My PC randomly reboots or produces a VIDEO_TDR_FAILURE (0x116) blue screen in certain games, especially Unreal Engine 5 or DX12 titles. The crash points to nvlddmkm.sys, with WinDbg showing dxgkrnl timeout and GPU reset activity. My system has a Ryzen 7 5700G, MSI RTX 3060 Ventus 2X 12G OC, 16GB DDR4 running at 2400 MHz, an MSI B550 Gaming Plus motherboard, and a Corsair CV650 650W Bronze power supply running Windows 11 Pro. GPU temperatures stay around 70–74°C. Fortnite is unstable in DX12 but becomes reliable in DX11 or Performance Mode, while other games run normally. I have already tried DDU, multiple NVIDIA Studio driver versions, DISM, SFC, Windows Memory Diagnostic, and disabling OBS. Does this point more toward a driver or software problem, a power-delivery issue, or a failing graphics card? What tests should I run before considering an RMA?
3 Answers
A 650W supply is sufficient on paper for an RTX 3060, but the wattage recommendation does not prove the PSU is healthy. Check Windows Event Viewer for Kernel-Power events, reseat the GPU and power cable, and test with another known-good PSU if available. OCCT or another GPU/VRAM stress test can help, but run separate GPU, VRAM, and power tests while watching for artifacts, driver resets, or reboots. Testing the card in another computer, or testing a different GPU in this system, is the most useful way to separate the graphics card from the rest of the machine.
The clean driver reinstall and normal temperatures are encouraging, but they do not conclusively clear the GPU. Check the actual minidumps rather than relying only on the failing module name, because nvlddmkm.sys is often where a broader GPU or hardware timeout is reported. Also test with stock BIOS settings, temporarily run the RAM at default JEDEC settings, and verify that the motherboard is not applying an unstable memory or CPU profile. If only one or two games fail, game-specific DX12 bugs or shader issues remain plausible; if multiple independent stress tests fail or the card fails in another PC, an RMA becomes much more justified.
The fact that it only happens in particular games does not completely rule out hardware. Different games use different DX12 features, shader compilation paths, VRAM patterns, and transient power loads, so a marginal GPU, memory subsystem, driver interaction, or PSU can appear stable elsewhere. Before an RMA, try lowering the GPU power limit or core clock slightly and disable features such as hardware-accelerated GPU scheduling and overlays. If a modest underclock makes the affected games stable, that is useful evidence of marginal hardware or power delivery, even though it is not a permanent fix.
The GPU is stable in several other demanding games, including FF7 Remake, Diablo IV, Baldur’s Gate 3, Alan Wake, Alien: Rogue Incursion, Days Gone, and Wuchang, even while recording with OBS. Fortnite also works when streaming through OBS. The failures seem concentrated in certain games and workloads rather than occurring under every kind of GPU load. Does that make a hardware fault less likely?