A community leaderboard of AI inference on Intel hardware. How fast do local models actually run on Arc cards, Core Ultra chips, Xeon, and Gaudi? Not in a vendor's slide deck.
People sign in with GitHub, register the rigs they run models on, and post the decode tokens per second they get for a model at a given quantization on a given runtime. Results are self-reported and confirmed by the community. The hardware database is open and grows by pull request.
Powered by Cascadia.
A rig is a machine, and it decomposes into parts. A result can describe the whole rig or one unit inside it — a single card, the CPU cores, the iGPU, or the NPU. A Core Ultra chip therefore holds three separate numbers, and the boards keep them apart.
A board is one model at one quantization. Each rig, or each part at a given quantity, appears once per board at its best decode tok/s; the earliest run wins a tie. Boards mix runtimes, and filter by runtime when you want to compare like with like.
Every result starts self-reported. Once enough signed-in members confirm it, it becomes community-verified. Implausible numbers get flagged, and past a few flags an entry is hidden pending review.
Two things grow by pull request, and both are checked automatically when you open one:
| You want to | Go to |
|---|---|
| Post a benchmark result | results/ — add results/<your-handle>/<name>.json, the check verifies your linked GitHub account, and merging imports the result automatically |
| Add a GPU, CPU, model, or runtime | frontend/src/catalog/ — one file, one entry |
You can also submit a result straight from the site, no git involved. CONTRIBUTING.md has the full picture, including how to run the app locally.
frontend/ Vite + React 19 + Tailwind 4 single-page app
src/catalog/ The hardware, model, quant, and runtime catalog — source of truth
src/lib/api/ Typed API contract, with supabase, mock, and live adapters
src/og/ Open Graph card rendering, shared by the app and the edge functions
api/og/ Vercel edge functions that serve the share images
scripts/ Catalog seeding, mark generation, OG previews, validators
results/ Benchmark results submitted as files, one folder per GitHub handle
supabase/ Migrations and database smoke tests
docs/ API contract and Supabase environment notes
Node 22 or newer.
npm --prefix frontend install
npm --prefix frontend run devThat is the whole setup. With no frontend/.env the app runs on generated mock data — every page, every
flow, no credentials, no backend — on port 5173.
To point it at a real Supabase project instead, copy frontend/.env.example to frontend/.env and fill in
the project URL and publishable key. docs/SUPABASE.md covers the hosted environments.
All run from the repo root.
npm --prefix frontend run dev # dev server on :5173
npm --prefix frontend run build # typecheck and build to frontend/dist
npm --prefix frontend run typecheck # tsc, no emit
npm --prefix frontend run catalog:validate # check the catalog
npm --prefix frontend run results:validate -- results/<handle>/<file>.json
npm --prefix frontend run og:preview # render the share cards to frontend/.og-preview/Vite, React 19, Tailwind 4, react-router, and shadcn/ui components on Base UI primitives, dark mode only. Supabase for auth (GitHub OAuth), Postgres, and rig photos, with row-level security enforcing ownership. Deployed on Vercel; the Open Graph cards are rendered by edge functions with Satori.
- docs/API.md — the typed API contract the front end is written against
- docs/SUPABASE.md — environments, migrations, and OAuth setup
MIT, for the code and the data in this repository.
Two things the licence does not cover. The brand assets in frontend/public/logos/ (the Intelinside lockup
and the Cascadia wordmark) are not licensed for reuse, so a fork needs its own name and mark.
The vendor logos beside them, and every hardware, model, and runtime name in the catalog, are trademarks of their owners, used here only to identify what a result ran on.
Intelinside is an independent community project. Intel, Arc, Core Ultra, Xeon, and Gaudi are trademarks of Intel Corporation. We are not affiliated with, endorsed by, or sponsored by Intel Corporation, or by any other vendor or project named here.