we have a laptop with Ubuntu 10.4 and the CUDA 5.0 driver+toolkit, equipped with a Quadro 1800M and 1GB of video memory. A few days ago, one of our apps stopped working and started throwing out of memory errors. So we investigated a little and inserted some cuMemGetInfo() calls into the code.
Turns out that right after cuda context creation, we only get 372 MB free memory out of 1024 MB of available device memory. And this is pretty much after a clean boot.
Turning off desktop compositing made a small difference (maybe 60MB more memory), also removing the desktop background image resulted in a small gain. But we’d really like to have some 900MB available, not just 372MB.
Does anyone know why a Linux system would suddenly report such little available memory for CUDA? Are there any apps for Linux (Ubuntu) that would show video RAM usage, like GPU-z does on Windows systems?