Passthrough for eGPUs on M-series Macs
A very unofficial, very experimental passthrough runtime for getting NVIDIA* RTX 30/40/50-series GPUs in Thunderbolt eGPU enclosures doing useful work on macOS through Linux virtualization.
* Compatibility/purpose labels only. PEGPU is not endorsed by, sponsored by, affiliated with, or encouraged by Apple, macOS, NVIDIA, Linux, Thunderbolt, QEMU, UTM, Debian, llama.cpp, or any other named company, protocol, project, or maintainer.
To disable SIP:
csrutil disableRestart your Mac normally.
Important: Understand and consider the security risks of disabling SIP before proceeding.
Scott J. Goldman found the wild bit: a way to pass Thunderbolt 3/4 eGPU PCIe devices into a Linux VM on Apple Silicon by combining customized QEMU with a macOS DriverKit host extension. PEGPU builds on that base and tries to turn the whole thing into something you can actually live with.
Native macOS UI, UTM-derived GUI display integration, ANGLE and CocoaSpice support, sidecar metrics, embedded file manager, SSH terminal, VM setup, port routing, web UI tabs, llama.cpp RPC orchestration, llama-swap style runtime management, and less ritual around starting the thing. PEGPU is distributed under the permissive Apache 2.0 license and uses components with their own licenses and notices.
It owns the patched QEMU runtime, the DriverKit/VFIO mechanics, firmware, guest DMA driver package, VM launch shape, and the UTM/QEMU-side mechanics that make accelerated visual sessions possible. PEGPU Machine is based on QEMU* and distributed under GPL 2.0. Components carry their own licenses and notices.
GUI. The internal Linux desktop is rendered inside the macOS app through the UTM-derived SPICE, ANGLE, and CocoaSpice path, so setup and normal GUI work stay accelerated without using the passed-through NVIDIA card for the desktop.
External displays. Secondary-click in the GUI tab to turn on external display sessions on eGPUs. There is no fixed limit: each GPU has its own external display session, and every monitor connected to that GPU shares the same session so Linux/NVIDIA settings can extend or mirror them. Move into a session with the app-shown shortcuts for that GPU, from ⌥⌘2 through ⌥⌘9; ⌥⌘1 always releases control back to the Mac. The main Mac-driven GUI does not lock the pointer because it uses absolute mouse mirroring directly on the screen.
Runtime. The Runtime view starts and stops the PEGPU Machine server, shows whether the VM/control routes are alive, and keeps Mac, guest, network, disk, and VM GPU telemetry in one place. It also launches the recipe-based Debian setup and NVIDIA/CUDA helper install flow, because the download does not ship a bundled VM or preinstalled GPU driver stack.
Models. The Models view manages model files and llama.cpp launches. Pick the Mac backend, the Linux/eGPU backend, or a split run where llama.cpp RPC bridges host and guest; TurboQuant is only the optional llama.cpp build/flag path for KV-cache quantization, not a separate GPU target. The UI assembles the real launch details: devices, RPC endpoint, GPU layers, tensor split, and cache type. Tip: Models does not repeatedly probe SSH for devices. With the VM running and GPU passthrough active, click Refresh Devices once to surface all connected GPUs.
Chat. The Chat tab talks to the currently loaded model through the local OpenAI-compatible route. It is there to test the selected model, backend, and split inference path from inside PEGPU before you point another app at the same runtime.
Files. The Files tab is a two-pane Mac and Linux file manager for the shared runtime paths. It handles browsing, drag/drop copy or move, transfer jobs, mount repair, and Finder handoff, so moving models, downloads, outputs, and VM files does not require falling back to ad hoc scp commands every time.
A managed guest and host llama.cpp* runtime can split work across macOS* unified memory GPUs and NVIDIA* GPUs inside the VM. In my test, a Qwen* 3.5 122B MoE LLM quantized as Q4_K_M reached 28.14 tokens per second split across two devices: an M1 Max* with 64 GB unified memory handling 32 layers, and an RTX* 3090 with 24 GB VRAM handling 16 layers. When a workload runs fully on the passed-through eGPU, it runs at the card's normal Linux* speed too, including ComfyUI* video generation models such as Wan* 2.2.
