By Web Asset House · data reviewed 2026-10-22
A useful comparison starts with the exact model and runtime. A unified-memory computer and a discrete graphics card can expose very different portions of their advertised memory to inference. Treat total system RAM and dedicated GPU VRAM as separate resources when the machine uses a discrete GPU.
Unified memory: reserve room for the rest of the machine
Our default budget on unified-memory systems is 75% of installed capacity. That is a conservative planning setting, not an operating-system limit or a maker guarantee. The calculator allows adjustment. A larger shared pool may hold a model that exceeds a smaller discrete GPU’s VRAM, while memory bandwidth still constrains estimated generation speed.
Discrete GPU: capacity before acceleration
A GPU-only estimate applies when all weights, KV cache, and runtime overhead fit its usable VRAM. Offloading part of a model to host RAM changes the data path. The estimator does not silently extend GPU bandwidth to system memory. Host-memory capacity and bandwidth must be supplied before a conditional offload estimate can be shown.
Software is part of the decision
llama.cpp provides multiple hardware backends, including Metal, CUDA, Vulkan, and SYCL. Backend availability is not universal support for every model or feature. Check the runtime build, driver, model format, and exact accelerator. A memory-fit verdict is a screening step, not a compatibility certification.
A practical comparison
- Keep the model, quantization, context, and concurrency identical.
- Compare memory headroom in the calculator.
- Look for a published benchmark with the same model format and workload.
- Check the full computer price, power requirements, and software you already use.
Primary sources
- Ollama: memory, context, local and cloud behavior
- llama.cpp: supported build backends
- Hugging Face: GGUF documentation
Questions, answered.
Does system RAM add directly to GPU VRAM?
No. A discrete GPU and host RAM have different bandwidth and transfer paths. A runtime must explicitly support offload.
Is a memory fit a compatibility guarantee?
No. Model architecture, runtime, operating system, drivers, and backend support must also be checked.