https://bugs.kde.org/show_bug.cgi?id=499560

--- Comment #17 from Kai Krakow <[email protected]> ---
(In reply to David Redondo from comment #16)
> In the end the image needs to be displayed on your screen so it has to end
> up in VRAM.

Not necessarily. What ultimately matters is that a scanout-capable buffer ends
up on the display. That does not imply that all source images or intermediate
surfaces must permanently reside in GPU VRAM.

For example, when Plasma is forced into software rendering, plasmashell itself
does not meaningfully allocate VRAM, yet the compositor still successfully
presents the final image on screen.

So while keeping resources in VRAM is often beneficial for performance, it is
not a strict functional requirement for every surface in the pipeline.

In addition, on current NVIDIA Linux drivers, VRAM eviction and overcommit
behavior is quite limited. This means aggressively allocating render resources
in VRAM can lead to long-lived pressure without effective reclamation.

Given that, seeing ~1.7 GB attributed to plasmashell alone appears excessive in
practice. Even accounting for high-DPI and multiple surfaces, one would
normally expect significantly lower steady-state usage.

-- 
You are receiving this mail because:
You are watching all bug changes.

Reply via email to