Nvidia RTX 5090 runs on macOS with new custom driver from Tiny Corp

Apple and Nvidia ended their partnership many years ago, leaving Mac users without official GPU support. This fallout effectively killed CUDA on the platform and forced developers and researchers toward Apple’s own Metal framework. However, a new open-source driver from Tiny Corp has finally sorted that bit out by bringing Nvidia Blackwell hardware back into the macOS ecosystem.

The project uses a custom kernel extension called Tiny GPU. It allows external GPUs, such as the RTX 5090, to interface directly with Apple Silicon Macs over Thunderbolt 5 or USB4. This is a big technical jump on its own, as it avoids the need for virtual machines altogether. In the demo by Alex Ziskind, an RTX 5090 with 32 GB of VRAM was successfully paired with a Mac Mini M4 Pro (curr. $1399 on Amazon for the 24 GB/512 GB variant, here's our detailed review).

While the connection is stable, the current software stack is in its early stages. The driver relies on the Tiny Grad compiler rather than native Metal or CUDA optimizations. This creates a performance gap during heavy compute tasks. When running the Llama 3.1 8B model, the setup hit roughly 7.48 tokens per second. While this is a big win for compatibility, it is still slower than native Llama CPP on Metal, Alex says, which is nearly ten times faster on equivalent hardware.

Regardless, the real value in this project is its potential for future optimization. The current bottleneck is not the Thunderbolt 5 cable, which handles the model weight transfer efficiently, but the efficiency of the autogenerated kernels. For simple chat interfaces, the Blackwell setup is snappy, giving time-to-first-token speeds that are three to four times faster than native Metal solutions.

The installation process includes the approval of a system extension and running a Docker-based compiler pipeline. Clearly it isn’t a replacement (yet) for streamlined Metal workflows, but still, it is the first functional one in years.