Omniverse system requirements: GPUs, servers and Isaac Sim specifications for digital twins

Omniverse and Isaac Sim system requirements at a glance

The system requirements of NVIDIA Omniverse and Isaac Sim start with an NVIDIA RTX GPU, a card with RT cores. Isaac Sim’s requirements page states that “GPUs without RT Cores (A100, H100) are not supported”, and the Omniverse developer guide says that running its SDKs on non-RTX GPUs “is provided without any support guarantees”. NVIDIA’s recommended workstation for Omniverse Kit has one RTX PRO 6000 Blackwell, a processor with 16 or more cores, 128 to 256 GB of RAM and 1 TB of NVMe storage. Its recommended RTX PRO Server has eight RTX PRO 6000 Blackwell Server Edition cards with 768 GB of GPU memory in total.

The figures below come from the technical requirements in the Omniverse developer guide, updated on 8 October 2026, and from the Isaac Sim requirements page, updated on 18 September 2026. Kit is the SDK on which Omniverse applications are built.

Omniverse developer guide, technical requirements (8 October 2026); Isaac Sim requirements (18 September 2026); x86_64 figures. Card memory from NVIDIA’s RTX PRO 6000 product page.

NVIDIA adds to the Kit table that the RAM “depends on the workload size and Operating System requirements”. On Linux, both products support Ubuntu 22.04 and 24.04. Kit also has an arm64 specification with the same RTX PRO 6000, RAM and storage as the workstation, on Ubuntu only.

Why the H100, H200 and A100 are not Omniverse GPUs

The Omniverse RTX Renderer uses hardware features that differ by GPU generation, and the developer guide lists them per architecture. For the Hopper H100, H200 and H800 (compute capability 9.0) and the Ampere A100, A30 and A800 (8.0), the table marks the OptiX denoiser as supported. It marks DLSS Ray Reconstruction, DLSS Frame Generation, Shader Execution Reordering, Opacity Micro-Map and Motion BVH as not supported on these GPUs. NVIDIA explains the visible effect: “Without DLSS Ray Reconstruction support, images appear significantly noisy when using the RTX Real-Time render mode.” Blackwell RTX PRO cards (12.0) and Ada Lovelace cards (8.9) have every feature in the table, with Shader Execution Reordering listed for DX12 and Motion BVH for Vulkan.

Feature support from the GPU feature table of the Omniverse developer guide (8 October 2026), mapped by compute capability as NVIDIA lists it for each card; memory from NVIDIA’s product pages and our GPU page.

The H200 NVL therefore belongs in a node that trains or fine-tunes AI models, while rendering and simulation run on RTX cards. NVIDIA’s L40S page lists 142 third-generation RT Cores and vGPU support, and names “OpenUSD-based 3D and simulation workflows built on Omniverse” among its uses.

Isaac Sim requirements and GPU memory

Isaac Sim is NVIDIA’s robotics simulator built on Omniverse, and its requirements are stricter than the Kit minimum: 32 GB of RAM and a 16 GB GPU at the low end. The ideal tier asks for 48 GB of VRAM, which the RTX PRO 5000, the L40S and the RTX PRO 6000 meet. Its footnotes say that “More RAM and VRAM is recommended for advanced usage of Isaac Sim”, and that “Isaac Lab usage will require additional RAM and VRAM for training.” A GPU with less than 16 GB “may be insufficient to run a complex scene rendering more than 16MP per frame”, and NVIDIA notes that workflows with many sensors are affected most by VRAM limits.

By our arithmetic, eight simulated 1080p cameras at about 2.07 megapixels each come to about 16.6 MP per frame, above the 16 MP in NVIDIA’s footnote. The sensor set-up therefore belongs in the GPU sizing alongside the scene, and the minimum tier is a floor rather than a target.

The Isaac Sim container is supported on Linux only, and an internet connection is needed to reach the online assets and to run some extensions, which matters for an isolated network; our guide to an air-gapped server with offline installation shows what has to be mirrored there. Isaac Sim’s arm64 builds are supported only on the DGX Spark, tested with DGX OS 7 and driver 580.159.03, and NVIDIA lists one limitation for Isaac Sim 6.0 there: cuRobo and cuMotion are not supported. The Omniverse developer guide also lists a known issue on Linux arm64, crashes in OpenGL interop and the Hydra engine, fixed in driver 580.173.02, 595.84 or newer, so check which driver your DGX OS release brings.

Drivers and operating systems for Omniverse

The developer guide lists the drivers “validated with the latest Omniverse releases”. On Linux, for Blackwell data-centre cards, these are 580.95.05 on the R580 branch and 595.58.03 on R595; for workstation cards the table also lists 570.169 on R570. Isaac Sim was tested on 595.58.03, and its page recommends the latest production branch driver for a new GPU or driver problems on Linux. NVIDIA adds that “Newer drivers may function properly, but have not been fully tested for stability and compatibility.” Both validated versions are older than 580.178.04 and 595.91.07, which NVIDIA’s security bulletin of 30 September 2026 names as the fixed releases, so test a patched driver before rollout.

On Linux, DLSS Ray Reconstruction on Blackwell cards needs Kit 106.5.3 or later, according to the same guide. How the R580 and R595 branches differ, and how to pin a server to one, is covered in our guide to NVIDIA driver branches and CUDA versions.

