What happened
c0dejedi released nbd-vram, a Linux utility converting NVIDIA GPU VRAM into system swap space. This daemon allocates VRAM via the CUDA driver API, exposing it as a Network Block Device (NBD) over a Unix socket. This approach bypasses NVIDIA's P2P API restrictions on consumer GeForce GPUs, which typically block direct CPU access to VRAM. Tested on an RTX 3070 Laptop, it achieved approximately 1.3 GB/s sequential throughput, tripling addressable memory by prioritising VRAM over slower SSD swap.
Why it matters
This utility extends usable memory for platform engineers and developers on Linux systems with NVIDIA GPUs, particularly those with soldered RAM. It provides a faster memory overflow mechanism, routing memory pressure to underutilised GPU VRAM at around 1.3 GB/s, outperforming NVMe SSDs. This mechanism addresses the constraint of fixed system memory, offering a software-defined solution to leverage existing hardware resources for improved system responsiveness during memory-intensive operations.
