GPU for EM simulation in 2026: CST Studio Suite and HFSS cards, solvers and FP64

Which GPU for CST Studio Suite and HFSS

For CST Studio Suite the solver decides the card: the time domain solver runs on every card in CST's list, including the NVIDIA RTX PRO 6000 Blackwell with 96 GB. The integral equation solver, the frequency domain solver with domain decomposition, the multilayer solver and the asymptotic solver need strong FP64, which among the cards we supply means the NVIDIA H200 NVL. For HFSS, Ansys publishes which solvers use a GPU but, in the documents we opened, no card list with a precision rule.

The CST facts below come from the CST Studio Suite GPU Computing Guide 2026, which Dassault Systèmes publishes as a PDF; its change history runs to 12 June 2026, and we read it on 25 October 2026. The guide marks some cards with a footnote: "GPU device with reduced double precision performance; please check chapter 2 Supported Hardware for affected solvers."

CST Studio Suite GPU Computing Guide 2026, supported hardware table, change history to 12 June 2026, read 25 October 2026. Rows limited to cards we supply; the guide also lists B200, H200 SXM, H100 and older cards.

Three gaps in the list matter when ordering. The RTX PRO 6000 Blackwell appears on the "Workstation" platform; the guide names no edition, and its specification row (96 GB, about 1,800 GB/s) matches the Workstation and Max-Q Editions with their 1,792 GB/s, while the Server Edition has 1,597 GB/s. Neither the Server Edition nor the Max-Q is named. The RTX PRO 5000 appears with 48 GB only, and the L4 does not appear. Before you buy a card that is not named, ask Dassault Systèmes or your CST reseller whether it is supported for your release.

Solver by solver: where FP64 is needed

Chapter 2 of the guide lists the solvers with NVIDIA GPU support and adds the precision requirement where there is one. We read the FP64 rates beside it from NVIDIA: the H200 NVL computes 30 TFLOPS of FP64, while the RTX PRO 6000 Server Edition computes FP64 at 1/64 of its 120 TFLOPS of FP32, about 1.9 TFLOPS by our arithmetic, a factor of about 16.

CST Studio Suite GPU Computing Guide 2026, chapter 2; FP64 rates from NVIDIA's H200 page and RTX Blackwell PRO architecture whitepaper. Card column is our reading.

The guide's own example for the frequency domain solver names the B200 and the H100, neither of which we supply; the H200 NVL is on its list without the reduced-precision mark. On AMD cards the guide supports only the time domain FIT solver. For a team whose work is mostly time domain, an RTX PRO 6000 Workstation or Max-Q Edition is the card; a team that also runs large MoM or MLFMM models plans an H200 NVL for those, or runs them on the CPU.

Mesh cells per GB for the time domain solver

CST gives one sizing rule, for the time domain solver with FIT: "available memory divided by 100", illustrated with "a GPU with 16 GB of Memory will be sufficient for a simulation with 160 million meshcells". That is 10 million mesh cells per GB. The guide adds that memory use depends on the features used, so the figures below are upper estimates.

CST rule of thumb from the GPU Computing Guide 2026; cell counts are our arithmetic from the nominal card memory and are upper estimates.

The guide supports "up to 16 GPU devices in a single host system", and its footnote recommends contacting CST "before purchasing a system with more than four GPU cards". It also states that cards "of different generations (e.g. "Ampere" and "Volta") can't be combined in a single host system for GPU Computing". A Blackwell RTX PRO 6000 and a Hopper H200 NVL therefore belong in separate hosts for CST, even though our guide to H200 NVL and RTX PRO 6000 in one platform shows such mixes for other software. Running several simulations on one card at the same time "will deteriorate the performance", and CST cannot use GPUs in "exclusive mode".

We supply the RTX PRO 6000 in all three editions and the H200 NVL, as cards or in workstations and servers built for CST; send us your solvers, the largest mesh and the number of engineers, and our team replies with a configuration.

