RTX 5070 users chase final 450MB of VRAM

This is the kind of small systems problem that matters a lot in local LLM work: a few hundred megabytes can decide whether a model fits cleanly or forces harsher compromises.

• –NVIDIA positions the RTX 5070 family as AI-capable hardware, but Linux desktop sessions can still keep a slice of VRAM tied up in compositor and driver state.
• –Moving display output to an iGPU does not automatically make the discrete GPU fully headless; GNOME, Wayland, and the NVIDIA stack may still retain buffers or contexts.
• –For LLM users, the real fix is often a lighter desktop, a TTY-only or headless session, or a dedicated inference box rather than expecting GNOME to release every last megabyte.
• –The post is a useful reminder that practical usable VRAM is often lower than the advertised total, especially on consumer cards doing double duty for desktop and compute.