WDDM, TCC and MCDM: NVIDIA driver models on Windows, how to switch, which one you need

What WDDM, TCC and MCDM are and how they differ

WDDM, TCC and MCDM are the three driver models an NVIDIA GPU can run under on Windows. WDDM is the display driver model: the card drives monitors and serves graphics and compute together. TCC and MCDM are compute-only: TCC runs the card with quicker kernel launches and no display, and MCDM, the newer model, also makes the card usable from WSL2, which TCC does not.

NVIDIA's nvidia-smi documentation, read on 25 October 2026, states three things: "On Windows, the TCC, WDDM and MCDM driver models are supported", "The TCC driver model is optimized for compute applications i.e. kernel launch times will be quicker with TCC", and "The WDDM driver model is designed for graphics applications and is not recommended for compute applications."

On Linux there is nothing to choose. The same page states that "Linux does not support multiple driver models", and both fields in nvidia-smi -q, Current and Pending, always read "N/A" there. This article is therefore about Windows workstations and Windows servers with RTX PRO or data-centre cards, the typical case being a simulation or CAE package that only ships for Windows or behaves differently by mode.

NVIDIA nvidia-smi documentation (driver model section) and CUDA on WSL User Guide (last updated 12 October 2026), both read 25 October 2026; NVIDIA developer forum, staff reply of 11 February 2025, on MCDM and WSL2.

Which model a card starts in

For data-centre cards the answer is in NVIDIA's Windows driver release notes. The R595 data-centre driver 597.30, released on 22 October 2026, lists among its limitations that "the default TCC mode in the NVIDIA driver does not support IOMMU-based isolation", which Windows features such as DMA protection rely on, and names the L40, L40S, L20, L4 and RTX PRO 6000 Blackwell Server Edition as affected. The R580 release 582.70 of June 2026 carries the same note. Both releases support Windows Server 2025 and Windows Server 2022, and the R595 notes list the RTX PRO 6000 Server Edition, the RTX PRO 4500 Server Edition, the RTX PRO 6000 Workstation and Max-Q editions, the H200 NVL, the L40S and the L4 among their GPUs.

NVIDIA Data Center GPU Driver release notes 595.112.07 (Linux) / 597.30 (Windows), page updated 21 October 2026, and 582.70 (Windows, R580); nvidia-smi documentation; all read 25 October 2026.

The same notes add a second fact that decides many Windows plans: "for any WDDM 2.0+ based functionality on Data Center GPUs, vGPU is required". In plain terms, a passive server card under the bare-metal data-centre driver is a compute device; graphics through WDDM on such a card is a vGPU case, with the vGPU (formerly GRID) driver and a vGPU licence, which our guide to NVIDIA vGPU on Hyper-V with Windows Server 2025 covers for virtual machines. The notes do also state that data-centre products support one display of up to 2560 × 1600.

For workstation cards, the card that drives the monitors is in WDDM, and it has to stay there: nvidia-smi's -dm "will fail if a display is attached". CST's GPU guide puts it from the user side: TCC "is not available for Quadro cards which are connected to a display/monitor". We found no NVIDIA page that lists TCC or MCDM support per RTX PRO Blackwell workstation card, so on a multi-GPU workstation the reliable check is nvidia-smi -q on the card without a monitor.

We supply the RTX PRO 6000 Server Edition, the L40S and the L4 for Windows Server hosts and the RTX PRO workstation cards for Windows workstations; ask us for a card layout per host, and our team replies with the cards and the driver model each one runs in.

Enabling TCC and switching with nvidia-smi -dm

The switch is one nvidia-smi option, run from a command prompt with administrator rights. NVIDIA documents the syntax as -dm TYPE (or --driver-model=TYPE) and a forcing variant, -fdm TYPE, that changes the model even with a display attached. The numeric values are not on NVIDIA's documentation page; CST's GPU Computing Guide 2026 gives 1 for TCC and 0 for WDDM, and an NVIDIA staff reply on NVIDIA's developer forum gives 2 for MCDM (written with -g), recommending a driver of the 570 series or later.

CST Studio Suite 2026 GPU Computing Guide (Dassault Systèmes, change history to 12 June 2026); NVIDIA developer forum thread "Enabling WSL2 on TCC GPU (L4) by switching to MCDM", February 2025; NVIDIA nvidia-smi documentation; all read 25 October 2026.

Four details from NVIDIA's page prevent surprises. Without -i the change "will impact all GPUs". The driver is restarted "for all GPUs on the system", whichever card changed, so running jobs on other cards stop. With --no-driver-restart the change waits for a manual restart or reboot, and a reboot is needed anyway "if the driver restart fails". CST's guide simply tells users to reboot after switching each card. Older guides, such as Altair's nanoFluidX 2022.1 help, write the command with -g for the GPU; a user in the forum thread above reported in 2025 that -g was no longer accepted and used -i.

The most common error is about rights. In a 2022 thread on NVIDIA's forum, a workstation card with no monitor attached returned "Unable to set driver model for GPU …: Insufficient Permissions" for an administrator account; the switch worked from a command prompt started with "Run as administrator", followed by a reboot.

What changes with the model

Display and remote desktop. In TCC and MCDM the card drives no monitor. For remote work, CST's guide states that GPU computing over RDP works with professional and data-centre cards under its conditions, among them a LAN licence and no monitor connected to the GPU, and that cards whose display link is used cannot run CST's GPU computing in a remote desktop session.

