About VRAM Calculator

Who runs this

VRAM Calculator is built and maintained by Bedino, a small independent web project that makes fast, single-purpose calculators. It is one of a family of tools, each on its own domain, each doing one thing. Contact:[email protected].

What this is

A set of calculators for sizing large-language-model inference against real hardware: how much GPU memory a model needs, whether it fits your card, how fast it will run, how much the context window costs, and how quantization changes all of that. Training, multi-GPU splits, serving concurrency and cost sizing live on the sibling sitefinetune-calc.com. No account, no sign-up, no paywall.

How it works

Every calculation runs entirely in your browser. The model, GPU and quantization presets are bundled into the page as plain data — nothing you enter is sent anywhere or stored. The site is deliberately light so pages load fast.

Accuracy

The formulas are standard — parameter counts times bytes per weight, the memory-bandwidth roofline for decode speed, and a KV-cache model that branches on the attention type: the standard equation for GQA/MHA, the latent-cache formula for MLA (DeepSeek, Kimi, GLM), a windowed cache for sliding-window models (Gemma 3, gpt-oss), and attention-layers-only for hybrid attention/SSM (Qwen3-Next, Nemotron-H). What ages is the preset data, so it is regenerated from Hugging Face model configs every few days. Each tool shows the date its presets were last verified, and every field also accepts fully custom values, so a missing or stale preset never blocks you. The estimates carry a deliberate safety margin; treat them as "fits in practice", not exact minimums. Spotted something wrong or out of date? Tell us at[email protected].

How the site is funded

VRAM Calculator is free and paid for by advertising. Ad and analytics scripts load only after you accept via the consent banner — details in theprivacy policy.