Welcome to poorpaul.dev

We're excited to launch poorpaul.dev — the companion site for Poor Paul's Benchmark (PPB).

What is Poor Paul's Benchmark?

PPB is an open benchmarking framework for local AI inference on consumer, prosumer, and small-business hardware. It measures real-world throughput and latency metrics across models, quantizations, hardware, runtimes, and workload settings.

Every benchmark result is normalized and published to a public Hugging Face dataset that anyone can download, analyze, and build on.

Why poorpaul.dev?

The raw dataset on Hugging Face is great for data scientists and researchers who want to run their own analysis. But for everyone else, it can be hard to quickly answer questions like:

• Which GPU gives the best throughput for my favorite model?
• How does latency scale as I increase the context length?
• What hardware should I buy if I care about time-to-first-token?

poorpaul.dev makes those answers easy to find with:

• A curated leaderboard showing the best results per model + GPU combo
• Interactive charts for exploring throughput, latency, and hardware comparisons
• Benchmark writeups and analysis articles (like this one!)

What's next?

We're just getting started. Expect more analysis posts, more hardware in the dataset, and more ways to explore the data. If you want to contribute benchmarks from your own hardware, check out the CLI on GitHub.

Happy benchmarking! 🚀