Workstation or RTX PRO Server for a digital twin

A designer who builds and reviews OpenUSD scenes works on a workstation. NVIDIA’s Kit specification names the RTX PRO 6000 Blackwell for it, and the three RTX PRO 6000 editions compared shows which edition suits a tower and which a rack. Where one machine has to hold several GPUs for simulation and rendering at once, the 300 W Max-Q edition allows up to four cards in one tower, as our article on four RTX PRO 6000 Max-Q cards in one workstation explains.

A shared twin that runs continuously, generates synthetic data or serves several teams belongs on a server. NVIDIA positions its RTX PRO Server for industrial digitalisation, robot simulation and synthetic data generation, and its Kit server specification is that design: eight RTX PRO 6000 Server Edition cards, 64 or more CPU cores at a base clock of 3 GHz or more and 1 to 2 TB of RAM. NVIDIA’s reference architecture for industrial facility digital twins, updated on 8 October 2026, uses RTX PRO 6000 Blackwell GPUs through “scalable data center deployments” or NVIDIA-Certified workstations, with Isaac Sim for robot simulation and sensor data streamed into the twin. Which manufacturers list the Server Edition, and for how many cards, is set out in RTX PRO 6000 Server Edition compatible servers.

We build GPU nodes for virtual workstations and rendering with RTX PRO Blackwell, L40S or L4 cards, and check the rack, power and airflow before we quote. Describe your scenes, simulations and number of users in the form below.

Streaming Omniverse apps with Kit App Streaming

Viewers who only look at a twin do not need a discrete GPU. NVIDIA’s minimum for a streaming client is an integrated GPU, 2 GB of RAM and 1 GB of storage. NVIDIA describes Kit App Streaming as cloud-based streaming of a Kit app from a Kubernetes-driven platform, with the rendering on the server.

Its requirements page, updated on 8 October 2026, asks for GPU worker nodes with “RTX based GPUs”, 8 to 16 or more vCPUs and 32 to 64 GiB or more of memory each, on Kubernetes 1.29. The number of GPU worker nodes is given as “1 per stream”, and the footnote reads “Your cluster will be limited to only as many streams as you have available GPU nodes.” Size the cluster from peak concurrent viewers, not from named users. With the chart defaults, the same page says, the security groups must allow TCP 41001 to 41025 and UDP 41026 to 41050. The documentation index says that Kit App Streaming “can be deployed across any cloud provider or on-premises infrastructure”. It also states that “Omniverse Streaming on NVCF replaces Omniverse Kit App Streaming (OKAS) as the Reference Architecture”, so confirm with NVIDIA which streaming stack applies to an on-premise cluster.

Omniverse Enterprise licensing and support

NVIDIA’s licensing guide for NVIDIA AI Enterprise, updated on 2 September 2026, states that “NVIDIA Omniverse, included with NVIDIA AI Enterprise, is a collection of libraries and microservices” for industrial digitalisation. The guide does not use the name Omniverse Enterprise and does not list which Omniverse applications or services the inclusion covers, so have the quote state it for Isaac Sim, Kit apps and streaming. The same guide says of NVIDIA AI Enterprise that “Customers are required to deploy it on NVIDIA-Certified Systems in order to be supported.” Its licensing page states that the software “is licensed on a per-GPU basis”, so a server with eight cards that runs it counts eight GPUs, as our AI Enterprise licensing guide explains.

Designers on virtual workstations with graphics vGPU profiles need NVIDIA RTX vWS, counted per concurrently active user rather than per GPU, and of the RTX PRO 6000 editions only the Server Edition supports vGPU.

NVIDIA’s Omniverse page lets developers use the documentation, GitHub repositories and sample content for development, research, testing and evaluation, and says that for production or enterprise use “support is available through NVIDIA Enterprise Support” for customers with a support entitlement. Building Isaac Sim from its GitHub repository prompts the user to accept the Omniverse Licensing Terms. Which terms cover a given production use is set by NVIDIA’s licence agreements, and that reading is a matter for the company’s legal department.

GPU configurations by digital twin use

The configurations below use hardware we supply and are our reading of NVIDIA’s published specifications, not NVIDIA recommendations, and the scene, sensor count and number of viewers decide the final build.

Our reading of the Omniverse developer guide (8 October 2026), the Isaac Sim requirements (18 September 2026) and the Kit App Streaming requirements (8 October 2026); not an NVIDIA sizing.

We supply the hardware in this table, as cards, in workstations or in servers built to order. Tell us which use comes first and how many people will work in the twin, and we reply with a configuration and quote within one business day.

What we supply

We supply the RTX PRO 6000 in its Workstation, Max-Q and Server editions, the RTX PRO 5000, 4500 and 4000, the L40S, the L4, the H200 NVL and the DGX Spark, listed on our professional NVIDIA GPUs page, on one EU contract and invoice with manufacturer warranty. We build AI servers to order, including GPU nodes for virtual workstations and rendering with RTX PRO Blackwell, L40S or L4 cards, assembled and burn-in tested, and we check the rack, power and airflow before we quote. Operating system, drivers, CUDA and a container runtime are installed on request. NVIDIA AI Enterprise and vGPU licences come on the same invoice as the hardware.

FAQ