GPU workstation builds by task: eight configurations from the RTX A400 to the RTX PRO 5000

Eight GPU workstation builds, from the CAD seat to local AI

A GPU workstation build starts from the task, and for most tasks one card settles it. An office CAD seat takes an NVIDIA RTX A1000 with 8 GB or an NVIDIA RTX PRO 2000 Blackwell with 16 GB, large assemblies and BIM the 24 GB NVIDIA RTX PRO 4000 Blackwell, rendering a 32 GB card, and local language models the 48 GB NVIDIA RTX PRO 5000 Blackwell or two to four RTX PRO 4000. The eight builds below name the card, the count and what the chassis has to provide.

NVIDIA datasheets 3279761 (May 2024), RTX A1000 (April 2024), 5322351 (June 2026), 5323450 (June 2026), 5108623 (April 2026), 2898211 (August 2023) and 5349550 (June 2026). The assignment of cards to tasks is ours; the board power of several cards is our sum.

Workstation configuration for an office CAD seat

Autodesk's system requirements for AutoCAD 2027, read 10 November 2026, name a 2 GB GPU in the basic line and an 8 GB GPU in the recommended line; for AutoCAD LT 2027 the lines are 1 GB and 4 GB. Dassault publishes no memory figure for SOLIDWORKS 2026. Its certification list, read the same day, has desktop rows for the RTX A1000 and the RTX PRO 2000 Blackwell, and each row certifies one computer model with one card and one driver.

Both boards are 2.7 in high, the half-height format usually called low profile. The RTX A1000 meets the 8 GB figure in one slot: NVIDIA's datasheet gives a maximum power consumption of 50 W and names no auxiliary power connector. The RTX PRO 2000 is two slots wide, draws at most 70 W and, by NVIDIA's quick start guide, uses no power cable. Its 16 GB also lie above the 12 GB of AutoCAD's separate table for large datasets and point clouds. For 2D drafting and AutoCAD LT the 4 GB NVIDIA RTX A400 is the smaller single-slot option.

To order, see buy the RTX A1000 and buy the RTX PRO 2000 Blackwell. The applications are in GPU for SOLIDWORKS 2026 and GPU for AutoCAD, Revit and Inventor, a fleet of office PCs in the RTX PRO 2000 Blackwell fleet rollout.

Workstation GPU for Revit, Inventor and large SOLIDWORKS assemblies

Revit 2027 names 4 GB for its entry-level configuration, 6 GB for Value and 8 GB for Performance (page of 7 April 2026). Inventor 2027 recommends an "8 GB/16 GB GPU with 240 GB/s Bandwidth", while its table for large assemblies says "8 GB GPU with 106 GB/S" (page of 11 May 2026). Autodesk does not define that bandwidth, so we compare memory only.

By memory a 16 GB card meets every one of these figures. We build this seat around the RTX PRO 4000 Blackwell all the same, for reasons of our own: its 24 GB leave room when a renderer runs beside the model, and it needs one slot. The board is 4.4 in high and 9.5 in long, has a total board power of 145 W and one PCIe CEM5 16-pin connector; NVIDIA's quick start guide lists an adapter from one 8-pin PCIe cable in the box. See GPU for CAD workstations and buy the RTX PRO 4000 Blackwell.

We supply every card on this page across the EU and build workstations to order; send us the CAD application, its version and the tower you plan to use, and we answer with a written quote within one business day.

Workstation for rendering and visualisation with 32 GB on one card

Real-time visualisation tools state modest figures: Enscape 4 GB as the minimum and 12 GB for VR (page of 19 March 2026), Twinmotion 2026.1 at least 6 GB and 12 GB in its high-end tier, Lumion 2026 6, 10 and 16 GB. A 16 GB card meets these memory figures; Lumion and Twinmotion also set benchmark scores, which we did not check. Two statements push a rendering seat towards one larger card. Chaos writes for V-Ray GPU that the scene must fit fully into GPU memory, and Enscape uses one GPU even when several are installed.

