I saw a message these days and was stunned for a moment. I only found out that Apple can have an external graphics card, and I never expected that Apple would also have a day of great kindness. Mac can now connect external AMD/NVIDIA graphics cards. It's not a virtual machine, it's not installing Windows, it's in macOS, take a Thunderbolt cable, plug in RTX 4090, and run AI inference. This matter was not officially announced by Apple, but it was actually released. Wait, hasn't Apple already cut off the GPU? Yes, cut it off. In 2020, Apple Silicon was released, and Apple quietly closed the door to GPUs. The official logic is that a unified memory architecture does not require an external graphics card, and the built-in one is sufficient. Then the era of AI arrived. The TinyCorp team is the group of people who developed the Tinygrad framework and managed to implement the graphics card driver in user mode. Without touching the kernel, there is no need to turn off SIP. Apple has no reason to block this thing, and even if it is blocked, it cannot be blocked. This Nima is the standard operation for smart people to find gaps. Think about it, Apple did not 'approve' this matter. It's just, there's no way to say no. The community bypassed all the doors that could be closed and entered through the window. The current result is that a Mac Studio M3 Ultra, with an external RTX 4090, can run a 7B model from 35 tokens/s to 120+directly. Mac manages memory, Nvidia manages computing power, and each side does what they are good at. I never imagined that this combination could be implemented on macOS six months ago. The pattern of AI computing power has always been driven by user needs, not determined by manufacturers' roadmaps. Apple believes that unified memory is the end point, but developers tell it with actual scores that the memory is large enough but not fast enough. If there are vacancies in the market, there will be people filling them in. It's not that Apple lost, it's that no one won this game, technology is winning. I haven't started running it myself yet, but seeing the screenshot of how pip install tinygrad can be used, I'm really excited. The group of people who were waiting for GPUs five years ago have received a better answer today. How to connect? Hardware List: Any M1/M2/M3/M4/M5 chip Graphics card: NVIDIA RTX 4090/AMD RX 7900 XT Adapter: ADT-UT3G (TB to PCIe x16) Power supply: graphics card independent power supply (ATX or eGPU dock) pip install tinygrad python -c "from tinygrad import Device; print(Device.DEFAULT)" python -c "from tinygrad import Tensor; print(Tensor.ones(4,4).realize())" A single thread connects the Mac and GPU ecosystems together. This thing is pretty good. What can I do after connecting it? Mac memory is large enough, but inference speed is limited by the built-in GPU. After external RTX 4090: Core value: Mac is responsible for memory and orchestration, while external GPU is responsible for brute force computing. Both take advantage of each other's strengths. Mac Studio M3 Ultra 256GB+external RTX 4090. Mac loads a billion parameter model into unified memory and hands over inference to 4090. Memory belongs to Apple, computing power belongs to Nvidia - this is the optimal solution. The memory advantage of Mac and the computing power advantage of NVIDIA are finally no longer mutually exclusive. It was unimaginable six months ago for a device to possess the top-level capabilities of two ecosystems simultaneously. Method transferred from official account, Just Jason