What this is, and what it is not

Every published figure names its source: the exact sentence, the URL, and the date it was read. Model architecture comes from the publisher’s own config.json at an immutable revision, recorded with its SHA-256. Device capacity and bandwidth come from the manufacturer’s specification. Where a figure is not published anywhere, the calculation is refused and the missing field is named.

This costs something. It means some questions get an honest “cannot be answered” where a competitor prints a confident number, and a scorecard that shows the formats this engine handles worst on the same table as the ones it handles best. That trade is the product.

• No quality rankings. There is no leaderboard saying which model is smarter. That is a benchmark question and this is a hardware tool; borrowing someone else’s benchmark table and presenting it as ours would break the rule above.
• No hosted-model pages. A model you cannot download has no memory footprint on your machine, so there is nothing here to calculate about it.
• The answer is not behind a login. The diagnosis, the recommendation and the sources are reachable directly, and every endpoint of the API answers on the free plan too; the paid plans add volume, your own branding and support.
• Referral links never set the order. The cloud pages link to GPU and API providers, and some of them pay a commission when a reader signs up. Every table is sorted by the provider’s live price, a provider that pays nothing is listed first whenever it is cheaper, and the terms of each programme are published in full.
• Analytics is a choice. Google Analytics loads only where you allow it: on by default in the United States with a one-click opt-out, and behind consent everywhere else, with a Global Privacy Control or Do Not Track signal treated as a refusal. Vercel page views remain cookieless, and the embedded widget is never counted at all — the privacy page says exactly what is recorded.

The changelog records what changed and when, including the times a number here was wrong and got corrected.

LLM Bottleneck is built and maintained by Ander Pascal. One person, not a company and not an editorial board: the catalogue, the engine, the calibration and the corrections are all the same pair of hands, and the rule above is what keeps that from being a problem — every figure names a document you can open without taking my word for anything.

Where a number here turns out to be wrong, it gets fixed and the fix is written down in the changelog with what it used to say. The cases this engine handles worst are published on the accuracy page rather than left for someone else to discover.

I also maintain PC Bottleneck Calculator, an independent tool for estimating CPU and GPU balance in games. It has its own calculation engine and scope; its results are not used by LLM Bottleneck.

The most valuable message anyone can send is a wrong number. If a figure here disagrees with a source you have, that is a defect and it will be fixed and recorded in the changelog — send the source rather than the conclusion and it will be fixed faster.

These are monitored by a person, not a ticketing system, so a reply takes as long as it takes. There is no support contract behind any of them, and saying otherwise would be the sort of claim this site exists to avoid.