The build is one 32 GB card: the NVIDIA RTX PRO 4500 Blackwell Workstation Edition has 32 GB of GDDR7 at 896 GB/s and a total board power of 200 W; the NVIDIA RTX 5000 Ada Generation has 32 GB of GDDR6 at 576 GB/s and 250 W. Both are 4.4 in high, 10.5 in long and two slots wide, with a 16-pin power connector.

In Blender Open Data (crowd-sourced, read 10 November 2026, Blender 4.5.0) the median of 17 submissions for the RTX PRO 4500 Workstation Edition is 8,486.57 samples per minute; the median of 9 submissions for the RTX 5000 Ada is 5,677.13. The RTX 5000 Ada stays the right order where a fleet is validated on Ada or where the card must be on NVIDIA's vGPU list, which names it from release 16.1. Scenes beyond 32 GB take the RTX PRO 5000 Blackwell with 48 GB.

More in GPU for Enscape, Twinmotion, D5 Render and Lumion; to order, buy the RTX PRO 4500 Blackwell or buy the RTX 5000 Ada. Render nodes for LED walls are a separate build: RTX PRO 4500 Blackwell for Unreal Engine and virtual production.

GPU for a video editing workstation

The readme of DaVinci Resolve Studio 21.2 of 4 November 2026 asks for at least 4 GB of VRAM, 12 GB for background rendering and 16 GB for its advanced AI tools. Adobe's page for Premiere, as displayed on 10 November 2026, names 4 GB of GPU memory as the minimum and 8 GB as recommended. 16 GB is the smallest card size that reaches every one of these lines.

The cards differ in their video engines: one encoder and one decoder on the RTX PRO 2000, two of each on the RTX PRO 4000 and the RTX PRO 4500 Workstation Edition, three on the RTX PRO 5000. NVIDIA's Video Encode and Decode GPU Support Matrix of 16 October 2026 lists AV1 encoding and an unrestricted number of NVENC sessions for the Blackwell workstation cards and the RTX Ada cards it contains, and 4:2:2 encoding for Blackwell only. The RTX A1000 and the RTX A400 are limited to 8 sessions and have no AV1 encoding. Whether an editing program uses these functions is for its maker to state.

The compact build is an RTX PRO 2000; the build with two encoders and 24 GB is the single-slot RTX PRO 4000. See GPU for Premiere Pro and DaVinci Resolve and NVENC session limits.

AI workstation build for local LLMs with one 48 GB card or several 24 GB cards

By our sizing rule a card has a budget for weights plus KV cache of 0.9 times the memory the driver reports, less 3 GiB. We scale the reported memory from the 95.6 GiB the driver shows on the 96 GB card; that scaling is an assumption, and the driver on the delivered card decides.

NVIDIA datasheets 5349550 and 5323450 (June 2026); budgets by our sizing rule, 40.0 GiB per 48 GB card and 18.5 GiB per 24 GB card.

The difference lies in the pools. One RTX PRO 5000 holds a model and its context in one memory space: Qwen3.6-35B-A3B in FP8, a file of 34.88 GiB on Hugging Face (read 10 November 2026), fits and leaves about 5.1 GiB for context. On the RTX PRO 4000 each card addresses only its own 24 GB. gpt-oss-20b in MXFP4 (12.82 GiB) fits one card and leaves 5.7 GiB. The 34.88 GiB file does not fit one card; two cards sum to 37.0 GiB, but the serving software would have to split the model over PCIe, and every split costs traffic between the cards. NVIDIA's datasheets list no NVLink for either card.

Our published sizing for Qwen3-14B in FP8 (15.2 GiB of weights, 8,192-token conversations with an FP8 KV cache) gives 5 concurrent conversations on one RTX PRO 4000 and 20 on four cards as four copies, against 39 on one RTX PRO 5000 with 48 GB. These counts say what fits in memory; no published measurement of token rates exists for either card, so the figures to compare are the memory bandwidths, 1,344 and 672 GB/s per card.

Four RTX PRO 4000 take four single slots and four 16-pin cables, or four 8-pin PCIe leads for the adapters; at x16 they occupy 64 PCIe 5.0 lanes, which is why we build four-card workstations on Threadripper PRO or Xeon W platforms. By our estimate two cards draw about 0.75 kW at the wall and four about 1.1 kW (board power plus one 350 W processor, plus 15 per cent; a measurement on the delivered system confirms it). NVIDIA's quick start guide for the Blackwell cards asks for at least as much system memory as GPU memory and recommends twice as much.

