Rubin Shrestha
Rubin Shrestha
Blog

RTX 5070 Ti is stable under Linux/Vulkan, but Windows 10 and 11 show UI artifacts and DXGI_ERROR_DEVICE_HUNG immediately after the NVIDIA driver loads

by |Published

I am dealing with a very strange RTX 5070 Ti problem that appears only under Windows.

I have already replaced hardware, performed multiple completely clean Windows installations, tested different operating systems and NVIDIA driver branches, changed PCIe and BIOS settings, and isolated as many variables as I reasonably can.

At this point, I am trying to determine whether this is:

  • a defective GPU block used specifically by WDDM, DWM, DXGI or DirectX;
  • a compatibility problem between the GPU, motherboard and Windows driver;
  • a VBIOS or firmware problem;
  • or some other issue I have missed.

Hardware

  • GPU: ASUS TUF Gaming GeForce RTX 5070 Ti OC, 16 GB
  • GPU PCI ID: 10DE:2C05
  • Subsystem ID: 1043:89F4
  • Current VBIOS: 98.03.58.00.4E
  • CPU: AMD Ryzen 9 5950X
  • Motherboard: MSI MAG B550M Mortar, board ID MS-7C94
  • RAM: 128 GB DDR4
  • PSU: Cooler Master MWE Gold 850 V3, ATX 3.1, using the native 12V-2×6 cable
  • Windows SSD: Kingston NV3 500 GB, model SNV3S500G
  • Linux SSD: Kingston 1 TB, model SNVS1000G
  • Main display: 2560×1440, 165 Hz
  • Dual-monitor setup, although different display connections and simplified display configurations were also tested

Exact symptoms under Windows

Windows works normally while using Microsoft Basic Display Adapter.

As soon as I install an NVIDIA driver and it takes control of the display, parts of the Windows graphical interface begin malfunctioning.

The GPU does not disappear from the PCIe bus. Device Manager recognizes it without a warning icon, and nvidia-smi reports:

  • the correct RTX 5070 Ti model;
  • 16 GB of VRAM;
  • WDDM mode;
  • normal idle clocks;
  • approximately 32 °C;
  • approximately 15 W at idle.

However, some Windows system applications show severe graphical corruption or stop rendering correctly.

Windows Explorer often remains usable, which makes the failure look selective rather than a complete GPU or driver crash. I have also received the Windows notification that an application was blocked from accessing the graphics hardware.

When I launch Assassin’s Creed Black Flag Resynced, the game can initialize and reach its opening screen, but it then crashes with:

DXGI_ERROR_DEVICE_HUNG

Error code:

0x887A0006

I have also seen nvlddmkm-related events, including Event ID 153, during previous Windows installations.

The important point is that the problem starts immediately when the NVIDIA driver takes control of the display.

It happens:

  • before installing the game;
  • before connecting Windows to the internet;
  • before Windows Update can install anything;
  • without third-party antivirus or security software;
  • and even on a completely clean and offline Windows installation.

Operating systems tested

I reproduced the same problem on:

  • Windows 11 25H2, build 26200.8875
  • Official Microsoft Windows 10 22H2 Brazilian Portuguese x64 ISO

The Windows 10 installation was performed from a freshly created USB drive onto a clean GPT SSD.

The machine remained completely offline.

The Windows 10 test sequence was:

  1. Clean installation onto the Kingston NV3 500 GB SSD
  2. No internet connection
  3. No third-party security software installed
  4. AMD B550 chipset driver installed
  5. Restart
  6. NVIDIA driver installed manually
  7. Restart
  8. The graphical corruption appeared immediately

Therefore, Windows Update did not install or replace the NVIDIA driver.

The problem also cannot be exclusive to Windows 11, because the same behavior occurs on a clean Windows 10 installation.

NVIDIA drivers tested

I originally reproduced the problem with NVIDIA driver 610.74.

I then tested the older 572.47 driver using the following procedure:

  1. Disconnected the computer from the internet
  2. Booted Windows into Safe Mode
  3. Removed the current NVIDIA driver using DDU
  4. Restarted normally
  5. Installed only the NVIDIA Graphics Driver and PhysX
  6. Did not install the NVIDIA App or any optional components

The same graphical corruption appeared while the older driver was being installed, as soon as it took control of the display.

Restarting did not change the result.

Therefore, the problem is not exclusive to the newer 600-series driver branch.

Linux behavior

The same GPU works extremely well under Ubuntu with the proprietary NVIDIA driver.

I have tested:

  • Dota 2 through Vulkan at high frame rates
  • gpu-burn without errors
  • FurMark 2 using Vulkan
  • FurMark 2 using OpenGL
  • FurMark Vulkan Knot
  • 2560×1440 with 8× MSAA
  • Sustained high GPU utilization

Example FurMark results:

  • Vulkan: approximately 334 FPS average, around 50 °C, zero artifacts
  • OpenGL: approximately 164 FPS average, around 64 °C, zero artifacts
  • Vulkan Knot: approximately 71–110 FPS depending on the scene, around 56–62 °C, zero artifacts

The card remains stable, cool and free of visible corruption.

There are no black screens, NVIDIA Xid errors, driver resets or rendering artifacts during these Linux tests. Vulkan and OpenGL performance are excellent.

I understand that Vulkan and OpenGL do not necessarily exercise every WDDM, DWM, DXGI, DirectComposition, display-engine, scanout or power-management path used by Windows and DirectX.

However, the card can clearly sustain heavy shader, VRAM and power loads under Linux.

Hardware already replaced

I originally suspected power delivery, so I replaced the previous PSU with a Cooler Master MWE Gold 850 V3 ATX 3.1.

