edit: Works as expected! Thanks good, I would not be able to pay for a new GPU.

vram_lock – GPU VRAM Corruption Finder / “Lock” Tool

vram_lock is a minimal C++ utility that tries to find VRAM allocations that do not reliably hold data and then keeps (“locks”) the suspected-bad chunk allocated so it can’t be reused.

What it’s for

This tool is useful when a GPU has VRAM errors (often due to aging hardware, unstable overclocks/undervolts, or marginal memory). By finding a slice of VRAM that fails to retain data and then keeping that slice allocated, you can sometimes reduce or eliminate crashes/artifacts by preventing the driver from reusing the problematic pages.

It can also be used to make a marginal VRAM overclock usable by effectively leaving the least-capable part of memory behind: if one region of VRAM is the first to fail at a given memory clock, vram_lock may find that region and keep it allocated so the driver is less likely to place new allocations there.

How it works

It works by repeatedly allocating a configurable “slice” of VRAM on a selected GPU, writing a deterministic byte pattern (0xA5), then copying the same slice back to host memory multiple times and comparing the results.

If the copies ever differ for the same allocation, the program treats that slice as corrupted/unstable, frees all previously successful allocations, keeps the faulty allocation resident (“locks” it), and then sleeps forever so the bad pages cannot be reused.

GPU 0 (NVIDIA GeForce RTX 4090)
Slice size: 64 MiB (67108864 bytes)
Slices held (locked faulty): 1   OK: 358   Faulty locked: 1   In-progress: 0
Slices held (allocations): 1
Map entries: 359
Total held: 64 MiB
Elapsed: 47s
Next slice index: 359
Last status: STOP: cuMemAlloc failed (likely OOM). Freeing all OK slices; keeping only faulty locked.
Last check: mismatch detected between repeated readbacks

VRAM slice map ('#'=allocated OK, 'X'=faulty locked, '?'=in-progress, '.'=freed after OOM)
     0: ..................X.............................................
    64: ................................................................
   128: ................................................................
   192: ................................................................
   256: ................................................................
   320: .......................................

cuMemAlloc failed at slice #359: CUDA_ERROR_OUT_OF_MEMORY (2) - out of memory
Sleeping forever holding only faulty VRAM allocations.

---

Command-line usage

./<binary> [gpu_index] [slice_mebibytes]
Defaults: gpu_index=0 slice_mebibytes=512

Examples:

./vram_lock              # GPU 0, 512 MiB slices
./vram_lock 1            # GPU 1, 512 MiB slices
./vram_lock 0 1024       # GPU 0, 1 GiB slices

---

Requirements

• NVIDIA GPU with a working CUDA driver
• CUDA Driver API available at runtime (libcuda)

Platforms / binaries

This project is intended to run on Linux and Windows (where a CUDA driver is available).

A prebuilt Windows binary is available here: https://github.com/Keeo/vram-lock/blob/master/vram_lock.exe

---

Native build & run

Build however you prefer (Makefile/CMake/manual). A typical manual build looks like:

g++ -O2 -std=c++17 vram_lock.cpp -o vram_lock -lcuda
./vram_lock

Note on cuCtxCreate signature

This project uses the 4-argument form:

cuCtxCreate(&ctx, nullptr, 0, dev);

Some older CUDA headers expose a different signature (e.g. 3 arguments). If you see a compile error around cuCtxCreate, adjust the call to match your installed CUDA headers.

---

Docker build & run

This repo includes a Dockerfile:

docker build -t vram_lock .
docker run --rm --gpus all vram_lock

To test if your GPU is faulty, you can also use gpu_burn:

https://github.com/wilicc/gpu-burn
docker run --rm --gpus all gpu_burn

---

Cleaning up / stopping

Press Ctrl+C to terminate the program.

Note: the program does not currently install a signal handler to explicitly free allocations on SIGINT; on exit, the OS/driver will reclaim resources when the process terminates.

---

License

Distributed under the terms of the MIT License (see LICENSE).