Disclaimer

Last updated 9 August 2026.

What the numbers are

Every figure published here was measured on one specific machine, running one specific software stack, in one specific configuration, on the date recorded in the result file. Inference throughput moves with driver versions, server builds, quantisation, context length, batch size and thermal conditions — this site has an article about a system update that shifted every measurement on the machine by a few percent.

They are measurements, not guarantees, and they are not a prediction of what identical hardware will do for you. The methodology page sets out how each number is produced and, more usefully, what it does not tell you.

The results may also simply be wrong. Two articles here carry published corrections for measurement errors I made and later caught. If something looks wrong to you, tell me — that is the fastest way it gets fixed.

Running any of this is your own risk

Articles include commands, container images, environment variables and server flags. They are provided as a description of what I did, not as instructions warranted to be safe on your system. Serving models consumes very large amounts of memory, driver and firmware updates can leave a machine unbootable, and a command that is harmless here may not be on a differently configured host. Back up anything you cannot lose, understand what a command does before running it, and accept that the consequences are yours.

Not purchasing advice

The hardware discussed here is expensive, and nothing on this site is a recommendation to buy any of it. Treat these numbers as one input among several, verify anything a purchase depends on, and remember that a machine which suits my workloads may be a poor fit for yours.

Commercial relationships

This is a commercial site. Some posts contain affiliate links, and each such post opens with a disclosure. The privacy page explains what those links do, and about explains why they do not reach the measurements. No manufacturer has supplied hardware, funded an article, or reviewed anything here before publication.

Trademarks and affiliation

NVIDIA, DGX and CUDA are trademarks of NVIDIA Corporation. All other product names, model names, logos and brands appearing on this site are the property of their respective owners, and are used only to identify the hardware and software being measured.

Local LLM Labs is not affiliated with, endorsed by, sponsored by or associated with NVIDIA Corporation or any other company whose products are measured here. Use of a product name does not imply any relationship with its owner.

No warranty

This site and everything on it is provided as is, without warranties of any kind, express or implied. To the fullest extent permitted by law, Local LLM Labs and its publisher accept no liability for any loss or damage arising from reliance on anything published here, including damage to hardware or data and decisions made on the basis of these measurements.

Questions about any of this go to [email protected].