A recent project explored connecting a high-end NVIDIA RTX 5090 GPU to an M4 MacBook Air via a Thunderbolt eGPU setup. While macOS lacks native drivers for NVIDIA GPUs on Apple Silicon, the author successfully passed the GPU through to a Linux virtual machine running on the Mac. This approach leverages Linux's support for NVIDIA cards and Thunderbolt tunneling to enable the GPU for tasks like AI inference, though performance in the tinygrad AI stack was found to be significantly slower than native Metal inference. AI
Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →
IMPACT Demonstrates a workaround for using powerful external GPUs with Macs for AI tasks, though current performance limitations highlight ongoing challenges in hardware integration.
RANK_REASON The article details a technical experiment and benchmark of using external hardware with a specific software stack, rather than a product release or major industry development. [lever_c_demoted from research: ic=1 ai=0.7]