Interactive · Failure Walkthrough
The Idle/Wake Race — Step by Step
Walk the exact five-step sequence behind an nvlddmkm.sys crash that only strikes at idle — then flip CUDA to OpenCL and watch the same steps end clean. It’s the step-by-step companion to the GPU Power States demo and the real-world crash write-up.
5-step sequence
CUDA vs OpenCL
Use-after-free
BSOD finale
Interactive · Failure Walkthrough
The Idle/Wake Race — Step by Step
This is the actual sequence behind an nvlddmkm.sys crash that only strikes at idle. Step through it, then flip the processing mode from CUDA to OpenCL and watch the same sequence end without a crash.
OS / Kernel
GPU Driver
nvlddmkm
nvlddmkm
App + VRAM
(Resolve)
(Resolve)
cached VRAM held
1 / 5Idle
The fix maps to three Resolve settings. CUDA → OpenCL changes how the app holds GPU memory, so the wake doesn't trip over a stale reference. Manual GPU selection stops the app from grabbing onto phantom display adapters. And fully quitting the app means nothing is pinned across an idle/wake cycle at all — no held memory, no race. Three switches, the same root cause closed three ways.
Plain-English glossary
- CUDA
- NVIDIA's system for letting apps use the GPU for heavy computation. Powerful, but fussy about how memory is held.
- OpenCL
- A vendor-neutral alternative to CUDA that manages GPU memory differently — and sidesteps this particular crash.
- Race condition
- Two things happening at almost the same instant that step on each other, with the outcome depending on which wins.
- Use-after-free
- Code reading memory that was just released or rebuilt — so what it finds is garbage, and it crashes.
- nvlddmkm.sys
- NVIDIA's kernel driver — the part of Windows that talks directly to your graphics card.
- dwm.exe
- The Desktop Window Manager — it draws your Windows desktop, and gets taken down when the GPU driver crashes.
Hitting this exact crash on your own PC?
Grid City does in-home & remote crash diagnostics across NYC and Long Island.
Get in touch