If your PC, laptop or mini PC has no graphics card, you can still run a chatbot locally. Many modern processors have integrated graphics (an iGPU), and local AI tools can use it through Vulkan, a graphics and compute API that's built into most GPU drivers. On some AMD Ryzen AI systems you can also reserve a larger chunk of system RAM as dedicated graphics memory. This guide is based on Ollama's and LM Studio's documentation and AMD's Variable Graphics Memory FAQ. Forge PC hasn't benchmarked integrated graphics for AI, so we don't quote speeds; expect it to be slower than a graphics card, and much slower with big models.
What you need
A processor with integrated graphics and a current graphics driver from Intel or AMD (or your laptop maker).
Enough RAM. An iGPU has no memory of its own: it borrows system RAM. LM Studio's system requirements recommend at least 16GB of RAM on Windows. Our guide to how much RAM you need for local AI goes into more detail.
Dual-channel memory if possible. Integrated graphics uses system RAM, so memory bandwidth matters; two matched sticks run in dual channel, as explained in single vs dual channel RAM.
Ollama or LM Studio. Both are free. Our guide to running AI models locally with Ollama and LM Studio covers installation.
Step 1: Update your graphics driver
Ollama's documentation says that on Windows, "most GPU vendors drivers come bundled with Vulkan support and require no additional setup steps". In practice, that means installing the latest Intel or AMD graphics driver is the main preparation. On Linux, Ollama says most distributions need extra Vulkan components, with separate instructions for Intel and AMD.
Step 2 (AMD Ryzen AI only): Set Variable Graphics Memory
AMD describes Variable Graphics Memory (VGM) as a BIOS-level feature, introduced with Ryzen AI 300 series processors, that lets you reallocate a percentage of system RAM to the integrated graphics as dedicated graphics memory. AMD's FAQ explains the difference from the usual arrangement: on typical Windows systems, 50% of RAM is already shared with the iGPU, and that shows up as shared graphics memory in Task Manager. VGM instead creates one contiguous dedicated block, which AMD says lets AI apps designed around VRAM run without managing shared memory.
To turn it on, AMD says to:
Right-click the desktop and open AMD Software: Adrenalin Edition.
Click the Performance tab, then the Tuning tab.
Choose a Variable Graphics Memory level from the drop-down, then restart.
AMD's recommendations are Default for casual users, Medium for gamers, High for power users and Custom (75%) for AI enthusiasts. It warns that VGM subtracts from the RAM the CPU can use and "may deteriorate system performance if used incorrectly". If you have 16GB of RAM, giving most of it to the iGPU leaves little for Windows, so start conservatively.
On other systems, the amount of memory reserved for the iGPU may be set in the BIOS, and options vary by motherboard. You can check how much dedicated and shared GPU memory Windows sees with our guide to checking VRAM usage.
Step 3: Run a small model in Ollama
Ollama uses Vulkan for additional GPU support on Windows and Linux, and its documentation says Vulkan is enabled by default when the backend is installed. Install Ollama, then in a terminal run a small model from Ollama's library (one with a few billion parameters) using ollama run followed by the model name. While it's answering, open Task Manager's Performance tab and check that the GPU shows activity.
If your iGPU causes problems on a PC that also has a graphics card, Ollama's docs say you can keep Vulkan enabled and set GGML_VK_VISIBLE_DEVICES to the discrete GPU's index, or disable Vulkan with OLLAMA_VULKAN=0.
There's a relevant change on the way: Ollama's v0.40.17 pre-release includes a change to "use mmap for Vulkan integrated GPUs". It's a pre-release, so wait for the stable version unless you like testing.
Step 4 (alternative): Use LM Studio
LM Studio is a desktop app with a model browser and a GPU offload slider. On Windows it needs a CPU with AVX2 support on x64. Note that LM Studio's own requirements recommend at least 4GB of dedicated VRAM for GPU acceleration, which most iGPUs don't have unless you allocate it, for example with AMD's VGM. If performance is poor, try fewer GPU layers; our guide to GPU offload in LM Studio explains the setting.
Which models to expect
Model size is the limit. A quantised model needs roughly its file size in memory, plus extra for context. Start with small models (around 1B to 8B parameters) at 4-bit quantisation; our GGUF quantisation explainer shows what the labels mean. AMD's FAQ says its top-end Ryzen AI Max+ 395 with 128GB of memory and up to 96GB of dedicated graphics memory can run 4-bit models up to 128B parameters, but that is a very different machine from an ordinary laptop or desktop iGPU.
Our verdict
Integrated graphics is a perfectly good way to try local AI: update your graphics driver, install Ollama or LM Studio, and start with a small 4-bit model. On AMD Ryzen AI systems, Variable Graphics Memory is the most useful setting, but don't starve Windows of RAM. If you find yourself wanting bigger models or faster replies, compare CPU vs GPU for local AI and our best GPUs for local LLMs.