CST GPU requirements: drivers, TCC and remote work

The guide names verified drivers: version 581.42 on Windows and 580.95.05 on Linux. Other driver versions "might also work", but it is "highly recommended to use the versions verified by CST", and a different driver must meet the guide's "minimum requirement for the CUDA compute capability of 12.8", which we read as a driver that supports CUDA 12.8. On Linux the NVIDIA driver has to be reinstalled after a kernel update, and the DC Solver Server restarted after a driver installation, "as otherwise the GPUs cannot be detected properly".

On Windows the driver model matters. The guide says that "When applicable the GPUs should be operate in TCC mode", requires TCC for the A-solver on the GPU and for combined MPI and GPU computing, and notes that TCC is not available for a card connected to a monitor. A Windows workstation for CST therefore works best with a separate display card and one or more compute cards in TCC, which our article on WDDM, TCC and MCDM driver modes sets up step by step. GPU computing over Remote Desktop works with professional and data-centre cards as long as no monitor is connected to the GPU, under the guide's licence conditions; with a display attached, "GPU Computing can't be used in a remote desktop session".

GPU computing is licensed separately. The guide states that CST uses acceleration tokens under the CST Studio Suite licence model, or SimUnit tokens or credits under the SIMULIA Unified Licence, and that the number of GPUs one simulation can use depends on the tokens in the licence. How many tokens one GPU takes is not in the guide; your Dassault Systèmes contact confirms it.

HFSS GPU acceleration, Maxwell and Icepak

Ansys's "GPU Accelerator Capabilities" document for release 2025 R1 states that the HFSS frequency domain solver, the time domain solver and the SBR+ solver support "NVIDIA CUDA-enabled workstation, data center, and server cards", and that Maxwell and Icepak do the same. The "What's New in Electronics Desktop" document for 2026 R1 adds three GPU items: the "cuDSS GPU Direct Matrix solver enhanced to support frequency sweeps and multi-GPU distribution (Beta)" for HFSS, a "distributed frequency sweep with GPU acceleration" for the Maxwell A-Phi solvers, and native Joule heating "for CPU and GPU (Beta)" in Icepak.

Neither document gives a card list with precision marks or an FP64 rule for HFSS, so we do not transfer the rules of CST, or of the Mechanical APDL solver in our FEA guide, to HFSS. Ask Ansys which cards are qualified for the HFSS solvers and the release you run, and run your own largest model on the candidate card before a larger purchase. HFSS licensing for GPUs was not stated in the documents we opened.

Workstation or GPU server for electromagnetic simulation

Our reading of the CST Studio Suite GPU Computing Guide 2026; the guide states that passive server GPUs need the extra fans of a server chassis, while workstation GPUs are actively cooled.

Time domain workstation. One or two RTX PRO 6000 cards in the Workstation or Max-Q Edition hold time domain models of up to about 960 million mesh cells per card by CST's rule; NVIDIA allows up to four Max-Q cards in one workstation. Under Windows add a small display card, so the compute cards run in TCC. Our comparison of the RTX PRO 6000 Workstation, Max-Q and Server Editions covers power and cooling.

FP64 server. Where the integral equation or frequency domain solvers carry the large models, a server with two or four H200 NVL gives each card 141 GB and 30 TFLOPS of FP64. By our estimate it draws about 2.0 kW at the wall with two cards and about 3.4 kW with four, with two AMD EPYC 9355, as in our guides to the two-card and four-card H200 NVL server. Our FP64 GPU server comparison explains why the B300 is not an FP64 alternative.

Shared server for several engineers. A rack server with RTX PRO 6000 Server Edition cards would suit a team that sends time domain jobs to a central machine, but CST does not list that edition. If CST confirms it for your release, a two-card server with two AMD EPYC 9135 draws about 1.8 kW by our estimate; otherwise the H200 NVL server is the listed server option. If you send us the solver mix, we check the card against your CST release before we quote. Our comparison of GPU servers and workstations for a team covers the shared-node layout, and CFD codes, which follow different rules, are in our CFD simulation guide.

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 Workstation, Max-Q and Server Editions, the RTX PRO 5000 and 4500 and the H200 NVL as cards, and build workstations and servers to order, assembled and tested before delivery, with the verified driver installed on request. Every component carries the manufacturer's warranty, and the assembly is covered by us, on one EU contract and invoice. Quantities and lead time are confirmed in writing with the quote.

FAQ