Someone built a full Metal driver stack to run NVIDIA GPUs on macOS Sequoia

NVIDIA dropped official macOS support after Mojave, and Apple Silicon made the prospect of revival seem permanently closed. This driver reopens it — not through reverse engineering Apple's proprietary GPU stack, but by building a complete alternative path from Metal down to the hardware.

The architecture is the interesting part. A Metal driver plugin intercepts Apple's AIR shader IR and translates it to SPIR-V, which then feeds Mesa's NVK Vulkan back end (patched for macOS against a specific Mesa commit). Beneath that sits NVIDIA's own open GPU kernel modules from the r610 release, ported to run as macOS kernel extensions. The GPU firmware is NVIDIA's unmodified GSP blob at version 610.57.04. Nothing here pretends to be Apple's GPU stack — it replaces it entirely for the NVIDIA card while leaving the rest of the system alone.

The honesty in the README is notable. Physical validation is explicitly limited to an RTX 5060 under macOS 15.7.x and 15.8.1. The device table covers GTX 16, RTX 20/30/40/50, TITAN RTX, and various Quadro and workstation cards, but the README is direct that listing a device is not the same as qualifying it. macOS 26 Tahoe carries a separate NVAccel build but is explicitly not a verified upgrade path.

The installation story has two tracks. The manual route involves downloading a tarball, verifying SHA-256 checksums on every file, running an install script, and then navigating OpenCore NVRAM settings with specific SIP flags, boot-args, and BAR resize values. The 1401 Mac app wraps all of that — it finds the OpenCore partition that actually started the machine, shows every change before making it, backs up the config, and adds a removal entry to the boot picker that cleanly reverses the install. The crash reporting path is also thoughtful: the app strips names, serial numbers, and addresses before sending logs, and shows a SHA-256 for each file it transmits.

The OpenCore configuration requirements are precise and non-obvious. The BAR settings that work for running the driver are the opposite of what the macOS installer needs, because the installer has no NVIDIA driver and runs on firmware video. The 1401 app handles that switch automatically; manual installers have to manage it themselves.

Licensing is noncommercial for the core components — personal, research, and educational use is permitted, commercial use is not. The shader translator carries LGPL-3.0 from its metal2vulkan lineage, Mesa changes are MIT, and the firmware is under NVIDIA's redistributable terms.

For anyone running an Intel Mac or an OpenCore Hackintosh with a recent NVIDIA card, this is the only working path to Metal acceleration on Sequoia. The scope is real and the caveats are stated plainly.