Use your Nvidia GPU's VRAM as swap space on Linux
The open-source project nbd-vram, published on GitHub by developer c0dejedi, introduces a mechanism to treat Nvidia GPU VRAM as swap space on Linux. It leverages the Linux kernel's Network Block Device (NBD) driver to create a block device backed by GPU memory, which the system can then use for swapping. This approach differs from traditional swap on disk or SSD by utilizing the high-bandwidth VRAM typically found on Nvidia GPUs. The tool is designed for systems where GPU VRAM is underutilized while system RAM is scarce, such as in multi-GPU workstations or AI development rigs. It requires the Nvidia proprietary driver and the nbd kernel module. Users can specify the amount of VRAM to allocate as swap. A concrete consequence is that developers running large language models or data processing pipelines on machines with limited system RAM but ample GPU memory may experience reduced out-of-memory errors and improved stability. However, performance will depend on PCIe bandwidth and GPU memory latency, and the tool is intended for experimental or niche use rather than production environments. The project is in early stages and may carry risks such as system instability if misconfigured.
Enables using spare GPU VRAM as swap, potentially reducing OOM errors for memory-intensive dev workloads.