Running out of video memory (VRAM) is one of the most common reasons a game stutters or a local AI model suddenly slows to a crawl. The first step is knowing what is actually using it. Windows can tell you more than most people realise, and a new free tool goes further, down to which AI model is sitting in your graphics card's memory.
Forge PC has not benchmarked these tools. The steps below follow Microsoft's DirectX team blog on how Task Manager reports GPU memory, Ollama's documentation and the GPU Taskman project page on GitHub, with background from Tom's Hardware and heise online.
Method 1: Task Manager (built in)
Task Manager is the quickest check and needs nothing installed.
Press Ctrl + Shift + Esc to open Task Manager.
Open the Performance tab and select your GPU. Microsoft's DirectX blog explains that this view shows data aggregated across all processes for your graphics card.
Look at the memory figures under the graphs. Microsoft splits video memory into dedicated and shared.
What those terms mean matters:
Dedicated GPU memory is, on a graphics card, your VRAM: the memory physically on the card. Microsoft says the figure shown is the memory currently consumed across all processes.
Shared GPU memory is normal system RAM that the GPU or CPU can use. Microsoft notes that Windows only allows the GPU to use up to half of physical memory at any instant, so on a 16GB PC the GPU can use up to 8GB of system RAM.
If dedicated memory is close to full and shared memory is climbing during a game or AI job, data is spilling into much slower system RAM. That is usually when performance falls off a cliff.
See VRAM use per app
The Performance tab shows the total. To see which app is using what:
Open the Details tab.
Right-click any column header and choose Select columns.
Tick the GPU and video memory columns you want, such as Dedicated GPU memory.
Microsoft's blog says these per-process counters include both memory used only by that process and memory shared between processes, so the numbers can add up to more than the card's total.
Method 2: Ollama's own check (for local AI)
If you run language models with Ollama, it can tell you directly whether a model fits in VRAM. Run this in a terminal while a model is loaded:
ollama ps
Ollama's FAQ explains the PROCESSOR column: 100% GPU means the model is loaded entirely onto the graphics card, while a split such as 48%/52% CPU/GPU means part of it is in system memory. A split is the clearest sign that the model, or its context window, is too big for your card. Choosing a smaller, more compressed version of the model is often the quickest fix; our quantisation explainer explains how.
Method 3: GPU Taskman (free, open source)
GPU Taskman is a new terminal tool by Scott Hanselman, a vice president at Microsoft, released under the MIT licence. Its GitHub page describes it as "an htop for what's inside your GPU", and it tries to answer the question Task Manager cannot: when Task Manager says "15 GB used", used by what?
According to the project page and heise online's report (12 November 2026), it shows:
per-process GPU memory and engine activity
GPU-accessible system RAM alongside VRAM
detected runtimes such as CUDA, HIP, Vulkan, Direct3D and ONNX Runtime
for GGUF models used by tools such as llama.cpp and Ollama, the model name, parameter count and quantisation, plus estimated memory for the weights and the KV cache
Install it
The GitHub page gives two install options if you have Python 3.11 or later:
uv tool install gputaskman
or
pipx install gputaskman
Then run gputaskman (or the shorter gtm). If you do not have Python, the project's Releases page offers portable Windows ZIPs for x64 and ARM64. heise online notes that you must keep the whole package together, including the _internal folder, when you extract it.
Read it with care
Hanselman describes GPU Taskman as early, experimental software. The GitHub page says device and process counters are measurements, but classifications and GGUF-derived breakdowns are estimates, and it does not dump VRAM or list a game's textures. heise online adds that unavailable counters are shown as unknown rather than zero. Tom's Hardware reports that Hanselman's own screenshots showed the Windows desktop window manager as the biggest VRAM user on his system, a useful reminder that background apps take a share too.
What to do with the numbers
Games: if dedicated memory is maxed out, lower texture quality first. Our VRAM guide explains how much memory current games want.
Local AI: if ollama ps shows a CPU/GPU split, try a smaller model or quantisation; our quantisation explainer covers the trade-off. Close browsers and other GPU-hungry apps before loading a big model.
Ongoing monitoring: for temperatures and frame rates while gaming, see our HWiNFO and Afterburner guide.
Our verdict
For a quick answer, Task Manager's Performance tab shows how much dedicated VRAM and shared system memory your graphics card is using, and the Details tab can break that down by app. Ollama users should run ollama ps: anything other than 100% GPU means the model does not fully fit. For the deepest view, GPU Taskman is a free, open-source tool that links VRAM use to specific apps and even AI models, though its model breakdowns are estimates and it is still experimental. Forge PC has not benchmarked these tools; the steps follow Microsoft, Ollama and the GPU Taskman project's documentation.