Shared GPU memory: what it is, and why it slows games and AI

Where to see it

Open Task Manager (Ctrl+Shift+Esc), choose Performance and then your GPU. Under the graphs are three figures:

• Dedicated GPU memory: in use, and how much the card has — 8 of 8 GB on an 8 GB card, for example.
• Shared GPU memory: how much system memory the graphics are using, and how much Windows would let them use. That second figure is usually half of the PC’s memory: 16 GB on a PC with 32 GB.
• GPU memory: the two added together.

To see which programs use which, choose Details, right-click a column heading, choose Select columns and tick Dedicated GPU memory and Shared GPU memory.

The shared figure is a limit, not memory taken

The 16 GB of shared GPU memory on a 32 GB PC is not reserved: the rest of Windows can use it all until the graphics actually need some. A few hundred megabytes in use is normal — drivers keep small things there. What matters is whether it climbs into gigabytes while a game or an AI model runs.

Graphics built into the processor are different: they have no memory of their own, so everything they use is shared memory, apart from a small amount the firmware sets aside, which Task Manager shows as their dedicated memory.

Why using it is slow

A graphics card reads its own memory at 200 to 1,800 GB a second, depending on the card. Shared memory has to come across the PCIe slot, which carries about 32 GB a second at PCIe 4.0 ×16 and 64 at PCIe 5.0 — and less on a card with fewer lanes. So anything the card has to fetch from shared memory arrives five to fifty times more slowly.

• In games, a card that runs out of its own memory keeps some textures in shared memory and fetches them as it needs them: sudden stutters, textures that load late, and frame rates that drop in busy scenes. Lowering texture quality is the setting that fixes it.
• In AI, a model has to be read in full for every word it writes, so a model that spills into shared memory writes many times more slowly — often more slowly than if the processor worked on the part that did not fit. With NVIDIA cards this happens when a program asks for more memory than the card has, because NVIDIA’s driver lends it system memory rather than refusing. Ollama using the CPU instead of the GPU covers how to tell, and the NVIDIA Control Panel setting that stops it.

Can you turn it off, or make it bigger?

Neither helps in the way it sounds. Windows offers no switch for shared GPU memory, and registry changes found online that raise the figure do not add memory to the card: they only move the point at which the slow borrowing starts. What does help:

• Free the card’s own memory. Browsers, Discord, launchers and animated wallpapers keep some of it all the time. The Details columns above show how much; close what you do not need, or switch off Use graphics acceleration when available in the browser.
• Ask less of it. Lower texture quality in games; in AI apps, a shorter context length or a smaller compression of the model.
• For NVIDIA and AI programs, set CUDA - Sysmem Fallback Policy to Prefer No Sysmem Fallback in NVIDIA Control Panel, under Manage 3D settings, so the program fits itself to the card instead of overflowing.
• On built-in graphics, the firmware sets how much memory is reserved for them (often called UMA frame buffer size), and AMD Software has Variable Graphics Memory on some Ryzen laptops. Raising it lets a game or a model use more memory as the graphics’ own; it does not make that memory any faster.
• A card with more memory is the real fix when a game or model needs more than it has. Best GPU for local AI compares what each card holds.

Shared GPU memory not being used?

That is normal and good: it means everything fits in the card’s own memory. Some AI apps let you choose whether to use it at all; leave that off unless a model will not load otherwise, and expect it to be slow when it is on.

How Bottleneck checks this

A scan lists the programs holding the graphics card’s memory. Bottleneck’s Check my local AI, on its AI models screen, sees whether a model run with Ollama, LM Studio, llama.cpp, KoboldCpp or Jan has overflowed into shared memory, and how fast it writes compared with what the PC should manage. Both are free; Pro explains the cause and the fix.