Sizing is free of charge: tell us the model, its precision and the number of people who will use it, and we propose the build. Further reading: buy the RTX PRO 5000 Blackwell with 48 GB, four RTX PRO 4000 in one workstation, buy an RTX PRO 4000 workstation, RTX PRO 4000 Blackwell for local LLMs, buying an AI workstation for LLMs and DGX Spark alternatives.

Image generation and the four-display desk

For image models we compare the file sizes with 0.9 times the card memory; no speeds were measured. On the 16 GB RTX PRO 2000, SDXL base 1.0 (6.46 GiB as a pipeline) and Stable Diffusion 3.5 Medium with its T5 encoder in FP8 (10.84 GiB) fit completely. FLUX.2 klein 4B needs 14.87 GiB with text encoders and VAE and fits the 24 GB RTX PRO 4000; Stable Diffusion 3.5 Large with the 16-bit T5 encoder comes to 25.97 GiB and fits the 32 GB RTX PRO 4500 Workstation Edition. The file tables are in ComfyUI on RTX PRO 2000, 4000 and 4500.

For a desk with four displays, the RTX A400 and the RTX A1000 each have four Mini DisplayPort 1.4a outputs and drive four displays at 4096 × 2160 and 120 Hz, in one slot at 50 W; 8K is listed only for the RTX A1000, for two displays at 30 Hz. Both datasheets list Mosaic, and neither card is compatible with NVIDIA's sync board. See four monitors with the RTX A400 and RTX A1000 and buy the RTX A400.

What the chassis must provide: slots, power connector, airflow

NVIDIA datasheets as under the first table; connectors and adapters from NVIDIA's quick start guides for the RTX PRO Blackwell cards and the RTX Ada cards. "None named" means that the datasheet names no auxiliary power connector.

Slots. The 2.7 in boards are half-height; the 4.4 in boards need a full-height slot, the dual-slot cards also 10.5 in of free length. See low-profile and single-slot NVIDIA RTX cards.

Power connector. NVIDIA calls the connector on the larger cards PCIe CEM5 16-pin; the power supply needs such a cable or the free 8-pin PCIe leads for the adapter in the box. NVIDIA publishes no power-supply wattage for these cards, so the board power is added to what the system already draws; see RTX power connectors and power supplies.

Airflow. NVIDIA's datasheets describe the cooling of all seven cards as active and give neither a fan type nor an airflow figure; what can be planned is the heat, which equals the power drawn at the wall. The NVIDIA RTX PRO 4500 Blackwell Server Edition, which we also supply, is passive, takes its air from a server chassis and belongs in none of these towers; see RTX PRO 4500 or RTX PRO 4500 Server Edition.

In a branded tower the maker's own option list decides which card fits and how many: RTX PRO cards in Dell, HP and Lenovo workstations.

Three more Ada cards for the same builds

Three more cards we supply fit these builds. The NVIDIA RTX 2000 Ada Generation has the format of the RTX PRO 2000, two slots and 70 W without a power cable, with 16 GB of GDDR6 at 224 GB/s (specs, order page). The NVIDIA RTX 4500 Ada Generation carries 24 GB at 432 GB/s in a dual-slot card of 210 W (specs, order page). The NVIDIA RTX 2000E Ada Generation puts 16 GB into a single slot at 50 W (specs, order page).

Eleven professional cards from 4 to 48 GB stand side by side in our NVIDIA RTX workstation GPU guide. Where the card in the tower today is an RTX A2000, A4000 or A5000, start from the upgrade map for RTX A-series cards.

From the build to a written quote

We supply the cards alone for towers you already have, and we build workstations to order, assembled and tested before delivery. The quote comes in writing within one business day, on one European contract with an invoice from Austria; manufacturer warranty is handled through us. Write to us with the task, the software versions and the number of seats, or send the same by e-mail to request@eurokommerz.at.

FAQ