GPU server for Isaac Sim and Isaac Lab: robot learning on RTX PRO 6000 Server Edition

What a robotics team needs from a shared GPU server

A GPU server for robotics runs NVIDIA Isaac Sim and Isaac Lab for a whole team on NVIDIA RTX PRO 6000 Blackwell Server Edition cards: robot simulation, reinforcement learning of control policies and synthetic sensor data, on hardware that stays in a server room instead of under desks. The card has to have RT Cores, because Isaac Sim does not support GPUs without them, and the Server Edition is the passive rack version of the card NVIDIA names as Isaac Sim's ideal GPU. A server with four of these cards runs four training jobs side by side or one job across all four; two cards suit a small team, and eight match NVIDIA's own Omniverse server specification.

This guide is for the head of a robotics or automation R&D team and for the IT staff who will run the server. Our guide to Omniverse and Isaac Sim system requirements covers the requirement tiers, drivers, streaming and licensing for digital twins in detail. Here the subject is training: how Isaac Lab spreads work over several GPUs, how much memory a task takes, and how many cards a team needs. Visual inspection and factory digital twins are covered in our guide to AI servers for manufacturing, and driving simulation with CARLA in our guide to AI servers for automotive R&D.

Why the card must have RT Cores

Isaac Sim renders its sensors with the Omniverse RTX renderer, and its requirements page excludes cards without RT Cores by name: the A100 and the H100. The README of the Isaac Sim repository on GitHub gives the A40 as the minimum data-centre GPU and an RTX 4080 as the workstation minimum. For a robotics server this decides the card type before any question of size.

Isaac Sim requirements page (last updated 14 October 2026) and Isaac Sim README on GitHub, read 24 October 2026; RT Core counts from NVIDIA's product pages as quoted in our engineering-firms guide; memory from NVIDIA's datasheets.

Isaac Lab 3.0 adds a second path. Its installation guide states that "Headless Newton simulation does not require RTX hardware", and its kit-less container image contains neither Isaac Sim nor Kit. The same image still includes the OVRTX renderer, which needs NVIDIA graphics driver capabilities, so camera-based tasks keep their need for an RTX card. NVIDIA publishes no Isaac Lab figures for the H200 NVL, and we would not plan simulation on it. Where a team also trains large networks on recorded or simulated data, the H200 NVL fits as a separate training node, as our comparison of the H200 NVL and RTX PRO 6000 for fine-tuning explains.

How Isaac Lab trains on several GPUs

Isaac Lab's multi-GPU guide, last updated on 24 October 2026, describes a simple model. Isaac Lab "starts one training process per GPU, gives each process its own simulation environments", and the processes synchronise their policy updates. Each process owns one Isaac Lab application with its vectorised environments, one copy of the policy and one rollout buffer. With PyTorch, train_multigpu wraps torchrun and synchronises gradients through DistributedDataParallel.

Three consequences follow for the hardware:

1. Cards are the unit of work. A four-card server runs one experiment on four cards with --num_gpus 4, or four separate experiments of one card each. A team that compares several reward or randomisation variants at once uses the second pattern.
2. The host must keep up. NVIDIA writes that simulation "does not move between GPUs", so host CPU, RAM and I/O can become limits as the GPU count grows. Each process steps its own physics, which is why we size processor cores per card.
3. Libraries and platform are fixed. Multi-GPU training supports RSL-RL, RL-Games and skrl, and multi-GPU and multi-node training need Linux and NVIDIA NCCL; Windows is not supported.

Across several servers, every node needs the same Isaac Lab checkout, dependencies, task configuration and training assets, and the nodes must reach each other on the rendezvous port. NVIDIA adds that "multi-node scaling depends heavily on the network between nodes". Two practical details from the same guide save time: rank 0 writes the videos to the usual place and every other rank writes to its own folder, and if a rendered job fails because CUDA and the renderer count the GPUs in different orders, CUDA_DEVICE_ORDER=PCI_BUS_ID is the documented fix. On a multi-card server we would set it in the job environment from the start.

