LemonSeed Studio runs unmodified Linux amdgpu driver on an iPad for LLM inference

An iPad app that embeds a Linux GPU driver

An open-source project called LemonSeed Studio has drawn attention after reaching the front page of Hacker News. According to the project's GitHub repository, it is a single iPad application that ships the upstream Linux amdgpu and amdkfd drivers — unmodified — packaged as a PCIDriverKit extension. The driver controls an AMD graphics card sitting in a Thunderbolt enclosure attached to the iPad, and the app uses that GPU to run large language model inference locally.

The project describes itself as an editor and IDE for iPad with on-device LLM inference on the external AMD GPU, which is an unusual destination for Linux kernel driver code.

What the app bundles

The repository is organised around three components, each pulled in as a git submodule:

• mac_linuxgpu, the GPU driver: the unmodified upstream Linux amdgpu and amdkfd driver, embedded in the app as a PCIDriverKit extension.
• LemonSeed Engine (LSE), the inference engine, which runs on that GPU through its HSA runtime and what the project calls HRX/Loom.
• amdgpu_mtopg, a GPU monitor providing live telemetry.

In the Linux world, amdgpu is the kernel-mode driver for modern AMD graphics hardware, while amdkfd exposes the compute side through the Heterogeneous System Architecture (HSA) interface that AMD's compute stack builds on. Carrying both means the app does not just light up the card — it gets the compute path needed to run inference workloads on it.

How the driver gets there

Apple does not permit third-party kernel extensions on iPadOS. The supported route for external hardware is DriverKit, Apple's userspace driver framework, with PCIDriverKit covering PCI devices attached over buses such as Thunderbolt. That is the mechanism this project uses: the Linux driver source is hosted inside an app extension rather than loaded into a kernel.

The repository does not explain in depth how the kernel-facing Linux code is adapted to DriverKit's APIs; that work lives in the mac_linuxgpu submodule. It does spell out the build arrangement: the driver fetches its own pinned Linux sources sparsely on first build, so the nested kernel submodule does not need to be cloned recursively. A standard git clone --recurse-submodules is enough to pull down the three top-level components.

Licensing

The project's original code is dual-licensed, available under MIT or GPL-2.0-only at the user's option. Bundled and submodule components keep their own licenses, with the details collected in the repository's THIRD_PARTY.md file — a sensible arrangement given that amdgpu itself is GPL-2.0 Linux kernel code.

Why it matters

For GPU and iPad developers, this is a notable proof of concept. It shows that a substantial real-world Linux driver — the full upstream AMD graphics and compute stack — can be hosted on iPadOS through Apple's userspace driver framework instead of being rewritten against Apple's native APIs. That matters to anyone building iPad accessories who already has Linux driver code to reuse.

The local-AI angle is distinctive too. Rather than running models on the iPad's own GPU through Metal, LemonSeed Studio pushes inference out to a desktop-class AMD card over Thunderbolt, using AMD's compute stack rather than anything Apple ships. The repository makes no performance claims and does not specify supported GPU models or iPad requirements, so it reads as an experimental project rather than a product — but as a demonstration of how far iPadOS's driver model can be stretched, it is a striking one.