# Warm-up

> Warm-up: Untimed runs before a benchmark starts timing, so one-off costs such as loading and kernel compilation don't land in the results.

Canonical: https://llmdojo.dev/glossary/warm-up

Also called warm-up runs.

Untimed runs before a benchmark starts timing, so one-off costs such as loading and kernel compilation don't land in the results.

Learn it in the lesson: [What a fair benchmark holds fixed](https://llmdojo.dev/learn/inference-optimization/benchmarks-and-capstone/what-a-fair-benchmark-fixes) (Inference and Model Optimization · Benchmarking Inference, and the Module Capstone).

## Interview questions about Warm-up

- [You time a GPU request by reading the clock right after calling the model and get 0.3 ms per token. What went wrong, and what else can silently skew a timing?](https://llmdojo.dev/interview/q/you-time-a-gpu-request-by-reading-the-clock-right-after-calling)
- [What does a fair inference benchmark hold fixed?](https://llmdojo.dev/interview/q/what-does-a-fair-inference-benchmark-hold-fixed)

## Related
- [MLPerf Inference](https://llmdojo.dev/glossary/mlperf)
- [Glossary of every term](https://llmdojo.dev/glossary)