NVIDIA GPU VRAM을 Linux 스왑 공간으로 사용 - 업그레이드 불가능한 노트북을 위한 메모리 확장 솔루션

Use your NVIDIA GPU's VRAM as swap space on Linux.

Built for hybrid graphics laptops with soldered memory and no upgrade path. The display runs off the integrated AMD/ATI GPU. The NVIDIA card sits idle most of the time, its VRAM completely unused. This puts that VRAM to work as high-priority swap.

Tested on: AMD/ATI + RTX 3070 Laptop (GA104M, 16 GB RAM, 8 GB VRAM), driver 580.159.03, kernel 6.17, Pop!_OS. Allocated 7 GB for swap. End result including zram and SSD swap: ~46 GB total addressable memory, tripled from stock. Overflow order: RAM fills, then VRAM absorbs the spill (PCIe), then zram compresses the rest (CPU), then SSD only if everything else is exhausted.

A small daemon allocates VRAM via the CUDA driver API, then serves it as a block device using the NBD (Network Block Device) protocol over a Unix socket. The kernel's built-in nbd driver connects to it and exposes /dev/nbdX. From there it's a normal swap device.

Data path: kernel swap subsystem - /dev/nbdX - nbd kernel driver - Unix socket - nbd-vram daemon - cuMemcpyHtoD/DtoH - GPU VRAM.

No kernel module to write or maintain. No NVIDIA kernel symbols. Survives kernel and driver updates without rebuilding anything.

The "obvious" approach is nvidia_p2p_get_pages_persistent, which pins VRAM pages in BAR1 so the CPU can access them directly via ioremap_wc. Every existing project that tried this route hits the same wall: the NVIDIA driver returns EINVAL on consumer GeForce GPUs. Both the persistent and non-persistent variants, both flag values. It's gated at the RM level for Quadro/datacenter SKUs only, regardless of driver version.

The other approach - directly ioremap_wc the BAR1 physical address without going through the P2P API - also doesn't work. The GPU's internal page tables only have ~16 MiB of BAR1 mapped (just the display framebuffer). Reads from the rest return zeros. mkswap appears to succeed, then swapon fails because the swap header isn't actually there.

The NBD approach sidesteps all of this. cuMemcpyHtoD and cuMemcpyDtoH work on any CUDA GPU without any special permissions.

• NVIDIA GPU with CUDA support (any consumer RTX/GTX card)
• NVIDIA driver with libcuda.so.1 (no CUDA toolkit needed)
• Linux kernel 3.0+ (nbd module, built into most distros)
• nbd-client package
• gcc, make

git clone https://github.com/c0dejedi/nbd-vram
cd nbd-vram
sudo ./install.sh
sudo systemctl start vram-swap-nbd

Verify:

swapon --show
# NAME       TYPE      SIZE USED PRIO
# /dev/nbd0  partition   7G   0B 1500

The service is enabled on install, so it comes up automatically on every boot.

Edit /etc/systemd/system/vram-swap-nbd.service:

Environment=VRAM_SETUP_SIZE_MB=7168    # how much VRAM to use
Environment=VRAM_SWAP_PRIORITY=1500   # swap priority (higher = used first)

The daemon tries the requested size first and backs off in 512 MiB steps if the GPU is short on memory - so it will grab as much as it can even if the display compositor is already loaded. VRAM_SETUP_SIZE_MB is the ceiling, not a hard requirement.

After changing, run sudo systemctl daemon-reload && sudo systemctl restart vram-swap-nbd.

The installer asks whether to enable power-aware management on first install. If enabled, the service automatically stops when you unplug from AC (or when battery drops below a threshold), and restarts when power is restored. Manual systemctl stop is always respected and won't be overridden.

To change settings after install, edit /etc/nbd-vram.conf. Changes take effect on the next poll (within 60 seconds) or immediately on the next AC plug/unplug event.

Allocates VRAM, connects the NBD device, does a 1 MiB write/readback check, activates swap, then prints teardown instructions. install.sh handles teardown automatically if a test instance is running.

To stress the full partition after the smoke test passes:

Writes the entire VRAM partition with zeros, verifies a sample read back, then auto-restores swap on exit.

Tested on RTX 3070 Laptop (8 GB VRAM), kernel 6.17, Pop!_OS. Compared against NVMe cryptswap (dm-crypt, PCIe 4.0). All benchmarks run with O_DIRECT to bypass page cache.

Three benchmarks are in benchmarks/. Each runs NVMe first, then starts the VRAM service and runs the same test against the block device. State is restored on exit.

sudo bash benchmarks/bench-throughput.sh   # sequential read/write (dd, 2 GiB, O_DIRECT)
sudo bash benchmarks/bench-iops.sh         # 4K random IOPS (fio, libaio, iodepth=32)
sudo bash benchmarks/bench-latency.sh      # per-operation latency (ioping, 20 requests)

fio and ioping are installed automatically if missing.

VRAM is slower for large sequential transfers. The bottleneck is the NBD + CUDA userspace round-trip - every block crosses a Unix socket and a cuMemcpy call, which adds overhead that NVMe's direct kernel block path doesn't pay. Sequential throughput is not the primary swap workload (the kernel swaps individual 4K pages, not 4 MiB streams) - see the IOPS and latency benchmarks below.

NVMe wins for sustained random I/O. At iodepth=32, NVMe can have 32 requests genuinely in flight simultaneously; the NBD+CUDA path serialises them through the daemon, so the depth advantage is reduced. The VRAM daemon also adds CPU overhead that the NVMe path does not pay. For continuous high-throughput swap pressure, NVMe is faster.

The picture changes for sporadic access - see the latency benchmark below.

VRAM is 27x faster average latency. The NVMe drive is physically capable of ~112 us (visible on the warmup request) but APST (Autonomous Power State Transitions) puts it to sleep between requests. At 1 request per second - the rate of sporadic swap access - it wakes cold almost every time and pays a ~9 ms penalty. VRAM has no power states and responds in 133-490 us consistently.

This is the scenario that matters most in practice. Memory pressure on a laptop is rarely a sustained GB/s flood - it is individual 4K page faults arriving seconds apart. Every one of those faults stalls waiting for the swap device to respond. At 9 ms per fault, NVMe swap is felt. At 335 us, VRAM swap is not.

MIT - Sean Lobjoit (c0dejedi)