GPU server for engineering firms: virtual CAD workstations, rendering, simulation and LLMs

GPU platform for an engineering firm: the short answer

An engineering or architecture firm can run virtual CAD workstations, overnight rendering, simulation and an internal LLM on one pool of GPU servers, if the cards handle both graphics and compute and the schedule keeps day and night loads apart. NVIDIA’s RTX vWS sizing guide, updated on 19 August 2026, presents the RTX PRO 6000 Blackwell Server Edition for “mixed-use environments that combine AI/ML compute with professional visualization or VDI workloads”. The same guide says the L40S covers “batch and real-time rendering, virtual workstations”. By day these cards serve virtual CAD workstations through NVIDIA vGPU, and at night the same cards render and solve.

For a mix of mostly light and medium users, one server with eight RTX PRO 6000 Server Edition cards carries about 40 concurrent CAD sessions at NVIDIA’s published densities, by our estimate; with heavy users only, it carries 16 to 32. Our example for 200 CAD users with 150 at peak needs five such servers, four in use and one as the failover reserve. Solvers that must run in double precision need an H200 NVL node beside the pool.

Estate sizing with office users is in our VDI GPU sizing guide, and the card’s profiles in our guide to RTX PRO 6000 Server Edition vGPU profiles.

Workloads and the GPUs that run them

An engineering firm’s GPU work falls into four kinds: interactive CAD and BIM sessions, final-frame rendering, simulation runs and language models. The first two need a graphics GPU, rendering also its RT Cores, and the last two need memory and compute. The RTX PRO 6000 Server Edition has 96 GB of GDDR7, 188 RT Cores and 120 TFLOPS of FP32, and the L40S 48 GB, 142 RT Cores and 91.6 TFLOPS, according to NVIDIA’s product pages read on 10 October 2026.

Users per GPU from NVIDIA’s RTX vWS sizing guide (example VDI configurations, 19 August 2026); cells per card from Ansys’s Fluent GPU buying guide (9 April 2026) for single precision with the segregated solver; LLM figure for gpt-oss-120b from our private ChatGPT server guide; FP64 from NVIDIA’s H200 page. SE stands for Server Edition.

The H200 NVL stays out of the desktop pool. NVIDIA’s H200 specifications list no RT Cores, and the GPUs its vWS sizing guide lists for virtual workstations are the RTX PRO 6000 and RTX PRO 4500 Server Editions, the L40S, the L4, the A40 and the A10.

Virtual CAD workstations on vGPU

NVIDIA’s RTX vWS guide sorts users into three types. Its typical deployment puts 6 to 8 light users on one GPU with an RTX PRO 4500 Blackwell DC-4Q profile, 3 to 6 medium users with DC-4Q or DC-8Q on the RTX PRO 4500 or RTX PRO 6000, and 2 to 4 heavy users with DC-12Q, DC-16Q or DC-24Q on the RTX PRO 6000. The VMs in that table have 4 vCPUs and 8 to 12 GB of RAM for light users, 8 vCPUs and 16 to 24 GB for medium users, and 12 or more vCPUs and 48 to 72 GB for heavy users. For the RTX PRO 6000 the guide lists profile sizes from 2 GB to 96 GB, and “1 GB profiles are not available on Blackwell GPUs”.

These are ranges because CAD users share a card’s compute, not only its memory. NVIDIA’s performance metrics page adds that many legacy CAD and CAE applications are “predominantly single-threaded” and that CPUs with “high clock speeds (generally above 3 GHz)” matter as much as the vCPU count, so the host processor belongs to the CAD sizing. During the pilot, a VM’s frame buffer “should not frequently exceed 90% or average over 70%”; above that, move the user to the next profile.

RTX vWS is counted per concurrent user: NVIDIA’s licensing guide states that “A CCU license is required for every user who is accessing or using the software at any given time”, and its feature table gives vWS CUDA and OpenCL support, which the night jobs below rely on.

We supply the RTX PRO 6000 Server Edition, the RTX PRO 4500 Server Edition and the L40S for virtual workstation hosts, with RTX vWS licences on the same invoice. Tell us how many CAD users you have of each type and which applications they run.

ISV certification for CAD on vGPU

NVIDIA lists “ISV certifications” as an RTX vWS feature that vPC does not include, and names Autodesk Revit, Dassault Systèmes CATIA and SOLIDWORKS among the applications vWS accelerates. Whether the CAD vendor supports its own software in a VM is a separate question.

Dassault Systèmes’ SOLIDWORKS system requirements, read on 10 October 2026, list VMware vSphere ESXi 8.0.3 U3g as a hypervisor for SOLIDWORKS 2026, and state that “SOLIDWORKS and eDrawings require a GPU for optimal performance in a virtualized environment” and that SOLIDWORKS Composer “is not supported in virtualized environments”. NVIDIA’s vGPU support matrix requires ESXi 8 Update 3g or later on 8.0 for the RTX PRO 6000 Server Edition, so the two lists meet at the same update. An undated Autodesk support article on virtual machines with roaming profiles states that “Virtualization setups are not something Autodesk supports” and leaves them to the customer and the virtualisation provider.

