AI server for automotive R&D: GPU CFD, AI surrogates, driving simulation and perception

What an AI server does in automotive R&D

An AI server in automotive R&D carries the GPU work of an engineering department: CFD on GPU solvers, AI surrogate models, driving and sensor simulation, perception training and an assistant over specifications and test reports. For most of this work, a server with four NVIDIA RTX PRO 6000 Blackwell Server Edition cards is the starting point, because the solvers and models run in single or mixed precision and each card brings 96 GB. Codes that compute in double precision, and training of large surrogates, point to the NVIDIA H200 NVL, so many departments end up with two card types.

This guide is written for the head of an engineering or simulation department at a carmaker or a supplier, and for the IT team that will run the servers. It explains the work in plain terms, what the solver and framework vendors publish about GPUs, and how four- and eight-card servers divide the jobs. Our detailed comparison of the GPU for CFD simulation carries the solver sources; here they meet the AI side of the same department.

The work, for the head of department

An aerodynamics team runs many variants of a body, a mirror or an underbody panel and compares drag and lift. A GPU solver shortens each run, as the Ansys measurement below shows, so the team can test more variants before a design freeze. A thermal team does the same for cooling air paths and battery housings.

Once a team has hundreds of finished simulations, it can train an AI surrogate on them: a model that predicts pressure and wall shear stress for a new shape without a full run, so that designers see a first estimate before they request a full run. The CFD runs stay the reference; the surrogate narrows down which variants deserve them.

Driver-assistance teams work with simulated traffic and sensors, cameras, radar and lidar, and train detection models on recorded and synthetic images. Every team also writes and reads specifications, test reports and change requests, where a language model on the department's own documents saves search time without sending confidential prototype data to an outside service.

These jobs share a property that matters for the server: a simulation or a training run keeps a card at full load for hours or days. The number of runs that should progress in parallel sets the card count.

CFD on GPUs: what the solver vendors publish

Ansys's Fluent GPU solver needs the whole case in GPU memory, and several cards in a server pool their memory for one case. Its buying guide, last modified on 9 April 2026, plans 1.2 GB per million hexahedral cells in single precision with the segregated solver and 3.6 GB in double precision with the coupled solver, as our CFD guide quotes.

GB per million cells from Ansys's Fluent GPU solver hardware buying guide (9 April 2026), as quoted in our CFD guide. Cell counts are our division of 384 GB and 564 GB of nominal card memory, rounded down, and are upper limits before process overhead.

The published measurement for these cards comes from Ansys itself. Its blog of 22 July 2026 ran Fluent 2026 R1 on a server with four RTX PRO 6000 Blackwell Server Edition cards against CPU baselines; on the DrivAer_50M case, about 50 million cells, the four cards delivered about 4.6 times the speed of a baseline of roughly 516 CPU cores. Ansys calls these configurations "a more accessible and cost-efficient option for smaller or mixed workloads"; the H200 NVL was not part of the test. DrivAer is a generic car body, so the figure guides external aerodynamics without forecasting every case.

Two Fluent details shape the budget. Ansys licenses the GPU solver by streaming multiprocessors, with 40 SMs included in the CFD Enterprise licence, and the guide counts 188 SMs for the RTX PRO 6000 Blackwell, so a four-card server has 752 SMs by our arithmetic; ask your Ansys channel for the licence count before you compare configurations. Ansys also asks for at least as much system memory as total GPU memory, which our four-card configuration with 768 GB meets twice over.

For Simcenter STAR-CCM+, Siemens's release post for version 2602 of 24 February 2026 added GPU-native Volume of Fluid and Mixture Multiphase solvers, and version 2606 added a GPU-native complex chemistry solver and GPU acceleration of the Lagrangian solver with spray models. Siemens wrote in June 2024 that one Power Session Plus licence covers unlimited CPUs or GPUs. The posts we read do not state the floating-point precision of the GPU-native solver, so confirm it with Siemens for the physics models you validate.

