoRA on a 70B model. The cards cover the middle step. By the figures on our workstation pages, LoRA on an 8B model takes about 16.7 GB of model states, which leaves about 7 GB on a 24 GB card and about 15 GB on a 32 GB card. QLoRA on a 70B model takes 42.8 GB and fits no single 24 GB or 32 GB card; on a 48 GB card about 5 GB would remain by our subtraction from the nominal 48 GB, and we do not plan on it. That job stays with the Spark.

Image generation with ComfyUI

The table counts model file, text encoders and VAE together against 0.9 times the card's memory. It is arithmetic from the file sizes; nothing here is measured, and we give no speed.

File sizes from Hugging Face, read 10 November 2026; our arithmetic. "all" = model, text encoders and VAE fit together; "model only" = the model file fits and the text encoder is unloaded or kept in system memory; "offload" = the model file alone exceeds 0.9 times the card's memory. The last column is the RTX PRO 4500 Workstation Edition.

The RTX PRO 5000 48 GB holds every row completely. FLUX.2 klein 4B is published under Apache 2.0 and SDXL under OpenRAIL++ with its use restrictions; Stable Diffusion 3.5 comes under the Stability AI Community License with its own conditions. The files and licences are set out in ComfyUI on the RTX PRO 2000, 4000 and 4500.

Windows, CAD software and displays on the same machine

The Spark runs DGX OS, which is Ubuntu 24.04 for Arm, and our DGX Spark page advises checking any binary-only dependency for an Arm build. A workstation with these cards is an x86 machine with the operating system IT already manages. NVIDIA's RTX Enterprise release notes 621.14 of 4 November 2026 list all four cards for Windows 11 and for Windows 10 22H2, with Release 620 as the last branch for Windows 10; the Linux driver lists contain them as well.

The same cards drive the displays. The RTX PRO 5000, the RTX PRO 4500 Workstation Edition and the RTX PRO 4000 have four DisplayPort 2.1b outputs, the RTX PRO 2000 four Mini DisplayPort 2.1b outputs. Dassault's SOLIDWORKS 2026 certification list, read 10 November 2026, has rows with desktop workstations for the RTX PRO 2000, 4000, 4500 and 5000 Blackwell, without separating editions or memory sizes; a row certifies one computer model with one card and one driver. A design office can run CAD and a local model on one such machine; our guide to GPU workstation builds from the RTX PRO 2000 to the 5000 goes through the builds.

What the workstation has to provide

Our estimate: board power from NVIDIA's datasheets plus one 350 W processor, plus 15 per cent for the rest of the system, to be confirmed by a measurement on the delivered system. The DGX Spark has a 240 W supply; NVIDIA's declaration gives 233.2 W maximum and 38.0 W idle at the wall.

Every build in the table stays far below the 2.94 kW we plan on a 230 V circuit with 16 A. The chassis has more to provide than the socket. The RTX PRO 5000 and the RTX PRO 4500 Workstation Edition take two slot widths and one 16-pin power connector each, with an adapter from two 8-pin PCIe leads in the box; four of the latter need eight free 8-pin leads or four native 16-pin cables. The RTX PRO 4000 takes one slot and an adapter from one 8-pin lead. The RTX PRO 2000 is half-height (2.7 in), the format usually called low profile, in two slot widths, with a maximum power consumption of 70 W and no power cable. A workstation with four RTX PRO 4500 Workstation Edition cards gives off about 1.3 kW of heat, roughly 4,500 BTU per hour.

For the platform, NVIDIA's configuration guide for NVIDIA-Certified Systems of 30 September 2026, written for servers, asks for at least six processor cores per GPU and system memory of at least twice the GPU memory; four cards at x16 take 64 PCIe 5.0 lanes. We build these workstations to order on Threadripper PRO or Xeon W platforms and supply the cards for towers that take them; send us the models, the number of people and the room the machine stands in, and we reply with a configuration. Branded towers and their option lists are the subject of RTX PRO cards in Dell, HP and Lenovo workstations.

The builds by task, with their buy pages

• Small models on a compact seat. One RTX PRO 2000 for files up to about 10 GiB: buy the RTX PRO 2000 Blackwell, RTX PRO 2000 Blackwell specs.
• One developer or a few users, models up to the 35B class in FP8. One RTX PRO 5000 48 GB, two for a second copy or a 32B model with more cache: buy the RTX PRO 5000 Blackwell 48 GB, RTX PRO 5000 with 48 GB or 72 GB, buy the workstation with 48 GB cards.
• Several people or models side by side, 14B class. Two to four RTX PRO 4500 Workstation Edition, one model per card: buy the RTX PRO 4500 Blackwell, RTX PRO 4500 Blackwell with 32 GB, buy the workstation with RTX PRO 4500 Workstation Edition cards.
• Four GPUs in four slots. Two to four RTX PRO 4000 for models up to about 14B in FP8 per card: buy the RTX PRO 4000 Blackwell, RTX PRO 4000 Blackwell specs, buy the RTX PRO 4000 workstation.
• One model above 40 GiB in one pool, or QLoRA on a 70B model. The DGX Spark, described on our DGX Spark page.

The whole workstation range is on our page to buy an AI workstation for LLMs and in our comparison of workstation GPUs for AI in 2026. The cards one class up are compared in RTX PRO 6000 alternatives with 48 GB and 32 GB, the deskside class above in DGX Station GB300 vs RTX PRO 6000 workstation, and three desktop systems in DGX Spark vs Mac Studio vs Ryzen AI Max. What a request should contain is listed in private AI sizing request: what we need to know.

A first consultation is free of charge. Tell us the task and the models you have in mind, or write to request@eurokommerz.at.

FAQ