Running out of VRAM: what happens and how to check

The settings menu draws a VRAM bar and yours is in the red. Or an overlay shows the card's memory almost full and you wonder whether that is why the game hitches. Or textures stay blurry for a second every time you turn around. All three get blamed on VRAM, and only some of them mean you are actually out of it.

What lives in VRAM

Video memory is the graphics card's own memory. In Microsoft's description of Task Manager's GPU figures, dedicated memory is "exclusively reserved for use by the GPU", and on a discrete card that is your VRAM. A game keeps there what it draws with: textures, shaders and geometry, which NVIDIA describes as constantly being transferred to the card as you move through a world, plus the render targets and depth buffers each frame is drawn into.

How much of it a game can use is not fixed. Microsoft's Direct3D documentation says the video memory budget can "fluctuate noticeably as background processes wake-up and sleep", and fluctuates dramatically when you switch to another application. Other programs that use the GPU hold memory on the same card.

What happens when a game runs out

A discrete card has two pools, its own memory and system memory, and Microsoft is blunt about the second: a game can overflow textures into system memory, but "the wrong resource in system-memory can severely impact frame rate". A program that does not stay within its budget can be "intermittently frozen" so other programs can run, and as an extreme last resort Windows itself can shift some of its data from video memory to system memory.

What you see depends on the game, because each one handles the shortage its own way. NVIDIA's VRAM explainer lists three outcomes: memory is purged and the scene reloads, with a noticeable hitch; only selected data is swapped, with no visible effect; or new assets load more slowly because they now come from system RAM. TechSpot's VRAM testing found that spread in practice. Halo Infinite removed textures and lowered detail; Forspoken and Hogwarts Legacy showed texture pop-in and high-quality textures that appeared and disappeared; in Avatar: Frontiers of Pandora the main symptom was less consistent frame times. Where the shortage hits performance, the cost can be large: in Horizon Forbidden West at 1440p with the very high preset, TechSpot measured the 16 GB RTX 4060 Ti 40% faster on average than the 8 GB version.

So the signs are hitches when you turn or enter a new area, textures that stay blurry or swap quality in front of you, and 1% lows that fall far below the average.

How to check VRAM use

• Task Manager. Open Performance and select your GPU. Dedicated GPU memory is how much of the card's own memory is in use; Shared GPU memory is ordinary system RAM the GPU is using. For a per-program view, open Details, right-click a column header, choose Select columns and add Dedicated GPU memory. Microsoft notes that the per-program figures can add up to more than the GPU uses, because memory shared between programs is counted more than once.
• An overlay. Xbox Game Bar's Performance widget shows VRAM, and AMD Software's metrics include GPU memory utilisation in MB; how to show an FPS counter covers where both live.
• The game's own estimate, if its menu has one, with the caveat in the next section.

Allocated is not the same as needed

Task Manager and the overlays show memory that is allocated, and that is not the same as what the game needs. NVIDIA's explanation is that a game allocates VRAM saying "I want it in case I need it", and that holding it "doesn't mean it actually needs all of it". It adds that only the game's developer can know precisely what is actively used, and that the estimate some options menus show "isn't always accurate". TechSpot notes the same from the other side: games tend to allocate more memory when more is available.

That makes a full bar weak evidence on its own. A card that sits near its limit with none of the symptoms above is not a problem to solve. The better test is behaviour: if the symptoms appear at one texture or ray tracing setting and go away one step lower, with nothing else changed, memory is the likely limit.

Which settings cost VRAM

The only figures worth repeating are measurements, and those are per game. TechSpot's 2024 testing recorded use rather than allocation, on a 24 GB RTX 4090 with at least 10 minutes of play per setting:

Three readings of those numbers, which describe these games rather than every game:

• Ray tracing adds memory. TechSpot's conclusion is that it "does increase VRAM usage substantially".
• Frame generation depends on the game. It raised both games above their ray-traced figures, while in The Last of Us Part 1 TechSpot found FSR frame generation "doesn't seem to increase VRAM usage significantly".
• Resolution added less in Ratchet & Clank than the step from medium to very high (8 to 10 GB) did: about 7% more at 1440p than at 1080p, and a further 8 to 10% at 4K.

For a card that is short, TechSpot's advice is to compromise on textures: "you can dial down things like textures and still receive a perfectly playable experience". They measured presets rather than texture quality on its own, so treat textures as a sensible first thing to try, not a measured ranking.

What to do

1. Confirm it. Watch Dedicated GPU memory in Task Manager while the problem happens, and note whether it sits at the card's total.
2. Change one memory-heavy setting at a time (texture quality, ray tracing, frame generation) and replay the same scene. If the hitches or texture problems go away, you have your answer; if nothing changes, look elsewhere.
3. Close other programs that hold GPU memory. The Dedicated GPU memory column in Details shows which ones.
4. Measure before and after on the same run rather than by feel; reading the benchmark covers how.

Integrated graphics and shared memory

Integrated graphics has no VRAM of its own. Intel's graphics memory FAQ says its integrated GPU "doesn't use a separate memory bank" and uses system memory instead. What follows from Intel's and Microsoft's descriptions:

• The small dedicated figure is not the limit. Depending on the GPU, it is memory the firmware or driver reserves from system RAM, as Microsoft describes, or, on Intel's driver, a 128 MB figure that Intel itself calls "fictitious" and reports for compatibility with programs that do not understand a unified memory architecture.
• Shared GPU memory is not taken from Windows in advance. Intel calls it "not an ongoing reservation", only the most the operating system lets graphics use at one time, and both Intel and Microsoft put that limit at half of system memory.
• More RAM raises the limit. Intel says so for processor graphics on 5th-generation Intel Core processors and later, and adds that the limit also depends on available memory, BIOS and system settings.

What does not help

• Treating allocation as need. A full bar is not a diagnosis on its own.
• Counting shared GPU memory as extra VRAM. On a discrete card it is ordinary system RAM, and Microsoft's warning about resources overflowing there applies.
• Resizable BAR as a fix for a small card. It changes how much of the card's memory the processor can reach at once, not how much memory the card has. Resizable BAR explained covers what it does do.

Related

• Why games stutter on Windows 11
• Sudden FPS drops in games
• How to show an FPS counter in any game
• Resizable BAR and Smart Access Memory explained
• DLSS, FSR and XeSS explained