WSL2. NVIDIA's CUDA on WSL User Guide (12 October 2026) makes GPU acceleration in WSL2 available "in WDDM mode" and not on GPUs in TCC mode. MCDM addresses this: NVIDIA staff wrote in February 2025 that MCDM makes such GPUs visible in WSL2 and that WSL2 support needs Windows 11 or Windows Server 2025. Users in the same thread who switched an L4 to MCDM on Windows Server 2022 with driver 572.13 reported a warning on the device in Device Manager after the reboot, so on Windows Server plan WSL2 compute on 2025. Our guide to CUDA on WSL2 for RTX PRO workstations covers the workstation side.

Security features. The IOMMU note above means that on the L40S, L4 and RTX PRO 6000 Server Edition the default TCC mode does not work with Windows DMA protection. If a server's security baseline requires it, raise this before the hardware is chosen.

Things no model changes. Some CUDA limits on Windows are platform limits. NVIDIA's CUDA Programming Guide states that the legacy CUDA IPC API "is only supported on Linux", and that on Windows, unified memory behaves as on devices without concurrent managed access, with a fallback to mapped memory where GPUs cannot map each other's memory. A multi-GPU code that relies on these features belongs on Linux, a decision our article on Windows Server or Linux for a GPU server works through.

What simulation software asks for

CST Studio Suite is the clearest case. Its 2026 GPU Computing Guide says "the GPUs should be operate in TCC mode" when applicable, because in WDDM "two sources (display and the application) are battling for the same hardware"; "the TCC mode is currently required for A-solver GPU computing"; for combined MPI and GPU computing on Windows "the TCC mode of the GPU hardware must be enabled"; and "GPUs running in WDDM instead of TCC mode might eventually fail during memory allocations". The guide lists the RTX PRO 6000 Blackwell workstation card, the RTX PRO 5000 Blackwell and, from 2026 SP5, the RTX PRO 4500 Blackwell among supported cards, with reduced double-precision performance marked, and the H200 NVL and L40S among server cards; it does not list the RTX PRO 6000 Server Edition. Card choice for CST and HFSS is the subject of our article on GPUs for EM simulation with CST Studio and HFSS.

Other vendors are less strict. Altair's nanoFluidX help for 2022.1 describes WDDM as the usual default on workstations and TCC as compute-only, and reports that its tests "have not shown a distinct advantage in using TCC mode as of version 2021.2": no significant difference on a desktop card, and about three times the run time in TCC on a mobile GPU. Check the GPU guide of each solver version you run; the mode it wants decides whether a workstation needs a second card. For CFD sizing see our comparison of GPUs for CFD simulation, H200 NVL and RTX PRO 6000, and for FEA GPUs for Ansys Mechanical and Abaqus.

We supply RTX PRO workstation cards and Server Edition cards, alone or in workstations and servers built to order; send us the solver, its version and the number of users, and our team replies with a card layout that fits the driver model the software requires.

Cloud instances and VMs stuck in MCDM or TCC

A thread on NVIDIA's developer forum from August 2026 shows the server-card side. A user on a public-cloud Windows instance with an RTX PRO 6000 Blackwell Server Edition found the card in MCDM, needed WDDM because a simulation package would not start otherwise, and could not switch it with nvidia-smi; DisplayModeSelector offered only the compute mode. Installing the GRID (vGPU) driver solved it, which is consistent with NVIDIA's note that WDDM functions on data-centre GPUs need the vGPU driver.

For your own servers the rule follows from the same note. A Windows VM that should show graphics on a Server Edition card, an L40S or an L4 gets a vGPU profile or the vGPU guest driver with a licence; a Windows compute VM or a bare-metal Windows Server can stay in TCC with the data-centre driver. Which hypervisors and profiles NVIDIA supports per card is in our NVIDIA vGPU hypervisor support by GPU overview.

Display mode is a different switch

The driver model is often confused with the display mode. NVIDIA's DisplayModeSelector tool switches a card between graphics mode, with display outputs active, and compute mode, with display outputs off. The RTX PRO 6000 Server Edition and the L40S ship in display-off mode and must stay in it for vGPU, per NVIDIA's vGPU validated-platforms pages, and MIG on the RTX PRO 6000 and RTX PRO 5000 requires Linux and compute display mode, as our MIG step-by-step guide sets out with the tool versions. The two settings are separate: DisplayModeSelector sets the display mode, nvidia-smi sets the driver model, and only the driver model is specific to Windows. In the forum thread above, a reply suggested the tool's --listgpumodes option to see which modes a card allows.

Which setup for which Windows task

our summary of the NVIDIA, CST and Altair statements cited above, October 2026.

On a workstation the clean layout for CST or a similar solver is two cards: one in WDDM for the monitors and one or more compute cards in TCC. NVIDIA allows up to four RTX PRO 6000 Max-Q cards in one workstation, so a solver workstation can carry several compute cards behind one display card. Our RTX PRO 6000 Workstation, Max-Q and Server Edition comparison has the power and cooling differences. For a rack server under Windows Server, the passive Server Edition stays a compute device in TCC, and graphics users reach it through vGPU.

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 and the H200 NVL as cards with manufacturer warranty, and build workstations and AI servers to order, assembled and tested before delivery, with the operating system and drivers installed on request. Quantities and lead time are confirmed in writing with the quote, on one EU contract and invoice.

FAQ