Use your Nvidia GPU's VRAM as swap space on Linux

Original Article ↗ Hacker News Discussion ↗

Using an Nvidia GPU’s VRAM as swap space on Linux is pitched as a way to reclaim otherwise idle video memory on machines with limited, non-upgradable RAM, especially laptops. Commenters see it as a clever, niche hack that can reduce SSD wear, but note that current user-space and NBD-based implementations are far from saturating PCIe or VRAM bandwidth and may complicate power management or gaming workloads. The broader exchange revisits when swap is beneficial at all, how kernel paging strategies work, and why GPU memory isn’t straightforward to treat as regular system RAM.

Motivation and Use Cases

  • Intended for machines with limited, non-upgradable RAM but relatively large, often-idle Nvidia VRAM (e.g., laptops in hybrid GPU mode, desktops with big gaming/AI cards).
  • Lets unused VRAM serve as swap instead of writing to SSD, potentially saving flash wear and exploiting otherwise wasted memory.
  • Seen as especially attractive when workloads (gaming vs. “productivity” / heavy RAM use) are not concurrent, so VRAM can be repurposed when GPU is idle.

Performance and Implementation Concerns

  • Reported throughput (~1.3 GB/s on a laptop RTX 3070) is far below theoretical PCIe/VRAM bandwidth.
  • Thread attributes this to:
    • User-space implementation using NBD, which is known to be relatively slow.
    • Extra copies via a bounce buffer and many kernel/user context switches per 4K page.
    • Limited queue depth and poor request coalescing with NBD.
    • ZRAM compression overhead (though some think this is minor).
  • Swapping to NVMe is described as a highly optimized, zero-copy, DMA-based path that can be faster in practice.
  • Suggestions: move to ublk or a custom kernel block driver to reduce overhead and increase concurrency.

Hardware and Architectural Limits

  • VRAM cannot simply be added to system RAM because most desktop GPUs are not cache-coherent with the CPU and must be mapped uncached or write-combined, making them very slow as “real” RAM.
  • Datacenter GPUs and future CXL-style devices can offer coherency, but latency remains much higher than DRAM.
  • Historical and alternative approaches exist (MTD/phram, vramfs, OpenCL/Vulkan-based ramdisks, Windows GpuRamDrive).

Swap Semantics and SSD Wear

  • Some see VRAM swap as a way to avoid SSD wear; others argue real-world swap usage usually doesn’t meaningfully shorten SSD life.
  • Broader debate on swap:
    • Some configure large swap (often for suspend-to-disk) and rely on it to prevent abrupt OOM.
    • Others prefer minimal or no swap, treating heavy swapping as equivalent to a “dead” system.
    • One view: swap’s main role is fair reclamation between anonymous and file-backed memory, not emergency RAM.

Usability, Risks, and Power

  • Concerns that using VRAM as swap may:
    • Prevent GPU power-gating, hurting battery life on laptops.
    • Compete with graphical or compute workloads and cause crashes if VRAM runs out (especially on dynamic Wayland compositors).
  • Dynamic enabling/disabling of VRAM swap on GPU pressure is desired but not fully solved; flagged as a to-do area.
  • Classic microkernel-style concern (swap daemon needing to page itself) is addressed by keeping its pages unswappable/pinned.