 |
| YouTuber Alex Ziskind paired an RTX 5090 (32 GB VRAM) with a Mac Mini M4 Pro (edited) |
Remember the good old days when you could slap an Nvidia graphics card into a Mac Pro and call it a day? That ended abruptly years ago when Apple and Nvidia’s partnership went up in flames. Since then, Mac users have been stuck watching from the sidelines as CUDA-powered AI and compute workflows became the industry standard. But something unexpected just happened: a tiny hardware company called Tiny Corp has quietly released an open-source driver that brings Nvidia’s latest Blackwell GPUs – yes, including the RTX 5090 – back to macOS. And no, it doesn’t involve running Windows in a VM.
The Breakup That Left Mac Users in the Cold
Let’s rewind. Apple and Nvidia’s relationship soured around the time of macOS Mojave (2018). Driver support vanished, CUDA was effectively banned from the platform, and Apple went all-in on its own Metal framework. For creative pros and researchers who relied on Nvidia’s compute dominance, that was a brutal pivot. Many simply abandoned the Mac for Linux workstations. The few who stayed had to wrestle with eGPU setups that were either broken or required hacky workarounds. For years, the consensus was clear: if you need serious GPU compute, don’t buy a Mac.
That consensus just got a lot more interesting.
Enter Tiny Corp and the “Tiny GPU” Kernel Extension
Tiny Corp, best known for its own AI inference hardware, decided to tackle the problem head-on. The result is a custom kernel extension called Tiny GPU – a low-level driver that lets external Nvidia GPUs talk directly to Apple Silicon Macs over Thunderbolt 5 or USB4. No virtualization layer, no PCIe tunneling tricks. Just a direct line of communication.
In a live demonstration by YouTuber and developer Alex Ziskind, an RTX 5090 with 32 GB of VRAM was successfully paired with a Mac Mini M4 Pro. If you want to check out the current pricing on that Mac Mini, here’s the Amazon listing for the 24 GB/512 GB variant – it’s the same model used in the demo. Ziskind’s detailed review of the Mac Mini M4 Pro is also worth a read if you’re considering this setup.
The video (embedded below) shows the RTX 5090 being recognized by macOS without the usual “unsupported GPU” tantrum. That alone is a minor miracle.
Performance: It Works, But Don’t Throw Away Your Linux Rig Yet
Before you start ordering a Thunderbolt 5 enclosure, let’s talk about real-world numbers. The current driver stack is very early software – it relies on a Tiny Grad compiler rather than native Metal or CUDA optimizations. That means performance is… mixed.
Running the Llama 3.1 8B model, Ziskind measured roughly 7.48 tokens per second. For context, native Llama CPP on Metal (using Apple’s own GPU cores) is nearly ten times faster on equivalent hardware. So if you’re doing heavy batch inference or training, this isn’t going to replace a dedicated Linux server.
But here’s where it gets interesting. For interactive workloads – like chatting with a local LLM – the time-to-first-token is reportedly three to four times faster than native Metal solutions. Why? Because the Thunderbolt 5 cable can shovel the entire model weight into the RTX 5090’s VRAM with very low latency. Once the model is loaded, the bottleneck isn’t the connection; it’s the efficiency of the autogenerated kernels. And that’s something that can be optimized over time.
Alex notes that for simple chat interfaces, the Blackwell setup feels “snappy.” The delay between hitting Enter and seeing the first word appear is noticeably shorter than on Metal. That’s a big deal for anyone running local AI assistants or real-time generation tasks.
The Real Value Is What Comes Next
Let’s be honest: 7.5 tokens per second on an RTX 5090 is embarrassingly low. That card can do ten times that on a native Linux system. But that’s not the point. The point is that the door is now open.
The current bottleneck is purely software. The Tiny Grad compiler is generating kernels on the fly without any architecture-specific tuning. Once the developers add support for CUDA core optimizations or start mapping compute graphs more intelligently to Metal’s memory model, those numbers will climb. And because the driver is open source, the entire community can pitch in.
Thunderbolt 5 is also a huge enabler here. With bandwidth up to 80 Gbps (and 120 Gbps in asymmetric mode), it’s more than capable of handling model weight transfers and even some compute offloading. The demo already proves that the cable isn’t the weak link – the compiler is.
How Painful Is the Installation?
Right now, this is not a plug-and-play solution. You’ll need to:
- Install the Tiny GPU kernel extension (which requires approving a system extension in macOS Security & Privacy).
- Run a Docker-based compiler pipeline to generate the appropriate shims for your Nvidia card.
- Manually configure your LLM or compute framework to target the Tiny Grad backend.
Alex’s video walks through the entire process – you can watch it here: Alex Ziskind’s full demonstration on YouTube. It’s about 20 minutes of terminal commands and config tweaks. So no, your grandma won’t be doing this. But for developers, researchers, and tinkerers, it’s surprisingly straightforward.
Who Is This Actually For?
Three groups immediately come to mind:
- AI researchers stuck on macOS – You can now prototype on your MacBook Pro and offload heavy inference to an external Nvidia GPU without rebooting into Linux.
- Creative pros with legacy Nvidia hardware – If you have a box full of older Quadros or RTX cards, this driver gives them new life on your Mac Studio.
- Open-source enthusiasts – Because this project proves that Apple’s walled garden isn’t as airtight as it seems.
It’s not yet for gamers (no DirectX translation layer) or for production training pipelines. But give it six months of community development, and that could change.
The Bottom Line
Tiny Corp has done what many thought was impossible: resurrected official-ish Nvidia support on macOS without CUDA emulation or virtual machines. The current performance is proof-of-concept territory, but the foundation is solid. Thunderbolt 5, Apple Silicon, and a determined open-source community might just turn the Mac into a viable Nvidia compute platform again – a decade after Apple and Nvidia’s bitter divorce.
If you’ve got an M4 Pro Mac Mini lying around and an RTX 5090 gathering dust (lucky you), this is your weekend project. For everyone else, keep an eye on this driver. It’s the most exciting thing to happen to Mac GPU computing in years.
Sources: Alex Ziskind on YouTube, Tiny Corp GitHub repository.
