LapTime

Test-drive local LLM hardware before you buy.

Visit the live site: https://laptime.run

LapTime turns benchmark numbers into something engineers can actually feel. Instead of staring at isolated tok/s figures, you can pick hardware, models, and workloads, then watch the tradeoffs play out across prompt ingest, time to first token, and streamed output.

Why it exists

Benchmark tables answer "what is faster?"

LapTime answers:

• How long will I wait before the model starts talking?
• Which part is slow: prompt processing, TTFT, or generation?
• Will this rig actually fit the model I want to run?
• How different will an Apple laptop feel from a GPU tower or a GB10 box?

Live product

• Production: https://laptime.run

Current features

• Interactive simulator for local LLM workloads
• Hardware, model, and workload selection with searchable controls
• Platform-aware hardware filtering
• Color-coded fit warnings for likely broken or risky combinations
• Segmented playback timeline for prompt ingest, TTFT, and token generation
• Side-by-side comparison view
• Broader model browser and source explorer
• In-app methodology section that explains measured versus estimated versus community-backed laps
• Cloudflare Pages deploys on every push to main

Hardware coverage

The catalog includes a mix of exact benchmark-backed entries and clearly marked estimate-based entries across:

• Apple Silicon laptops and desktops
• NVIDIA consumer, workstation, and datacenter GPUs
• GB10-class systems such as DGX Spark and partner variants
• AMD Strix Halo systems, including Framework, HP, ASUS, and community-tracked mini PCs

Data philosophy

LapTime tries to stay honest about source quality.

• Exact benchmark rows are labeled from structured sources like LocalScore
• Community/forum observations stay separate from high-confidence benchmark data
• Newer or harder-to-source systems are included as estimates only when labeled clearly
• Fit checks are guardrails, not guarantees; long context, KV cache growth, backend choice, and offload behavior still matter

Stack

• React 19
• Vite 8
• Plain CSS with a custom design system
• Cloudflare Pages for hosting
• GitHub Actions for automatic deploys

Local development

npm install
npm run dev

Quality checks

npm run lint
npm run build

Deployment

LapTime deploys to Cloudflare Pages from GitHub Actions on pushes to main.

Required repository secrets:

• CLOUDFLARE_API_TOKEN
• CLOUDFLARE_ACCOUNT_ID

Workflow:

• .github/workflows/deploy-cloudflare-pages.yml

If you need to create the Pages project manually, use:

• Framework preset: Vite
• Production branch: main
• Build command: npm run build
• Output directory: dist

Roadmap

• Expand benchmark ingestion from more structured sources
• Expand the new methodology and source-quality experience into deeper attribution pages
• Broaden buyer flows and hardware landing pages
• Improve context-aware fit modeling for long prompts and offload-heavy runs

Contributing ideas

If you have benchmark data, hardware ideas, or real-world validation feedback, open an issue or share the site with another engineer and tell us where the simulator feels right or wrong.

License

The code in this repository is licensed under the MIT License.

Benchmark data, source links, and third-party referenced materials remain subject to their respective licenses, terms, and attribution requirements. LapTime's code license does not imply relicensing of third-party data or source content.