NBD-VRAM Bypasses NVIDIA P2P API Limits to Enable Consumer GPU Swap on Linux
nbd-vram routes swap traffic through user-space CUDA copies to evade GeForce P2P restrictions documented in NVIDIA RM-level gating, achieving PCIe-bandwidth VRAM swap on unmodified consumer hardware.
The nbd-vram project allocates up to 7 GB of NVIDIA GPU VRAM as a Network Block Device swap partition on Linux kernels 3.0+, using CUDA memcpy calls over a Unix socket daemon rather than restricted P2P APIs. On an RTX 3070 Laptop with 16 GB system RAM, this configuration yields an effective 46 GB addressable space when combined with zram and SSD swap, with overflow ordered as RAM to VRAM to compressed RAM to disk. Primary source: https://github.com/c0dejedi/nbd-vram. Related coverage in Linux kernel NBD module documentation confirms built-in nbd driver support without custom modules. NVIDIA CUDA Driver API reference documents cuMemcpyHtoD/DtoH availability on all consumer CUDA GPUs since driver 580 series.
[AXIOM]: Consumer GPUs gain practical heterogeneous memory extension via existing CUDA paths, avoiding vendor SKU locks observed in prior P2P attempts.
Sources (3)
- [1]Primary Source(https://github.com/c0dejedi/nbd-vram)
- [2]Related Source(https://docs.nvidia.com/cuda/cuda-driver-api/index.html)
- [3]Related Source(https://www.kernel.org/doc/html/latest/admin-guide/nbd.html)