GPU memory and throughput in NVIDIA's benchmarks

NVIDIA's performance benchmark page for Isaac Lab (documentation version 3.0.0-beta2) lists memory use per task and frames per second on single and multiple GPUs. All runs used RL-Games on Ubuntu 22.04 with the PhysX backend, except the G1 task, which used RSL-RL.

Isaac Lab performance benchmarks, memory consumption table, documentation v3.0.0-beta2, read 24 October 2026; NVIDIA does not name the GPU for this table.

The table separates two kinds of task. State-based policies, which observe values such as joint positions, use only a few GB of a card even at thousands of environments, so their speed depends on GPU compute and on the host. Camera-based policies are different: 1,024 environments with one RGB camera need 16.7 GB, more than the 16 GB of the Isaac Sim minimum tier, and memory grows with the number of cameras and their resolution. A 96 GB card leaves room for larger camera tasks or for several sensors per robot; NVIDIA lists no larger camera run, so measure your own task before you fix the environment count.

Isaac Lab performance benchmarks, "Step, Inference, and Train" frames per second, L40 single GPU and one node with 4 × L40 (Intel Xeon Platinum 8362), read 24 October 2026; the factor is our arithmetic.

On these measurements, four cards in one node trained between 3.3 and 4.3 times as fast as one card. The figures were measured on L40 cards, not on the RTX PRO 6000 Server Edition, and the page gives no Blackwell figures. Take them as the scaling pattern, not as the speed of our configurations.

We build these servers with two, four or eight RTX PRO 6000 Server Edition cards; send us your tasks, sensors and the number of experiments you run in parallel, and our team replies with a configuration.

Server options by card count

The options follow the reference configurations on our AI servers page and NVIDIA's guide for certified systems: two processors, at least six cores per GPU and system memory of at least twice the GPU memory. NVIDIA's Omniverse Kit server specification, quoted in our Omniverse guide, asks for eight RTX PRO 6000 Server Edition cards with 64 or more CPU cores at 3 GHz or more and 1 to 2 TB of RAM, so for simulation we choose processors with high clock speeds as well as enough cores.

Reference configurations on our AI servers page; NVIDIA-Certified Systems configuration guide; power by our estimate (cards at 600 W, two processors, plus 15 per cent for the rest of the server), as in our sibling server guides.

Two cards suit a team of a few researchers who run one or two experiments at a time, or a company that starts with synthetic data for a perception model. Four cards are the size we would propose first for an R&D team: four experiments run in parallel by day, and one run across all four cards by night. Eight cards match NVIDIA's Kit server specification and carry several teams, or simulation and synthetic data generation for a digital twin next to training; our guide to GPU allocation in an eight-card server shows how to divide the cards between users. A team that trains large networks on recorded data adds the H200 NVL node, which has no RT Cores and therefore stays out of simulation.

The system memory in our reference configurations is far above the 3.7 to 7.5 GB of RAM per process in NVIDIA's table. It is there for the NVIDIA-Certified rule, for datasets and recorded trajectories, and for the Isaac Sim tiers, which ask for 64 GB of RAM per machine in the good and ideal tiers and more for advanced use. Each card needs its own processor cores to step physics, so we do not save on cores to fit more cards.

Workstations for scene building and debugging

Scene building, robot description files and debugging happen at a desk with a display. Isaac Sim's ideal tier is a workstation with an RTX PRO 6000 Blackwell, 16 cores, 64 GB of RAM and a 1 TB NVMe SSD. The RTX PRO 6000 Max-Q is the 300 W blower edition for workstations, and NVIDIA allows up to four of them in one, as our guide to a workstation with two RTX PRO 6000 Max-Q cards describes. In a rack, the Max-Q belongs only in a chassis whose maker lists it, and for continuous training we recommend the Server Edition. A common split is a workstation per developer for interactive work and the shared server for every run that takes hours. Our comparison of a GPU server or workstations for a team works through that decision with numbers.