The GPU is now powered through the PSU’s native 12V-2×6 cable.

This removed the following from the equation:

  • the old PSU;
  • the old PCIe power cables;
  • the NVIDIA power adapter.

The Windows storage device was also replaced.

Windows is now installed on a new Kingston NV3 500 GB NVMe SSD.

The same failure occurred after both replacements.

PCIe, BIOS and motherboard tests

The motherboard is using its latest available BIOS.

I tested the following BIOS configurations:

PCIe Gen3

I forced the main GPU slot from Auto/Gen4 to PCIe Gen3.

The problem remained exactly the same.

Resizable BAR disabled

I tested with:

  • PCIe Gen3
  • Resizable BAR disabled
  • Above 4G Decoding enabled

The problem remained exactly the same.

Above 4G Decoding disabled

I then tested with:

  • PCIe Gen3
  • Resizable BAR disabled
  • Above 4G Decoding disabled

The problem remained exactly the same.

After completing these tests, I reset the motherboard BIOS back to its default settings.

Because forcing PCIe Gen3 did not change the behavior, a problem caused only by PCIe Gen4 signal integrity appears less likely.

Disabling ReBAR and changing the Above 4G memory mapping also had no effect.

CPU and socket test

I removed the Ryzen 9 5950X, inspected and cleaned the contacts, and installed it again.

The GPU, RAM and NVMe devices were all recognized normally afterward.

The Windows problem remained exactly the same.

This makes a simple CPU socket or PCIe-lane contact problem less likely, although it may not completely exclude a deeper motherboard or CPU PCIe-controller issue.

Other tests already performed

  • Multiple clean Windows installations
  • Windows installed completely offline
  • AMD chipset driver installed before the NVIDIA driver
  • Latest motherboard BIOS installed
  • GPU reseated
  • CPU removed, cleaned and reseated
  • Display cables changed
  • HDMI tested
  • Different display connections tested
  • RAM tested without detected problems
  • GPU dual-BIOS switch tested in both Performance and Quiet positions
  • Both VBIOS banks report the same version
  • ROMs from both VBIOS positions were saved and produced matching hashes
  • Power limit reduced from approximately 300 W to approximately 250 W
  • MSI Afterburner power limit set to approximately 83%
  • nvidia-smi power-limit testing
  • Windows WPF hardware acceleration disabled
  • Multi-Plane Overlay disabled using OverlayTestMode = 5
  • MPO and WPF changes made no difference
  • PCIe Gen3 forced
  • Resizable BAR disabled
  • Above 4G Decoding disabled
  • Motherboard BIOS reset to defaults after testing
  • Older NVIDIA driver installed after DDU in Safe Mode
  • No NVIDIA App or optional driver components installed during the older-driver test

I also attempted to use the lower PCIe slot, but the card produced no video there.

That slot is chipset-connected and affected by lane sharing on this motherboard, so I do not consider that result conclusive.

An earlier 250 W test appeared more stable, but I could not reproduce that result on the latest clean Windows 10 installation.

Limiting the card to 250 W no longer changes the behavior.

Why I am confused

If the GPU core, VRAM, PSU or basic PCIe communication were badly defective, I would expect at least some instability, artifacts, NVIDIA Xid errors or crashes during sustained Linux Vulkan and OpenGL loads.

Instead, the card performs extremely well under Linux.

Under Windows, the failure appears as soon as the NVIDIA WDDM driver activates, even at desktop idle.

The Windows UI becomes selectively corrupted, while:

  • the GPU remains detected;
  • nvidia-smi continues working;
  • Device Manager reports no error;
  • and applications such as Explorer may remain usable.

A DirectX 12 game then crashes with DXGI_ERROR_DEVICE_HUNG.

The same behavior has now survived:

  • Windows 10 and Windows 11;
  • multiple clean installations;
  • two NVIDIA driver branches;
  • DDU removal in Safe Mode;
  • an offline installation;
  • a new PSU;
  • a new SSD;
  • PCIe Gen3;
  • ReBAR disabled;
  • Above 4G Decoding disabled;
  • CPU and GPU reseating;
  • reduced GPU power limits;
  • MPO disabled;
  • WPF hardware acceleration disabled;
  • both GPU BIOS switch positions.

Remaining possibilities

At this point, the remaining possibilities seem to include:

  • a defective GPU block used specifically by WDDM, DWM, DXGI or DirectX;
  • a display-engine, scanout or low-power-state fault that is not triggered under Linux in the same way;
  • a VBIOS and Windows-driver incompatibility;
  • a motherboard or CPU PCIe-controller problem that only appears with the Windows NVIDIA driver;
  • an undocumented RTX 50-series or Blackwell compatibility issue;
  • or another Windows-specific hardware interaction that I have not identified.

Questions

Has anyone seen an RTX 50-series GPU remain completely stable under Linux, including heavy Vulkan and OpenGL loads, while corrupting Windows system applications as soon as the NVIDIA WDDM driver loads?

Can a defective display engine, scanout block, DirectX-specific engine or low-power-state circuit cause this behavior while the CUDA, Vulkan and OpenGL workloads remain stable?

Are there any Windows logs, ETW traces, LiveKernelReports fields or NVIDIA diagnostic tools that could distinguish a defective GPU from a motherboard or CPU PCIe problem?

Is there any meaningful test left that does not require installing the GPU in another computer?

Would this evidence already justify an RMA, even though the card works normally under Linux?

I can provide:

  • Event Viewer logs
  • nvidia-smi output
  • VBIOS dumps
  • FurMark screenshots
  • LiveKernelReports dumps
  • GPU-Z reports
  • photographs or videos of the Windows corruption
  • any other diagnostic information that may help

You may also like