GPU for FEA in 2026: which card Ansys Mechanical and Abaqus need, and when FP64 decides

Which GPU Ansys Mechanical needs

For implicit structural analysis in Ansys Mechanical, the solver decides the card. The sparse direct solver and the eigensolvers built on it need a GPU with strong FP64, which among the cards we supply means the NVIDIA H200 NVL. The PCG and JCG iterative solvers and the mixed solver run well on any card Ansys lists, including the NVIDIA RTX PRO 6000 Blackwell, which Ansys marks as not recommended for sparse direct.

Ansys publishes these rules in its installation guide, under "Requirements for the GPU Accelerator in the Mechanical APDL Application". We read the 2026 R1 Linux page and the Windows page on 25 October 2026. The guide states that "only GPU cards with significant double precision performance (FP64) are recommended for use with the sparse direct solver" and the eigensolvers based on it, and it recommends at least 16 GB of memory on the card for meaningful acceleration.

Ansys 2026 R1 Installation Guide, "Requirements for the GPU Accelerator in the Mechanical APDL Application", Linux and Windows pages, read 25 October 2026. "From us" is our supply list.

The table separates two kinds of work. A team that solves most models with PCG or JCG gets GPU acceleration from an RTX PRO card in a workstation. A team whose models need sparse direct, for example modal analyses with Block Lanczos, plans an H200 NVL. Which solver a model uses is a decision for the analyst, and the solver choice in the existing models shows what the GPU will see.

FP64 rates behind the Y and N

NVIDIA rates the H200 NVL at 30 TFLOPS of FP64 and 60 TFLOPS on its FP64 Tensor Cores, with 141 GB of HBM3e at 4.8 TB/s. NVIDIA's RTX Blackwell PRO architecture whitepaper states that "The FP 64 TFLOP rate is 1/64th the TFLOP rate of FP 32 operations", which gives the RTX PRO 6000 Server Edition, rated at 120 TFLOPS of FP32, about 1.9 TFLOPS of FP64 by our arithmetic. For the solver Ansys reserves for strong FP64 cards, a factor of about 16 between the cards decides. Our FP64 GPU server comparison sets these rates beside the B300 and the cuBLAS FP64 emulation.

Ansys's table prints its own FP64 and FP32 figures, and its footnote warns that they come from public datasheets, "may contain inaccuracies" and should be checked with the maker. Three rows show why the Y and N columns, and NVIDIA's own figures, are the parts to rely on:

• The row "NVIDIA H200" carries 34 and 67 TFLOPS, NVIDIA's figures for the H200 SXM. Ansys names no form factor; by our reading the row covers the H200 NVL as well, which NVIDIA rates at 30 TFLOPS of FP64. Confirm this with your Ansys channel if a support case depends on it.
• The row for the RTX PRO 6000 gives 1.9 and 126 TFLOPS and 1,340 GB/s. NVIDIA gives the Workstation Edition 1,792 GB/s and the Server Edition 1,597 GB/s.
• The B300 is marked "Y" for sparse direct with the figures "75/1.2". NVIDIA's HGX table lists 10 TFLOPS of FP64 and 600 TFLOPS of FP32 for eight B300 GPUs, so 75 is the FP32 value per GPU and about 1.25 the FP64 value, by our division. That FP64 figure is below the RTX PRO 6000's 1.9, which Ansys marks "N". We cannot explain the mark from the documents, so treat the B300 as an AI system and ask Ansys before planning it for sparse direct.

The L40S is not listed at all, and Ansys's footnote adds that GPUs released after the software release are not listed. To use any unlisted card, Ansys documents the environment variable ANSGPU_OVERRIDE=1 and asks you to make sure the card is "sufficiently powerful, in terms of both double-precision compute power and on-card memory"; the page also warns that using the variable "with an underpowered CPU may actually decelerate your simulation".

How Mechanical APDL uses the GPU

The GPU does not replace the processors. The 2025 R2 Parallel Processing Guide states that "a GPU does not replace the CPU core(s) on which a simulation typically runs": the CPU handles most operations and "will automatically offload some of the time-intensive parallel operations performed by certain equation solvers". By our reading, a GPU server for FEA therefore still needs fast cores and enough system memory.

The 2026 R1 guide lists what the GPU accelerates. Static linear and nonlinear analyses with the sparse, PCG, JCG or mixed solver; buckling with Block Lanczos or subspace; modal analyses with those and with PCG Lanczos, PCG Block Lanczos, QR damped, unsymmetric and damped; full harmonic analyses with the sparse solver; full transient analyses with sparse, PCG, JCG or mixed; and the generation pass of substructuring, CMS included. For any other analysis type "the solution will continue but GPU acceleration will not be used". The same applies when the sparse solver switches to partial pivoting, which the guide names for mixed u-P elements, Lagrange-multiplier MPC184 elements, Lagrange-multiplier contact elements and certain circuit elements.

Ansys 2026 R1 Parallel Processing Guide, section 3.2, "Supported GPU Hardware, Analysis Types, and Features", read 25 October 2026.

