gregordinary/rockchip-npu-notes — explained in plain English
Analysis updated 2026-05-18
Look up how a specific RK3588 NPU operation, like matrix multiplication or convolution, actually behaves in hardware.
Avoid known hardware quirks and bugs when writing low-level software for this chip.
Cross-check hardware behavior observed through testing against other open source driver documentation.
| gregordinary/rockchip-npu-notes | aclark4life/home-depot-crawl | aetheria-labs1/storefront-skills | |
|---|---|---|---|
| Stars | 6 | 6 | 6 |
| Language | Python | Python | Python |
| Last pushed | — | 2014-08-10 | — |
| Maintenance | — | Dormant | — |
| Setup difficulty | hard | moderate | easy |
| Complexity | 5/5 | 2/5 | 3/5 |
| Audience | researcher | developer | pm founder |
Figures from each repo's GitHub metadata at analysis time.
Requires actual RK3588 hardware and low-level driver knowledge to make practical use of these notes.
rockchip-npu-notes is a collection of research notes about a specific piece of computer hardware called the RK3588 NPU, a chip built into some small computers that speeds up AI and machine learning tasks. The notes explain how this chip actually behaves at a very low level, based on work the author did while building software that talks directly to the chip on a real device. The README is upfront that these documents were written by an AI, mainly Claude Code, as part of a series of side projects, and it notes that accuracy is not guaranteed. The notes were gathered by reverse engineering the chip, meaning the author experimented directly with the hardware and observed how it responded, rather than reading an official manual, since a full manual for this chip is not publicly available in that depth. Each fact recorded in these notes is labeled with how confident the author is in it: some facts come from directly testing the hardware and watching what happens, and others are backed up by reading other open source projects that work with the same chip. When a fact matches both kinds of evidence, both labels appear together. The notes also carefully record things that do not work, since the author says those were the most expensive lessons to learn. The repository is organized as a set of separate documents, each covering one part of how the chip processes information, such as how it handles matrix multiplication, different numeric precision formats, and how certain operations are chained together for speed. It also documents specific known quirks and bugs in the chip's behavior that could otherwise cost someone significant time to discover on their own. This project is aimed at people building low level software that runs directly on this NPU hardware, rather than everyday users of AI tools. The README does not state a license for this repository.
A collection of AI-assisted reverse engineering notes documenting exactly how the RK3588 NPU chip behaves at a hardware register level.
Mainly Python. The stack also includes Python, RK3588 NPU.
Setup difficulty is rated hard, with roughly 1day+ to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
Don't trust strangers blindly. Verify against the repo.