For AI work, partly. For CAD and most simulation, no. A DGX Station GB300 can replace the AI side of a room of engineering workstations: training and running models, GPU data science, surrogate models and large-batch inference, shared by the whole team. It cannot replace the interactive engineering desktop. Its GB300 superchip has no display output, its Arm CPU cannot run many x86-only engineering applications natively, and its GPU slices do not support graphics. The realistic plan is hybrid: keep conventional workstations for CAD, visualisation and x86 tools, and centralise AI compute on the Station.
Where does each machine fit?
| Task | DGX Station GB300 | Conventional RTX PRO workstation |
|---|---|---|
| Training and fine-tuning AI models | Strong: 252 GB HBM3e GPU memory, 748 GB coherent in total | Limited by the card's memory |
| Serving AI models to many users | Strong: designed as a "Deskside Server for AI Teams"; MIG up to 7 instances | One user or a small team |
| Interactive CAD and 3D modelling | Needs an optional RTX PRO card; GB300 has no display output | Designed for it |
| MATLAB on Linux | Not natively: MATLAB R2026b needs an x86-64 CPU | Runs natively on x86 |
| Ansys Fluent and LS-DYNA | Arm builds listed for RHEL and Rocky Linux, solver-focused; Fluent GPU solver on Arm is beta; Ubuntu not listed | Full x86 support |
| Siemens STAR-CCM+ | Arm64 batch mode only, no official GUI or CAD import (secondary source; confirm with Siemens) | Full x86 support |
| NVIDIA Isaac Sim | aarch64 builds "currently only supported on DGX Spark" | Supported on x86 RTX systems |
| Graphics inside shared GPU slices | "No graphics APIs are supported" in MIG instances | Not applicable |
From NVIDIA's DGX Station datasheet and Development Guide (2026), NVIDIA's MIG User Guide (13 Oct 2026), MathWorks R2026b system requirements, Ansys' January 2025 Arm support announcement and Fluent 2025 R2 beta documentation, Isaac Sim requirements (14 Oct 2026) and a secondary mirror of Siemens' STAR-CCM+ 2406 documentation, all accessed 4 Nov 2026. Software support changes often, so check each vendor's current platform list. SourceIT has not tested these applications.
Why can't engineers just work on the Station directly?
Three reasons, all in the hardware design. First, the GB300 has no display output. The only video port on the back is a mini-DisplayPort for the management controller, which NVIDIA says "is intended only for system management". Monitors need an optional RTX PRO card in one of the PCIe slots.
Second, that card costs the AI side performance. The Station's 1,600 W budget is shared, and NVIDIA's power-sloshing documentation says the GB300's cap drops to 1,250 W with a 76–300 W RTX card and to 950 W with a card above 300 W, such as an RTX PRO 6000. NVIDIA promotes pairing the Station with an RTX PRO 6000 for "ray-traced visualization and simulation", but on a single box you trade AI headroom for graphics. More on this in Why Does a S$150,000 DGX Station Need Another Graphics Card?
Third, the Grace CPU is Arm. Engineering software has been slow to follow, as the table shows. For the full compatibility picture, see Your DGX Station Has 72 CPU Cores. Why Might Your Software Still Struggle?
A DGX Station centralises the AI an engineering team runs. It does not centralise the screen, mouse and x86 software an engineer works in all day.
What does the money buy either way?
Public pricing puts a DGX Station at around S$150,000 depending on OEM and configuration (US pricing for most OEMs of roughly US$93,000–125,000 is about S$119,000–160,000 before GST at USD/SGD 1.2802). For scale only, the same S$150,000 would buy about 25 HP Z2 Tower G1i workstations with an RTX PRO 2000 at S$5,788.00, about 11 HP Z4 G5 workstations with Xeon W3-2423, 64GB and RTX PRO 2000 at S$12,683.00, or about 29 Lenovo ThinkStation P2 Tower Gen 2 units with RTX PRO 2000 at S$5,018.54 (all listed 4 Nov 2026). A single NVIDIA RTX PRO 6000 Blackwell Workstation Edition with 96 GB is S$29,100.00 (listed 4 Nov 2026).
These are not like-for-like. None of those workstations comes close to the Station's 252 GB of GPU memory for large models, and the Station cannot give 25 engineers a responsive desktop. The comparison shows why most firms should not frame this as either-or: the workstations serve the people, and the Station serves the models.
How does a hybrid setup work in practice?
- Keep a workstation per engineer for CAD, meshing, post-processing and x86 tools like MATLAB. An RTX PRO 2000 or 4000-class card covers most CAD.
- Move AI jobs to the Station. Engineers submit training runs, surrogate-model fitting and inference from their desks over SSH, VS Code or a web notebook. HP's Z Toolkit, for example, handles discovery, pairing and remote access from a Windows 11 or Ubuntu client.
- Split the GPU sensibly. MIG gives up to seven isolated GPU instances for development and smaller models; see Can Seven AI Teams Share One DGX Station?
- Connect them properly. Engineering datasets are large. The Station's 10GbE port suits office networks; its two 400G ports need a matching switch and cabling.
- Keep heavy GPU visualisation on a workstation with a high-end card, such as an RTX PRO 6000, rather than adding that card to the Station and cutting its AI power budget.
When could a DGX Station replace some workstations?
When the "workstations" are really AI boxes. If engineers bought high-end GPU towers mainly to train or run models, and do their CAD on separate laptops or desktops anyway, consolidating that AI work onto one shared Station can reduce duplicated GPUs and idle hardware. The same applies to data-science teams running PyTorch, JAX or RAPIDS, all of which have Arm builds. If you want to compare GPU options first, read NVIDIA GB10 vs RTX PRO 6000 Blackwell and Best AI Workstation for Research Labs and Universities.
Buying in Singapore
SourceIT supplies the HP ZGX Fury G1n (factory-fitted with an RTX PRO 2000 or RTX PRO 4000 SFF display card), MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC, all price on request. For the engineers' desks, the HP Z2 Tower G1i U7-265 32GB 1TB RTX PRO 2000 is S$5,788.00, the HP Z4 G5 Xeon W3-2423 64GB 1TB RTX PRO 2000 is S$12,683.00, the Lenovo ThinkStation P2 Tower Gen 2 with RTX PRO 2000 is S$5,018.54 and the NVIDIA RTX PRO 6000 Blackwell Workstation Edition is S$29,100.00 (all listed 4 Nov 2026). Browse the AI Workstations and AI Servers collection. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Can I plug a monitor into a DGX Station? Only through an optional RTX PRO graphics card. The GB300 superchip has no display output, and the rear mini-DisplayPort is for the management controller only.
Can a DGX Station run CAD software? Only if the CAD application has an Arm Linux version and you add an RTX PRO display card. Most CAD and many engineering tools are x86 and Windows based, so they stay on conventional workstations.
Does adding a graphics card slow the AI side? It can. NVIDIA says the GB300's power cap drops from 1,600 W to 1,250 W with a 76–300 W RTX card and to 950 W with a card over 300 W, such as the RTX PRO 6000.
Can engineers share the DGX Station GPU? Yes. Multi-Instance GPU splits it into up to seven isolated instances, though graphics APIs are not supported inside MIG instances, so it is for compute, not remote desktops.
Should we buy a DGX Station or more workstations? Usually both, for different jobs. Workstations give each engineer a responsive desktop for CAD and x86 tools; a DGX Station gives the team shared capacity for large AI models.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