Software, versions and offline operation

On 24 October 2026 the latest tagged Isaac Lab release on GitHub is v3.0.0-EA of 16 September, built for Isaac Sim 6.1, and its release notes target general availability for late October 2026. The README maps the release/3.0.0 branch and main to Isaac Sim 6.2. The installation guide asks for Python 3.12 and lists driver 595.58.03 on Linux x86_64. That driver is older than 595.91.07, the R595 release NVIDIA's security bulletin of 30 September 2026 names as fixed, so test a patched driver before rollout; our guide to NVIDIA driver branches and CUDA versions explains the branches.

• Containers. Isaac Lab's Docker setup needs Docker Engine, Docker Compose and the NVIDIA Container Toolkit, tested with Docker Engine 26.0.0 and Compose 2.25.0. For HPC clusters the guide converts the image into an Apptainer image.
• Offline networks. Isaac Sim needs an internet connection to reach its online assets and to run some extensions. For an isolated robotics network, mirror the assets first, as our guide to an air-gapped server with offline installation describes.
• Licences. The Isaac Lab framework is released under BSD-3 and the isaaclab_mimic extension under Apache 2.0; isaaclab_mimic needs cuRobo, which has proprietary terms, and workflows that use Isaac Sim include proprietary components. The kit-less image is the only one without the Isaac Sim EULA. Which terms apply to your use is for your legal department to confirm.
• Delivery. The server arrives on request with the operating system, the NVIDIA driver, CUDA and a container runtime installed, on the branch your other hosts run.

Machinery Regulation and AI Act for robot makers

Regulation (EU) 2023/1230 on machinery repeals the Machinery Directive 2006/42/EC, and the Commission's machinery page states that it "applies on a mandatory basis as of 20 January 2027"; machinery placed on the market before that date follows the directive. Its recitals 54 and 55 explain that systems with self-evolving behaviour ensuring safety functions belong in Annex I and call for third-party conformity assessment, which should apply only to systems with "a fully or partially self-evolving behaviour using machine learning approaches", not to software that only executes programmed automated functions.

The AI Act links to the same products. Regulation (EU) 2026/1744, the Digital Omnibus on AI, moves the Machinery Regulation from Section A to Section B of Annex I of the AI Act, where the Commission's AI Act Service Desk now lists it. Recital 42 explains that the Commission should adopt delegated acts amending Annex III of the Machinery Regulation and that those acts "should apply by 2 August 2028". For high-risk AI systems related to products in Section B, the amended Article 2(2) of the AI Act states that "only Article 6(1), Article 60a and Articles 102 to 112 shall apply", and Articles 57 to 59 only in so far as the requirements for high-risk AI systems have been integrated in that legislation. Our article on the EU AI Act gives the timeline in general terms. For the GPU server this has a practical side: the training environment that produced a policy for a safety function is part of what a manufacturer may need to describe. Keep the versions of Isaac Sim, Isaac Lab, the task configuration and the container images with every training run. This is general information, not legal advice; your legal counsel or notified body confirms what applies to your product.

Power and room

By our estimate, the two-card server draws about 1.8 kW at the wall and fits one 230 V 16 A feed, which carries 2.94 kW at 80 per cent. Four cards at 600 W draw about 3.4 kW and need 32 A or three-phase. Eight cards draw about 6.3 kW by our estimate; Dell's datasheet gives about 8 kW per eight-card node as an indicative figure, so plan the feeds on the maker's figure. Reinforcement learning keeps every card at full load for hours, so the noise and heat belong in a server room, as our guide to GPU server power on 16 A, 32 A and three-phase feeds works through. A power estimate is confirmed by a measurement on the delivered system.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the RTX PRO 6000 in its Server Edition, Max-Q and Workstation Edition, the L40S and the H200 NVL, and build servers to order with two, four or eight cards, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us. Quantities and lead time are confirmed in writing with the quote. Tell us which robots, sensors and training libraries you use, and we propose the server and the workstations around it.

FAQ