How to choose GPU memory for a local AI PC

Start with the workload, not the GPU name

Choose the model family, quantization, intended context length and number of simultaneous requests first. GPU support varies by runtime and operating system. Check the current support list for the tool you plan to use, rather than assuming any card with enough VRAM will behave the same way.

Weights are only part of memory use

Model weights, the context or KV cache, temporary buffers and the runtime all compete for memory. A model file smaller than the advertised VRAM capacity can still exceed available memory during use. Longer context and concurrent requests can change the result. Measure the actual workload before making a performance promise.

Compare 12GB, 16GB and 24GB as distinct choices

These are useful shopping filters, not guaranteed model-size tiers. More VRAM can make a workload possible on one device, but architecture, memory bandwidth, compute and software support affect speed. A used 24GB card may offer more capacity than a newer 12GB card, while costing more in power, cooling or ownership risk.

System RAM is a separate decision

System RAM supports the operating system, model loading and other applications. Some runtimes can offload work between CPU and GPU, but system RAM is not an interchangeable extension of VRAM and can change performance substantially. Our 32GB and 64GB build targets are editorial starting points, not runtime minimums.

Do not add GPU capacities together blindly

Multiple GPUs need physical clearance, suitable motherboard lanes and slots, power capacity and a runtime that supports distributing your chosen workload. Two cards do not automatically present one combined memory pool. Start with a single GPU unless you have verified a reason and a supported configuration.

Sources & further reading

• Ollama GPU support
• Ollama model scheduling and memory

Reviewed October 22, 2026. Hardware and runtime support can change; confirm the exact current configuration.