Precision decides the card for the rest. NVIDIA's RTX Blackwell PRO architecture whitepaper states that "The FP64 TFLOP rate is 1/64th the TFLOP rate of FP32 operations", about 1.9 TFLOPS on the Server Edition by our arithmetic, while NVIDIA rates the H200 NVL at 30 TFLOPS of FP64. For structural, crash or in-house codes, ask the developers whether the GPU kernels compute in FP64; our FP64 GPU server guide compares the cards that carry full double precision.

Send us your solver, its version, the cell count of your largest case and its precision, and our team replies with a configuration that holds the case on the cards.

AI surrogates trained on simulation results

NVIDIA's PhysicsNeMo describes itself as "an open-source PyTorch framework for physics machine learning", under Apache 2.0, and its repository lists DoMINO for "Surface and volume field prediction for automotive aerodynamics". The DoMINO example, in the 25.11 documentation, predicts pressure and wall shear stress on the car's surface and velocity and pressure in the volume around it. It trains on DrivAerML, which NVIDIA describes as "500 parametrically morphed variants" of the DrivAer notchback with "mesh sizes of 160M points per example", an open dataset under CC-BY-SA.

NVIDIA PhysicsNeMo repository on GitHub and DoMINO example documentation (version 25.11), read on 24 October 2026.

The documentation states that over several releases the training time "decreased from about 5 days to just over 4 hours, with eight H100 GPUs", and its training guidelines list "A couple of days on a single node of H100 GPU" as the duration. We do not supply the H100, and no figure is published for the H200 NVL or the RTX PRO 6000 Server Edition, so treat the times as the published order of magnitude. What carries over is the pattern: surrogate training reads large meshes, runs for hours on several cards and shards data across them, which suits the eight-card H200 NVL training node with NVLink bridges or a four-card RTX PRO 6000 Server Edition server for smaller models. NVIDIA's MIG guide states that NCCL is not supported with MIG, so such training uses whole cards.

Your own results are the training data. A department's archive of Fluent or STAR-CCM+ runs on its own vehicles holds more value for a surrogate than a public dataset, and it stays on the department's server. Our fine-tuning server guide covers memory and multi-GPU training in general.

Driving simulation and perception training

CARLA describes itself as "an open-source driving simulator for use in autonomous driving applications", built as a client and server, with sensors "from cameras to radars, lidar and many more". Its documentation for the Unreal Engine 5 version requires Ubuntu 22.04 or 24.04 or Windows 11 and 130 GB of disk, and recommends "at minimum an NVIDIA RTX 3000 series or better with at least 16 Gb of VRAM". It adds that "A dedicated GPU, separate from the GPU used for CARLA, is highly recommended to handle large machine learning workloads", and its Docker variant runs off-screen without a display, with several servers on different GPUs.

On a four-card RTX PRO 6000 Server Edition server, that gives a simple layout: two cards run CARLA servers for test scenarios and synthetic sensor data, two train or evaluate the perception models. NVIDIA's Omniverse and Isaac Sim need RT cores as well, as our guide to Omniverse system requirements explains, so simulation and rendering belong on RTX PRO cards, while the H200 NVL, without RT cores, stays with training.

Perception training itself follows the rules in our guide to training vision models: detection models of tens of millions of parameters train data-parallel over PCIe, and memory per card goes to batch size and image resolution, where 96 GB per card helps with high-resolution camera frames. Sensor recordings set the storage, and the size depends on the cameras, the lidar and the hours recorded, so we size the NVMe and the file server from your logging setup.

An assistant over specifications and test reports

Engineering departments write and search large volumes of text: requirements, test specifications, test reports, change requests and supplier documents. A language model with retrieval over these documents runs next to the simulations. By our users-per-card rule, 0.9 × driver-visible memory − 3 GiB per card − weights, gpt-oss-120b on two RTX PRO 6000 Server Edition cards as one tensor-parallel pair holds 93 conversations of 32K tokens, and on one card 19. With the example values of our mid-sized company guide, about 4 requests in flight per 100 employees at the busiest moment, a department of 300 engineers has about 12, which one card carries. Document access rights and the query log are part of our Private AI/ML service.