Use external NVIDIA*-driven displays for 3D rendering, ComfyUI* video generation, random Linux* desktop things, or games. Use the internal UTM*-style window when you want smooth GUI output without spending the external GPU on the desktop.
PEGPU downloads the official Debian* cloud image and scripts the setup. It does not ship a bundled VM and it does not vendor NVIDIA* drivers or CUDA*. The helper only installs from NVIDIA* repositories inside Linux*.
XFCE* on Xorg* is light, old, and conservative in exactly the useful way. The Wayland* display path from the base project can be luck-based on external screens, especially with wider or stranger monitors, and failures around custom DMA remapping can turn ugly fast. Xorg* handles this more gracefully in my testing: internal SPICE*/GL stays usable, external NVIDIA* displays can run on one or more eGPUs, and the XFCE* tree only adds about 250 MB idle RAM compared with headless mode.
* Names including NVIDIA, RTX, CUDA, macOS, Apple, Mac, M1 Max, Linux, Debian, Wayland, Xorg, XFCE, UTM, SPICE, Qwen, llama.cpp, ComfyUI, and Wan are compatibility or purpose references only. PEGPU is not affiliated with, endorsed by, sponsored by, or approved by those companies, projects, maintainers, or trademark owners.
Stop the server from PEGPU, quit the tray runtime, or in the worst case kill QEMU from Activity Monitor. Do not yank the Thunderbolt cable while the VM server is alive. That is how you buy yourself a frozen Mac and a reboot.
If you stopped the server and want to launch with eGPUs attached: unplug the eGPUs, wait 5 seconds, plug them back in, wait another 5 seconds, then start the server. Most of the time the GPUs attach.
Start the VM headless or GUI, let Debian finish its first boot work, run the NVIDIA installer if you use NVIDIA cards, stop the VM completely, then plug in your eGPUs and launch the server with the devices attached.
Treat the external GPUs as either the graphics/rendering path or the compute path for a session. Heavy 3D rendering while loading AI models or other large compute jobs can overload PCIe passthrough and may kernel panic the Mac.
Thunderbolt PCIe passthrough on Apple Silicon is not on parity with mature Intel Thunderbolt passthrough paths. Expect bugs, occasional macOS kernel panics, and forced restarts while the driver and runtime improve.
If you mainly want VLMs, ComfyUI, or other AI workloads, use the PEGPU GUI display as your primary display. It uses the Mac's internal graphics for 3D rendering, so external cards are not competing for limited passthrough capacity. Advanced users can enable headless-only mode in Developer Options and control the VM through the embedded terminal or SSH.
PEGPU is experimental kernel-driver-adjacent software for Apple Silicon Macs with SIP disabled, and the DriverKit extension is ad hoc signed. The warnings below are expected for this kind of package, but they are still real security decisions.
csrutil disable, restart, and understand the security cost before continuing. PEGPU depends on an experimental DriverKit/VFIO path for Thunderbolt PCIe passthrough, so SIP must be disabled before macOS will allow the required host-side extension flow.
xattr -dr com.apple.quarantine ~/Downloads/pegpu*.pkg
.pkg. It installs the app plus the Machine components that let QEMU and the DriverKit extension talk to Thunderbolt PCIe hardware.
PEGPU Machine, approve the System Extensions prompt, and confirm the toggle is on.
PEGPU does not ship a pre-made VM, and it does not pre-install GPU vendor drivers, because licensing and distribution of VM images, NVIDIA drivers, and CUDA packages gets complicated fast. First startup can take half an hour depending on the Mac and network because it fetches a clean Debian cloud image from the official Debian repository, then uses scripts to install packages and drivers from their official repositories. NVIDIA/CUDA installation can add another 20-ish minutes, and only runs if you explicitly click the NVIDIA driver install helper and accept the script and terms. You can also ignore the helper and install drivers yourself through the GUI, SSH, or the embedded terminal. Headless mode skips the XFCE dependency tree if all you want is compute, SSH, and terminal work.
* These are credits to the upstream and original projects that PEGPU builds on, derives from, integrates, or was inspired by. PEGPU is not affiliated with, endorsed by, sponsored by, or approved by those projects, maintainers, companies, or trademark owners.
PEGPU is experimental, imperfect, and built in the open. If it helps, breaks, half-works, or does something strange on your monitor setup, that is useful signal. Issues, fixes, logs, and better ideas are welcome.