LemonSeed Studio: Bringing Linux AMD GPU Drivers to iPad

Project Author

LemonSeed Studio is an innovative project developed by Geramy that transforms the iPad into a powerful development environment. By leveraging a Thunderbolt connection to an external AMD GPU, it enables local LLM (Large Language Model) inference directly on iPadOS.

The core achievement of this project is the integration of the ~~standard cloud-reliant~~ unmodified upstream Linux amdgpu driver, functioning within the Apple ecosystem via DriverKit.

🛠️ Technical Architecture

The system is composed of three primary pillars that allow the iPad to communicate with high-performance hardware.

Component	Technical Role	Implementation Detail
`mac_linuxgpu`	GPU Driver	Unmodified Linux `amdgpu` + `amdkfd` embedded as a `PCIDriverKit` extension.
`LemonSeed Engine`	Inference Engine	Executes LLMs using the HSA runtime and HRX/Loom.
`amdgpu_mtopg`	Telemetry	Provides real-time monitoring of GPU performance.

System Data Flow

💻 Development & Setup

The project is structured into several key directories: Engine, Studio, Toolchain, docs, and third_party. To get the environment running, follow these steps:

Clone the repository (including all submodules).
Initialize dependencies for existing clones.
Build the project (Note: mac_linuxgpu will sparsely fetch pinned Linux sources during the first build).

Command Line Instructions:

# Clone the full project with submodules
git clone --recurse-submodules https://github.com/Geramy/LemonSeed-Studio

# If already cloned, update the submodules
git submodule update --init

Note: Because mac_linuxgpu handles its own nested kernel submodules via sparse fetching during the build process, you do not need to clone those specific nested components recursively.

📊 Performance Logic

The ability to run local inference can be conceptualized as a function of hardware throughput and driver efficiency. If we represent the total compute capacity as $C$ , the relationship is roughly:

$C = \int (\text{GPU}_{\text{TFLOPS}} \times \text{Driver}_{\text{Efficiency}}) \, dt$

By using the amdgpu driver, the project maximizes $\text{Driver}_{\text{Efficiency}}$ on non-native hardware.

📜 Licensing and Legal

The primary codebase is dual-licensed, allowing users to choose between:

MIT License
GPL-2.0-only

Important: The various bundled components and submodules maintain their own independent licenses. For a comprehensive list, please refer to the THIRD_PARTY.md file.

Project Tags: amdgpu code-editor driverkit egpu ide ipad ipados llm local-llm swiftui