Four configurations for an automotive R&D department

These start from the reference configurations on our AI servers page. Memory and processor cores follow NVIDIA's configuration guide for NVIDIA-Certified Systems: at least twice the GPU memory in system memory, at least six cores per GPU and two sockets.

Reference configurations of our AI servers page and the NVIDIA-Certified Systems configuration guide (30 September 2026). Wall power is our estimate: cards at 600 W plus two AMD EPYC 9355 (four cards) or 9555 (eight cards), plus 15 per cent for the rest.

Four RTX PRO 6000 Server Edition suit a department that starts with GPU CFD and wants the other jobs on the same machine: one Fluent case across four cards by night, CARLA and perception training by day, with a card for the assistant where needed. Eight RTX PRO 6000 Server Edition double the case size and fit NVIDIA's Omniverse server specification of eight Server Edition cards; our page on eight RTX PRO 6000 Server Edition covers the chassis. Four H200 NVL with a 4-way bridge carry the double-precision codes and surrogate training for a team; our four-card H200 NVL server shows what one NVLink domain holds. Eight H200 NVL is our training node, with 400G networking for a multi-node cluster.

A department that runs both kinds of work gets two servers, an RTX PRO 6000 Server Edition machine for simulation, sensor simulation and inference and an H200 NVL machine for FP64 and training. Our comparison of four RTX PRO 6000 or four H200 NVL sets the cards side by side.

Details for the IT team

1. Processors and memory. Each Fluent compute process takes about 95 MB of GPU memory and 210 MB of system memory, according to Ansys's guide, and Ansys asks for system memory of at least the total GPU memory. CARLA's servers and the data loaders of perception training use processor cores too, so we size the processors from the codes; our guide to EPYC or Xeon for a GPU server explains the trade-offs.
2. Remote pre- and post-processing. The Server Edition is the only RTX PRO 6000 edition with vGPU, from vGPU 19.0. NVIDIA's RTX vWS sizing guide places two to four heavy CAD users on one GPU, as our guide to GPU servers for engineering firms explains, and RTX vWS is licensed per concurrent user.
3. MIG. Up to four 24 GB instances on every RTX PRO 6000 edition and up to seven on the H200 NVL, for small single-GPU jobs. NCCL does not run on MIG, so multi-card CFD and training use whole cards.
4. Storage. Case files, results and training data set the storage: one DrivAerML example alone has a mesh of 160 million points. Local NVMe holds running jobs; finished results move to the department's file server. Our guide to NVMe storage for a GPU server covers drive classes.
5. Network. Our training node carries 400G networking for multi-node clusters. A single four-card server usually needs 25 GbE for users and a faster link only for its storage.
6. Software on delivery. The server arrives on request with the operating system, the NVIDIA driver, CUDA and a container runtime. Keep the driver at a release NVIDIA lists as fixed in its security bulletin of 30 September 2026, and check that your solver version supports it before you update.

Power, cooling and continuous load

A simulation or a training run keeps its cards near full power for hours or days. By our estimate a four-card server with two EPYC 9355 draws about 3.4 kW at the wall with the cards at 600 W, or about 2.7 kW for RTX PRO 6000 Server Edition cards at the 450 W cap, and an eight-card server with two EPYC 9555 about 6.3 kW. We plan a 230 V 16 A feed at 80 per cent, 2.94 kW, so a four-card server at full power needs a 32 A single-phase or three-phase feed. Dell's technical guide for the PowerEdge XE7740 limits the inlet temperature to 30 °C with H200 NVL cards, so the room counts as much as the feed. Our guide to GPU server power on 16 A, 32 A and three-phase works through redundant supplies.

Ordering the servers from us

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. Sizing is free. We assemble each server, test it before delivery and deliver it anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us and the assembly covered by us; other destinations go through our logistics partner. On request we install the operating system, NVIDIA driver, CUDA and a container runtime, and commission the server on site. Quantities and lead time are confirmed in writing with the quote.

To quote the exact configuration we need the solvers and versions, the precision each runs in, the cell count of your largest case, how many runs should progress at once, the frameworks for surrogates and perception, and the power feed at the rack position. Describe these in the form below, and we reply with a configuration and a written quote.

FAQ