About Bitpute

Bitpute exists because most GPU calculators give a different answer for the same model — and none of them show the working. One tool says a model fits, another says it doesn’t, and there is no way to check which one is right.

This site takes the opposite approach: every number on every page is produced by one small calculation engine whose formulas, constants and sources are laid out in full on the methodology page. If you disagree with a figure, you can trace it back to the arithmetic and the source datasheet that produced it.

What this site is

Bitpute is a small collection of calculators for people planning GPU hardware for AI work. It answers questions like: will this model fit on this card, at this quantization, with this much context? What will it cost to rent instead of buy? How much extra memory does fine-tuning need over plain inference?

The rule we hold ourselves to is simple: no number without a source. Weight sizes come from published bits-per-weight figures in the llama.cpp documentation. Architecture details (layer counts, KV heads, head dimensions) are read from each model's public config file, not guessed from the parameter count. The one thing that genuinely cannot be derived from first principles, runtime overhead, is labeled as an estimate right in the results, with the assumption printed next to it. The full working for every formula lives on the LearnEngineering Center page.

What this site is not

It is not a benchmark database, a shop, or a review blog. We don't sell GPUs, we don't take money from anyone who does, and no manufacturer has any say in what the calculators report. If a card looks bad for a given model, the calculator says so.

How we keep the numbers honest

• Every formula is written out on the Engineering Center page before it goes into a tool, so anyone can check the arithmetic by hand.
• Memory estimates are checked against the reference calculation engine on every build; the parity test fails the build if any page drifts from the engine. Corrections are welcome and are handled through the editorial policy.
• Tools ship one at a time. A calculator stays off this site until its outputs have been verified, which is why the roadmap on the homepage lists several tools as "soon" instead of pretending they are done.
• Each page carries a "formulas last reviewed" date in the footer. If model data or quantization figures change upstream, we update and re-check.

Mistakes

We make them. When a wrong number gets caught, by us or by a reader, the fix goes out and the change is noted in the changelog. If you spot something off, please tell us. A correction from a stranger is worth more to this site than a compliment.

Want to understand the calculation engine? Read how Bitpute works — every formula, constant and source written out — or why Bitpute exists.

Who makes this

Bitpute is a one-person project. Its author built the calculation engine behind the site and wrote the methodology. There is no content team and no outside sponsor — the working is shown for every number so you can verify it yourself rather than take our word for it.

The site does not perform physical-hardware benchmarks. GPU specifications come from manufacturer datasheets and model specifications from each model’s published card; both are cited on the methodology page. Every figure on this site is produced by the same reference calculation engine, and a parity test fails the build if any page drifts from it.

How the numbers are checked is set out in the methodology; how content is produced and corrected is in the editorial policy; and every material change is logged in the changelog.