GPU Driver Crash Fix: TDR, WHEA, Code 43

Your screen goes black mid-match, the fans spin up, and three seconds later the desktop pops back with a useless little bubble. Annoying. The fix depends on what actually crashed: the driver (TDR), the hardware (WHEA), or Windows refusing to talk to the card at all (Code 43). Start with a clean driver install, because nine times out of ten that's the problem.

Step 1: Kill the old driver properly

Windows' "Uninstall" button in Device Manager leaves residue. Registry keys, INF packages, and leftover DLLs in System32\DriverStore survive and poison the next install. Use DDU (Display Driver Uninstaller) instead.

Grab the latest DDU from Wagnardsoft. Version 18.0.9.1 or newer.
Boot into Safe Mode. Yes, you do need this. Windows won't fully release GPU resources otherwise.
Run DDU, select your vendor (NVIDIA, AMD, or Intel), click Clean and restart.
After reboot, install the driver fresh from the vendor — not through GeForce Experience's "Express" mode, and not through Windows Update.

If you're on NVIDIA, prefer the Studio Driver over Game Ready if you've been crashing. Same silicon, tested against workstation workloads, and you skip a lot of half-baked game-day profiles. The frame hit is somewhere between zero and two percent.

Step 2: Check the power and PCIe side

Half the "driver crashed" reports I see are the GPU browning out. A 3080 spiking to 380 W on a 650 W unit with a split 8-pin daisy chain will drop the rail for a few milliseconds, the driver times out, and Windows logs nvlddmkm or atikmdag as the bad guy. It isn't.

Run separate PCIe cables from the PSU. Not the pigtail. Ever.
Reseat the card. Twist gently when you press it into the slot — half-seated cards cause intermittent Code 43 like clockwork.
If your board has PCIe 4.0 and your riser or slot is marginal, drop to PCIe 3.0 in BIOS. It's a cleaner signal, and you lose nothing meaningful in games.

Step 3: The TDR registry value (use sparingly)

TDR — Timeout Detection and Recovery — is Windows giving the driver two seconds to respond before resetting it. That reset is the black screen. Raising the timeout doesn't fix a broken driver, but it does stop the reset if the driver is just slow under load.

reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers" /v TdrDelay /t REG_DWORD /d 10 /f
reg add "HKLM\SYSTEM\CurrentControlSet\Control\GraphicsDrivers" /v TdrDdiDelay /t REG_DWORD /d 10 /f

Reboot after. If crashes vanish, you've papered over a real problem — probably VRAM instability or an unstable core overclock. Undo it once you've found the root cause, or you're just hiding it.

Why any of this works

The reason step 1 matters is that a driver install is not one file. It's a kernel-mode component (nvlddmkm.sys, amdkmdag.sys, igdkmd64.sys), a user-mode DXGI layer, a shader compiler, and a pile of registry entries that tell Windows which one loads. When you install over an old version, Windows keeps the old kernel component registered if anything in the load path errors out. Then you reboot, Windows loads the stale kernel driver against a new user-mode layer, and you get a version mismatch that manifests as random TDRs. DDU removes the registry link entirely. That's why "reinstall the driver" fails but "DDU then install" fixes it — the missing step was deleting the pointer, not the binary.

Step 2 works because TDR is a timing budget. Your GPU has to service the request within 2 seconds. A power dip stalls the card mid-render, the request never returns, and the timeout fires. The log blames the driver because the driver is what Windows was waiting on. The card is what actually stopped talking.

Less common variations

Code 43 on a card that was working yesterday

Error 43 means Windows can't start the device. If it appeared out of nowhere, check in this order:

Windows Update pushed a driver. Roll back via Device Manager → Properties → Driver → Roll Back Driver. If greyed out, DDU and reinstall vendor driver.
Resizable BAR + old BIOS. Some B450 and X470 boards with ReBAR enabled throw Code 43 on RX 6000 cards. Update BIOS or disable ReBAR.
eGPU or passthrough. Code 43 in a VM usually means the hypervisor isn't hiding the fact that you're in a VM. Set hypervisor.cpuid.v0=FALSE in ESXi, or enable the kvm hidden state flag.
Dying card. Code 43 that survives a clean install in a different slot, on a different PSU, is the card telling you it's done. Bake it in the oven at your own risk, but the VRAM is usually the thing that failed.

WHEA errors instead of TDR

WHEA (Windows Hardware Error Architecture) means the CPU reported a hardware fault — often a PCIe bus error at the root port level. Event Viewer → Windows Logs → System shows these as Event ID 17 (corrected) or 18/19 (uncorrected). Corrected WHEA errors are usually PCIe signal integrity. Uncorrected WHEA means the link dropped. Fixes:

Set PCIe to Gen 3 in BIOS.
Check for a bent pin in the CPU socket. Genuinely. A single bent pin on an LGA1700 board will generate WHEA storms under PCIe load and nothing else.
Update chipset drivers. Old AMD chipset drivers on X570 were notorious for this.

Video Scheduler Internal Error (BSOD)

Stop code 0x00000119. Almost always a corrupted dxgkrnl.sys or a third-party overlay — Discord, MSI Afterburner, RivaTuner, NZXT CAM. Disable them one at a time. RivaTuner's OSD hook is the usual suspect on Windows 11 23H2 and later.

Driver crashes only in one game

Then it's the game, not the driver. Common culprits: DirectX 12 shader compilation on first launch (let it finish — don't alt-tab), anti-cheat kernel hooks, and engines that pre-allocate VRAM over the card's limit. Cap the framerate, drop texture quality one notch, and see if it stops. If yes, the game is over-committing VRAM and the driver is falling over — not a bug you can patch on your end.

Prevention

Don't install drivers through Windows Update. Set the device installation setting to "No" for automatic driver downloads in System Properties → Hardware → Device Installation Settings. Microsoft's certified driver is often six months behind and lacks game-day fixes.
Don't use Express install. Custom → Clean Install in the NVIDIA installer. Skip GeForce Experience entirely unless you actually use ShadowPlay.
Keep 5 GB free on C:. Driver installs extract there. I've seen installs silently fail with 800 MB free.
Undervolt, don't overclock. A stable undervolt is more stable than a stable overclock, and modern GPUs are power-limited long before they're clock-limited. Card runs cooler, boosts higher, crashes less.
Check your PSU age. Electrolytics degrade. A 7-year-old 750 W isn't a 750 W anymore. If you've had mystery crashes across multiple GPUs on the same PSU, that's your answer.

If you've done all of this and the card still TDRs on a clean Windows install with one monitor and no overclock, the GPU is failing. RMA it. No registry tweak fixes a memory chip that's decided to stop being a memory chip.