AI inference glossary
Benchmark metrics

Recipe

Also known as configuration, serving recipe

In plain English

A recipe is the complete set of choices behind one curve: which model, engine, image, precision, parallelism, and workload were run.

Technical definition

A recipe is the fully specified combination of model, inference engine and container image, numerical precision, parallelism strategy, chip system, and workload that produces one measured curve.

Engineering details

Every point on InferenceX belongs to a recipe, and a concurrency sweep across one recipe traces out its curve. Changing any element produces a different recipe rather than a variation of the same one, which is why an engine image bump is reported as its own result rather than folded into an existing line.

Why it matters

Peak chip specifications do not describe serving performance, and the same silicon can differ by multiples across recipes. Naming the whole combination is what makes a claim checkable: a number without its recipe cannot be reproduced or fairly compared against another vendor.

How to read it in InferenceX

InferenceX benchmark configs mainly track the published vLLM and SGLang cookbooks on upstream images, so results reflect what users can actually deploy rather than images tuned for the benchmark. Point tooltips expose the recipe behind each point along with links to the run provenance.