GPU Poor answers the two questions every local-LLM enthusiast asks first: can my GPU run this model, and how fast will it go? This free web calculator estimates the GPU memory a language model needs and the tokens per second you can expect. The open-source project has about 1,400 stars on GitHub.
Model cards list parameter counts, not whether a 24 GB card can actually hold a 70B model at a given context length once the KV cache and overhead are counted. GPU Poor turns that guesswork into a calculation. Pick a model and your hardware, choose a quantization scheme — GGML, bitsandbytes, or QLoRA-style formats — and the tool breaks down where the memory goes for inference and for training, so you can see the weights, cache, and overhead separately. It also estimates the generation speed you can expect, and accounts for popular inference frameworks such as vLLM, llama.cpp, and Hugging Face. The calculator runs entirely in the browser on the project’s site, requires no signup, and the underlying JavaScript is open source if you want to check or extend the math. The project was last updated in late 2024.
GPU Poor is completely free. The calculator is a static web page hosted on GitHub Pages, and the source code is open; there are no accounts, plans, or paid features.
Anyone planning a local LLM setup — hobbyists sizing a GPU purchase, developers checking whether a model fits their card, and teams estimating serving costs before committing to hardware they may not need.