NVIDIA and AMD Video Cards Back on Mac Thanks to an Independent Driver Package

Support for NVIDIA GPUs on macOS seemed like a distant memory after Apple decided to discontinue compatibility with CUDA drivers in 2018. Everything changed in recent hours when The Tiny Corp, a small company that develops workstations and GPU boxes, released a driver package that enables the use of AMD and NVIDIA GPUs on Mac: TinyGPU.

The software has received Apple's official approval, a detail that removes the need for complex workarounds. It requires macOS 12.1 or later, a USB4 or Thunderbolt connection, and GPUs based on AMD RDNA3 or NVIDIA Ampere architectures and later. Installation is limited to authorizing the kernel extension, with no system modifications needed.

A significant first test came from developer and YouTuber Alex Ziskind, who connected a GeForce RTX 5090 with 32GB of VRAM to a Mac mini M4 Pro. The result shows that even cutting-edge hardware based on Blackwell architecture can operate on macOS through a single cable.

However, it should be noted that this is not native support and does not represent a return of official NVIDIA drivers. TinyGPU uses an alternative pipeline based on tinygrad and a compiler managed via Docker, specifically focusing on AI-related workloads.

The benchmarks reveal both strengths and weaknesses. In a test with Llama 3.1 8B, the RTX 5090 reached about 7.48 tokens per second, still far from the theoretical potential of an RTX 5090. However, the time to first token shows an interesting behavior, with times up to four times lower compared to Metal in chat-type scenarios.

The difference with Metal on llama.cpp remains stark: the native solution is up to ten times faster. The limitation does not pertain to the bandwidth of the external connection, but to the efficiency of the kernels. With memory capable of 1.8 TB/s, the driver currently manages to utilize about 33 GB/s.

TinyGPU already handles hardware connection and driver installation, but code optimization remains the primary challenge. Blackwell-class GPUs do not yet demonstrate performance consistent with specifications, primarily due to the limited maturity of the software.

However, the project is not aimed at immediately competing with established solutions. The current goal is to build a complete technical foundation that includes drivers, memory management, and compilation. Once this infrastructure is established, Tiny Corp will be able to address optimizations.

The result is still significant: for the first time on Macs with Apple Silicon, it is possible to use NVIDIA and AMD GPUs through a third-party driver, a scenario that just a few months ago lacked any concrete implementation.