This table changes the server design. A sparse direct job in shared-memory mode uses exactly one H200 NVL, so a two-card server runs two such jobs side by side, or one distributed job across both cards. The 2025 R2 guide activates the GPU with -acc nvidia -na N on the command line, N being the number of GPUs per machine, and uses one GPU when only -acc is given. It also notes that the Mechanical APDL GUI on Linux is not supported with GPU acceleration, that GPU use must be the same on every node of a multi-node run, and that the program does not check whether another application already uses a GPU, so a shared server needs a scheduler that assigns cards.

We build servers with two or four H200 NVL for sparse direct work and RTX PRO workstations for iterative solvers; send us your model sizes, solvers and the number of analysts, and our team replies with a configuration.

Licences, drivers and Windows

Ansys counts HPC licences per CPU core and per GPU. The 2025 R2 Parallel Processing Guide states that "HPC licensing for Mechanical APDL is per CPU core and GPU (not per GPU compute engine)" and that up to four CPU cores and GPUs combined run without HPC licences, one core and three GPUs for example. A run with 32 cores and two GPUs therefore counts 34 by our reading. Licence terms change with releases and contracts, so confirm the 2026 count with your Ansys channel before you size the processors.

The 2026 R1 installation bundles the CUDA 13.0.2 libraries, so no separate CUDA installation is needed, but it does not install the driver. Ansys requires driver 580.95.05 or newer on Linux and 581.42 or newer on Windows. On Windows the guide recommends the TCC driver mode "when using Tesla series GPU cards" and warns that Remote Desktop, or a solve launched through RSM running as a service, may disable GPU use. Our article on WDDM, TCC and MCDM driver modes explains how to switch modes and which cards start in TCC; the choice of operating system is covered in Windows Server or Linux for a GPU server.

Abaqus on GPUs

Abaqus/Standard has accelerated its direct sparse solver on NVIDIA GPUs since release 6.11, as NVIDIA's whitepaper "Accelerating Abaqus Computations Using NVIDIA GPUs" (WP-07843-001_v03, August 2016) records. It describes the method: the solver offloads the dense matrix operations to the GPU, while assembly, tree traversal and the forward and backward solve stay on the CPU. For Abaqus 2016 it adds GPU support for the AMS eigensolver and a frequency response solver. The paper advises GPUs with "high double-precision capacity", reports significant speed-ups for models above a million degrees of freedom, and states for that release that a GPU "is counted as a single CPU core" when the licence tokens are calculated. GPUs are requested with the -gpus option, for example -cpus 8 -gpus 1.

The newer step is Abaqus/Explicit. A post in Dassault Systèmes' SIMULIA community of 23 January 2026 announced that "with the release of 2026 FD01, Abaqus is introducing GPU acceleration for Explicit analyses". The post names no cards, precision or licence terms, and Dassault's Abaqus 2026 system information was not reachable for us, so we give no card list for Abaqus 2026. Ask Dassault Systèmes or your reseller for the qualified GPUs of your release, and for Explicit, whether it runs in single or double precision on the GPU. By our reading the FP64 logic of the Ansys sparse solver applies to the Abaqus/Standard direct solver as well, so the H200 NVL is the card to plan for it.

FEA GPU server or workstation: configurations

Card memory from NVIDIA's product pages; wall power is our estimate (cards at 600 W, two AMD EPYC 9355 for the H200 NVL servers and two EPYC 9135 for the RTX PRO server, plus 15 per cent), as in our two- and four-card server guides, and is confirmed by a measurement on the delivered system.

Workstation for iterative solvers. One or two RTX PRO 6000 Workstation Edition or Max-Q cards accelerate PCG and JCG and also drive the CAD and post-processing screens. Up to four Max-Q cards fit one workstation, according to NVIDIA. If the models fit in 72 GB, the RTX PRO 5000 72 GB is the smaller choice, as our guide to buying an RTX PRO 5000 72 GB workstation explains. Graphics choices for the same desk are in our GPU for CAD guide.

Server with two or four H200 NVL. This is the configuration for sparse direct and the sparse-based eigensolvers. We quote host memory at twice the GPU memory by NVIDIA's certified-systems guidance, 768 GB for two cards and 1,152 GB for four, and at least six cores per GPU on two sockets; the FEA processes add their own memory on top, so we size it from your largest model. The NVLink bridges of the H200 NVL matter for NVLink-aware codes; we found no Ansys statement that the Mechanical APDL sparse solver uses them. Our guides to the two-card H200 NVL server and the four-card H200 NVL server cover chassis, power and cooling.

Mixed team. Where analysts solve iteratively at the desk and send large direct solves to a shared node, RTX PRO workstations plus one H200 NVL server cover both. Our comparison of GPU servers and workstations for a team works through the shared-node layout, and our CFD simulation guide covers Fluent, STAR-CCM+ and OpenFOAM, which follow different rules. AI surrogate models trained on FEA results are a separate workload, covered in our article on PhysicsNeMo for CAE.

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 H200 NVL and the RTX PRO 6000 in its Workstation, Max-Q and Server Editions, and build workstations and servers to order, assembled and tested before delivery, on one EU contract and invoice, with manufacturer warranty on every component and the assembly covered by us. Operating system and driver are installed on request. Quantities and lead time are confirmed in writing with the quote. Tell us which solvers your models use and how large they are, and we reply with a configuration and a written quote.

FAQ