I built LapTime while researching hardware and running into the same question over and over: what will this actually feel like to use?
There’s a lot of great benchmark data out there, but I kept wanting a more intuitive way to understand the practical experience across prompt ingest, time to first token, and generation speed.
So I built a small simulator that tries to make that easier to compare:
- Live app: https://laptime.run
- Hugging Face Space: LapTime - a Hugging Face Space by timmcg
- Source: GitHub - mcglothi/laptime: Local LLM benchmark playback simulator for comparing how hardware and models feel in practice. · GitHub
You can compare different setups side by side and get a rough feel for how the experience changes across Apple Silicon, NVIDIA GPUs, GB10 systems, and Strix Halo machines.
A few things I tried to be careful about:
- direct benchmark-backed rows are kept separate from modeled estimates
- community/forum references are labeled separately rather than presented as exact measured data
- source links are exposed so people can inspect where the assumptions come from
The goal isn’t to replace benchmark tables or original sources, just to make them easier to interpret when you’re trying to decide what to run or buy.
Would love feedback on:
- whether this framing is useful
- where the current estimates feel too hand-wavy
- what hardware/model comparisons would be most useful to add