About VRAMwise

The problem we work on

Running language models locally has become mainstream, but the first step is still guesswork for many people: a model download can be tens of gigabytes, and only after loading it do you find out that it does not fit — or that it fits at 4k context but crashes at 16k. Rule-of-thumb answers scattered across forums often ignore quantization differences and almost always ignore the KV cache, which scales with context length.

VRAMwise approaches this as an estimation problem with a transparent formula: model weights at a given quantization, plus KV cache at a given context length, plus runtime overhead, compared against the memory a card or Mac actually makes available.

What this site is — and is not

It is

A set of calculators and reference tables built on published specifications and openly documented formulas, with worked examples you can verify by hand.

It is not

A benchmark lab with every GPU on a shelf. Where we cite measured throughput figures, they come from clearly attributed community and vendor sources, and estimates are labeled as estimates.

How the numbers are maintained

GPU specifications are taken from vendor spec sheets, model sizes from official model configuration files, and quantized file sizes from their published repositories. New models and GPUs are added as they release. The full sourcing and calculation policy lives on the methodology page.

Found an error, or want a specific GPU or model added? Contact us — corrections are prioritized over new features.