NBD-VRAM launched on August 21 as an open-source MIT-licensed project that lets Linux users allocate part of NVIDIA GPU VRAM as conventional swap space [1]. It uses the CUDA driver and Network Block Device protocol on Unix sockets to expose VRAM as a block device "/dev/nbdX" that can be used for swap [1, 2].
The tool targets laptops with soldered memory and NVIDIA GeForce RTX GPUs that cannot upgrade RAM, allowing VRAM swap to reduce swapping to slower SSD storage [1, 2]. Unlike past methods relying on NVIDIA P2P APIs limited to datacenter GPUs, NBD-VRAM works with any consumer NVIDIA GPU supporting CUDA without requiring special kernel modules or driver hacks [1, 2].
Installation involves running a daemon that exposes VRAM as a block device. The daemon starts automatically on boot, making swap setup seamless [2]. A tested configuration used 7GB of VRAM swap on an RTX 3070 Laptop with NVIDIA driver 580.159.03, Linux kernel 6.17, and Pop!_OS. This setup combined VRAM swap with traditional RAM, zram, and SSD swap to increase effective memory about threefold [2].
NBD-VRAM's lightweight approach enables improved memory performance on laptops where hardware upgrades are not feasible. The project is available now under an open-source license, allowing users to implement VRAM swap on supported Linux systems [1].
