# AMDai

> Local AI on consumer AMD Radeon — ROCm notes from a real R9700 test rig.

The consumer AMD ROCm notebook: llama.cpp and Qwen on Radeon GPUs, grounded in a two-R9700 test rig. Daily driver, fast agent stack, GPU directory, and a quant + hardware calculator. Measured numbers come from amdai-bench/run@1 cards on a two-R9700 local desktop. Estimates stay labeled. Q3 is estimate-only. Q1/Q2 stay off. AMDai is not an AMD publication.

## Docs

- [R9700 Lab Notes](https://amdai.net/): The consumer AMD ROCm notebook: llama.cpp and Qwen on Radeon GPUs, grounded in a two-R9700 test rig. Daily driver, fast agent stack, GPU directory, and a quant + hardware calculator.
- [New models](https://amdai.net/new-models/): Daily Hugging Face Hub table of new open-weight text and vision releases. GGUF Q3–Q8. Runs-on from cited engine lists. Q4_K_M fit is a labeled estimate.
- [Lab log](https://amdai.net/lab-log/): Results log of measured and cited findings, plus informative failures. Dual R9700 speed and the hourly Idea lab. No invented tok/s.
- [Getting started](https://amdai.net/start/): Check a rig or budget, first-run Ubuntu setup, then a measured 35B recipe.
- [First-run setup](https://amdai.net/setup/): Ubuntu 24.04+ to first token on a Radeon box. llama.cpp Vulkan first.
- [GPU directory](https://amdai.net/gpus/): Radeon cards with gfx IDs, VRAM, official ROCm rows, and tested recipes.
- [Radeon AI PRO R9700](https://amdai.net/gpus/r9700/): Measured 35B llama.cpp on the house 32 GB card — one GPU vs two.
- [Where to buy the Radeon AI PRO R9700](https://amdai.net/where-to-buy-r9700/): US store listings for the 32 GB R9700. Dated Newegg snapshot. No invented price.
- [Inference engines](https://amdai.net/backends/): llama.cpp first. Vulkan daily driver. Cited engines, no invented benches.
- [Serve many users on two Radeon cards](https://amdai.net/serving/): Several users on 1–2 Radeon cards. llama.cpp -np and --split-mode first. Third-party figures labeled. Our lab 35B and Flash.
- [Quant + hardware calc](https://amdai.net/calculator/): Size a GGUF against Radeon VRAM. Estimates stay labeled. No prices.
- [Public API](https://amdai.net/api/): Static JSON at /api/v1/. GPU, model, stack, measured, and ROCm catalogs for assistants. MCP server in this repo.
- [API index.json](https://amdai.net/api/v1/index.json): Endpoint list, schema version, and as_of for the public static API.
- [Benchmarks](https://amdai.net/benchmarks/): Cited 35B llama.cpp matrix from local desktop amdai-bench/run@1 cards.
- [Official run cards](https://amdai.net/results/): amdai-bench/run@1 cards credited to AMDai, with the exact launch command.
- [How we test](https://amdai.net/methodology/): Measured vs estimate. Q3 is estimate-only. Q1/Q2 stay off.
- [FAQ](https://amdai.net/faq/): First-time Radeon questions. Points at measured pages.
- [Glossary](https://amdai.net/glossary/): Plain-English ROCm, GGUF, quant, and run-card terms.
- [About AMDai](https://amdai.net/about/): Independent notebook. Not an AMD publication.
- [Lab Notes digest](https://amdai.net/newsletter/): Email-only digest. Double opt-in. No tracking pixel.
- [Sponsor](https://amdai.net/sponsor/): Sponsorship slots for AMD/Radeon-relevant vendors: homepage banner, featured GPU and backend listings, newsletter mention. Labeled Sponsored. Contact for rates.

## Sections

- [Compare](https://amdai.net/compare/): 228 pages — 35B shopper questions, GPU-vs-GPU, and model-vs-model sheets.
- [Can my Radeon run it?](https://amdai.net/can-run/): 162 pages — Model × GPU fit from the same calculator sheets.
- [Harnesses](https://amdai.net/harnesses/): 4 pages — Local frontends, connect notes, quiz, and agent pages.
- [Official run cards](https://amdai.net/results/): 49 cards — amdai-bench/run@1 cards credited to AMDai.
- [GPU directory](https://amdai.net/gpus/): 27 pages — Radeon sheets with gfx IDs, VRAM, and recipes.
- [Recipes](https://amdai.net/recipes/): 7 pages — First-party launch recipes with sourced proofs.
- [Lab log](https://amdai.net/lab-log/): 101 entries — Measured and cited findings as of 2026-10-28.

## Full text

- [Lab log as Markdown](https://amdai.net/lab-log.md): every lab-log entry with metrics and sources.
- [llms-full.txt](https://amdai.net/llms-full.txt): full text of key pages.