ShimetaPi Software Repository

Repository URL: https://mirrors.shimetapi.cn/

Overview

The ShimetaPi Software Repository is the unified software distribution and update system for all ShiMeta Pi series development boards. It addresses the fundamental problem of hardware fragmentation — each chip platform (GK7206, RK3588, T507, etc.) has its own independent NPU architecture, multimedia framework, and C library baseline. The repository provides a single source for drivers, tools, SDKs, and application experiences, automatically adapting to each platform.

The Fragmentation Pain Points

Pain Point	Impact
Massive, high-barrier SDKs	Each chip platform has its own development package — toolchains, driver APIs, and media pipelines are incompatible
Model-to-deployment "last mile" gap	Even when the training platform can export compatible files, users don't know how to load models onto NPUs, connect cameras, or visualize inference results
Low-level driver burden	Developers waste time on MIPI initialization, VB buffer pool configuration, YUV image preprocessing, and other tedious details
Weak platform lock-in, diluted brand value	Fragmented tools lack a unified entry point; competitors can easily reuse open-source components

Design Goals

Unified distribution — All boards pull from the same source; architecture auto-routing eliminates manual selection
High encapsulation — Shields users from system base layer (kernel, firmware), hardware abstraction layer (NPU/camera/GPIO tools), and development layer (SDK/cross-compilation), enabling one-click AI inference and WebUI visualization
Cross-platform consistency — Core commands (shimeta_infer, shimeta_camera) maintain identical invocation and output formats across all boards
Safety and isolation — Strict repository isolation for the three C library baselines (uclibc for GK7206, glibc for RK/Most Allwinner, musl for T133) prevents cross-libc system crashes

Architecture

Repository Structure

  http://mirrors.shimetapi.cn                          ← Homepage (board listing)
  │
  ├── /repos/GK7206/                                   ← GK7206 board root directory
  │   ├── install.sh                                   ← One-click install script
  │   ├── install_models.sh                            ← SVP model install script
  │   ├── svp_models.tar                               ← SVP model pack
  │   └── bare/                                        ← Bare-metal install modules
  │       └── shimeta-gk7206_0.2.0_arm.tar.gz          ← Main install package (5.5MB)
  │
  ├── /repos/RK3588/                                   ← Coming soon

One-Stop AI Software Stack

The flagship product of the application layer, bridging the "model training platform → board deployment → inference visualization" last mile:

Model Training Platform (Cloud GPU) → Board Deployment (NPU Device) → Visualized Results (RTSP/WebUI)

Core tools currently delivered:

Pico G1 (GK7206)

Tool	Description
`shimeta_infer`	Single-image NPU inference, supports custom trained models
`shimeta_camera`	Real-time camera inference with custom models, MJPEG WebUI + HTTP `/api/status`
`shimeta_svp`	33 pre-compiled scene detection models for real-time camera inference, RTSP stream output

Future cross-platform evolution: commands like shimeta_infer xx.xmm.zip xx.png 0.5 0.5 will deliver a fully consistent experience across GK7206, RK3588, and T507.