"Out of video memory" error on startup — Intel UHD 620 (integrated graphics), CARLA 0.9.15 on Windows 10

[ardaergun4-commits]

Hello,
I'm trying to run CARLA 0.9.15 (pre-built Windows package) for a university project, but the simulator exits on startup with the following error:

Out of video memory trying to allocate a rendering resource. Make sure your video card has the minimum required memory, try lowering the resolution and/or closing other applications that are running. Exiting...

System specs:

OS: Windows 10 (build 19045)
GPU: Intel(R) UHD Graphics 620 (integrated) — 128 MB dedicated display memory, ~8 GB shared
Driver version: 27.20.100.9749
CARLA: 0.9.15 pre-built release, Python client 3.10.11

What I've tried:

Launching with reduced settings: CarlaUE4.exe -quality-level=Low -windowed -ResX=640 -ResY=480 (also tried 400×300 and -dx11)
Closing other GPU-consuming applications

The error occurs regardless of launch flags. I understand the documented requirement is a dedicated GPU with 6+ GB VRAM, so my questions are:

Is there any supported way to run the CARLA server on integrated graphics (e.g., a null-renderer or fully headless mode that skips GPU resource allocation), even with camera sensors disabled? My use case only needs the ray-cast lidar and vehicle physics.
If not, what is the recommended minimum GPU / cheapest practical setup (cloud instance type, etc.) for small-scale sensor simulations like this?

Thanks in advance for any guidance.

[darekwojciechowski]

No, there's no GPU-free mode. The server is an Unreal Engine app and it builds a render device before it ever reads your settings. -RenderOffScreen only skips the window, it still needs a real GPU. That's why the resolution flags did nothing — what fails isn't the framebuffer, it's the engine's baseline allocation, and your card reports 128 MB.

Careful with the one that looks like an escape hatch: no_rendering_mode is set by the Python client after the server is already up, so it can't get you past startup. It also kills ray-cast lidar, since that runs on the GPU too.

So "lidar + physics only" doesn't help. You need a dedicated NVIDIA GPU with 6-8 GB. For a university project a cloud box is usually cheaper than buying — an AWS g4dn.xlarge (T4) is the standard pick:

./CarlaUE4.sh -RenderOffScreen -quality-level=Low -nosound

Rule of thumb before you install anything: if nvidia-smi shows 6 GB or more, you're fine.