We found no current Siemens or PTC page we could read that lists vGPU support for NX or Creo. Ask the vendor or your reseller for the certified card, driver branch and hypervisor of the release you run, and test your own models in the pilot.

Rendering and simulation at night on the same cards

Without MIG, all time-sliced vGPUs on one card must have the same frame buffer size unless the card runs in mixed-size mode, as our VDI guide sets out from NVIDIA’s user guide. A card holds CAD desktops or one 96 GB VM, not both at once, and a render inside a desktop slows its neighbours. NVIDIA’s vWS guide sizes on the basis that “it is unlikely that all users will be rendering simultaneously”. Final frames and long solver runs move to the night.

1. Log off idle sessions and power off the desktop VMs after hours, at a time the firm sets, and keep a few cards for staff who work late.
2. Power on the render and solver VMs, each with one or more 96 GB Q profiles, on the emptied cards.
3. Let the render manager and the solver queue hand their jobs to these VMs.
4. Before office hours, drain the queues, power off the night VMs and start the desktop pool.

RTX vWS lists several vGPUs per VM among its features, so one render VM can take several whole cards and counts as one render machine. Each night VM takes one vWS licence while it runs, and the desktops have released theirs. Scene memory, scaling across cards and render node licences are in our GPU render server guide.

Simulation follows the solver’s precision. Ansys’s Fluent GPU buying guide calls double precision “The most important question to consider” and plans 1.2 GB per million cells for a typical single-precision case, about 80 million cells on one 96 GB card. The guide does not mention virtual machines, so ask Ansys whether the GPU solver is supported in a VM before you plan it on the shared cards. Double-precision cases go to the H200 NVL node, and our CFD GPU comparison of the H200 NVL and RTX PRO 6000 covers FP64 rates, memory per cell and licences.

Internal LLM and RAG for standards and project archives

The language model runs on its own cards day and night. Our private ChatGPT server sizing guide estimates about 19 conversations at 32K for gpt-oss-120b on one RTX PRO 6000. With that guide’s example usage values, a firm of 300 staff peaks at about 12 requests in flight, so one copy carries the load and a second copy on another server keeps it running when one fails.

The embedding model and the reranker for the standards and project archives fit MIG instances on a further card, since NVIDIA’s product page gives the Server Edition “up to four fully isolated instances”. The search has to respect who may see which project, as our article on RAG on company data explains. Whether the licence terms of purchased standards allow them to be indexed is a legal assessment for the company’s legal department.

NVIDIA’s AI Enterprise reference states that C-series compute vGPUs require an NVIDIA AI Enterprise licence. An LLM card on C-series profiles is therefore licensed per GPU, while the desktops stay on vWS.

Licence servers, data and management

vGPU licences come from the NVIDIA License System. NVIDIA’s documentation states that “A DLS instance is hosted on-premises at a location that is accessible from your private network”, as a virtual appliance or a container image, and describes registration on an air-gapped network, so the platform can serve its own vGPU licences without a cloud connection. The network licence servers of the CAD, rendering and simulation software run as VMs on CPU hosts beside it, with the hypervisor management and the queue managers.

Project data stays on the firm’s file or PDM storage on the data centre network. Each GPU server needs a BMC for out-of-band management and redundant power supplies.

Example platform for 200 CAD users

The example is our estimate from NVIDIA’s densities, not a reference design. The firm has 300 staff, of whom 200 work in CAD, with 150 sessions at peak: 90 light users at 8 per card, 50 medium users at 4 per card on DC-8Q and 10 heavy users at 3 per card on DC-16Q. That needs 12, 13 and 4 cards, 29 in all. We keep NVIDIA’s light-user figure for the smaller RTX PRO 4500 on the RTX PRO 6000, and a pilot may raise it.

Our estimate: users per card from NVIDIA’s RTX vWS sizing guide; eight cards per server as in NVIDIA’s RTX PRO Server specification for Omniverse; the H200 NVL joins 2 or 4 cards over an NVLink bridge at 900 GB/s per GPU, per NVIDIA’s H200 page.

The 32 cards in use by day fill four servers, and the fifth keeps them all running when one fails, since a vGPU desktop restarts only on a host with the same GPU type and room for its profile. Queue jobs on the reserve cards by day stop in that case. At NVIDIA’s upper VM sizes, the 150 desktops need about 3 TB of RAM, so each server needs about 1 TB to carry its share with one server down. The same table gives them 880 vCPUs, so the processors also need enough cores above 3 GHz.

We build GPU servers with up to eight RTX PRO 6000 Server Edition or H200 NVL cards to order and check the rack, power and airflow before we quote. Send us your user mix, renderer and solvers through the form below.

What we supply

We supply the cards in this guide, the RTX PRO 6000 Server Edition, RTX PRO 4500 Server Edition, L40S and H200 NVL, as professional NVIDIA GPUs, and build them into GPU servers to order for VDI, rendering and inference, assembled and burn-in tested, with manufacturer warranty on every component. RTX vWS and NVIDIA AI Enterprise licences come on the same invoice as the hardware, on one EU contract. Where the firm also wants the LLM and RAG platform built and supported, our Private AI/ML service covers it, with engineering by our partner Vixen.UNO.

FAQ