Top AI Product

Apple hasn’t allowed third-party GPU drivers on Apple Silicon. Not NVIDIA, not AMD, not anyone. When Apple dropped Intel chips in late 2020, eGPU support died with them. Six years of silence.

Then on March 31, 2026, George Hotz’s Tiny Corp announced that Apple officially signed and approved their TinyGPU driver extension. NVIDIA RTX 30/40/50 series and AMD RDNA3+ cards can now run on any Mac with a Thunderbolt or USB4 port. No SIP bypass. No kernel hacks. A standard toggle in System Settings.

This is the first Apple-approved third-party discrete GPU driver in the Apple Silicon era. That alone is historic.

What TinyGPU Actually Does (and Doesn’t Do)

Let’s be clear about scope. TinyGPU is not bringing gaming GPUs to Mac. You won’t be running Cyberpunk at 4K on your MacBook Pro. This is a compute driver built on top of tinygrad, Tiny Corp’s neural network framework (27,000+ GitHub stars). It enables AI inference workloads — running LLMs locally, specifically.

The driver uses Apple’s DriverKit framework, which means it runs in userspace rather than as a kernel extension. That’s likely why Apple was willing to sign it — DriverKit extensions can’t crash the system. Smart engineering choice by Tiny Corp.

Hardware setup is straightforward but DIY: you need an ADT-UT3G USB4-to-PCIe adapter (~$110-170 depending on where you buy it), a GPU with its own power supply, and any Mac running macOS 12.1 or later. Install the app, toggle the driver on, install the compiler (AMD’s compiler runs natively; NVIDIA requires Docker), and you’re running inference.

18.5 Tokens Per Second on a Mac Mini

The real question: does the performance justify the setup?

Tiny Corp’s own benchmark shows a Mac mini M4 connected to a Radeon RX 7900 XTX running Qwen 3.5 27B at 18.5 tokens per second. That’s a 27-billion-parameter model at interactive speed. For context, Apple’s integrated GPU on the M4 can’t touch models that large without aggressive quantization and significant speed penalties.

The bottleneck is Thunderbolt bandwidth. Even Thunderbolt 4 tops out at PCIe 4.0 x4 speeds — roughly 4 GB/s each way versus the ~60 GB/s you’d get from a native PCIe 5.0 x16 slot. So no, this won’t match a desktop workstation. But for Mac users who want to run 20B+ parameter models locally without buying a separate PC, the tradeoff is real.

The Long Road to Apple’s Signature

This didn’t happen overnight. Tiny Corp has been grinding on this for over a year:

• May 2025: Demonstrated the “world’s first” AMD GPU driven over USB3 from an Apple Silicon Mac, using a reflashed adapter and custom userspace drivers
• October 2025: Got NVIDIA RTX GPUs running on a MacBook Pro M3 Max over USB4 — first-ever NVIDIA discrete graphics on an ARM Mac
• March 31, 2026: Apple signs the DriverKit extension. No more SIP workarounds needed

The early versions required disabling System Integrity Protection, which is a non-starter for most users. Getting Apple to officially approve the driver is what turns this from a hacker project into a real product.

The Catch: Tinygrad Only

Here’s the limitation that matters most. TinyGPU only works with tinygrad. Standard CUDA, Vulkan, PyTorch — none of that runs through this driver. If your ML workflow lives in the PyTorch/CUDA ecosystem (which, let’s be honest, most do), TinyGPU doesn’t help you today.

The Hacker News crowd was blunt about this. One commenter called it “a good technical project, but honestly useless in like 90% of scenarios.” That’s harsh but not wrong for the current state. The value proposition is narrow: Mac users who are willing to adopt tinygrad for local LLM inference.

Compare this to what Mac users have been doing instead — running quantized models through llama.cpp on Apple’s Metal GPU, or just SSH-ing into a cloud GPU. Both work. Neither requires buying additional hardware. TinyGPU’s pitch is that you get significantly more compute power locally, but you’re locked into one framework to use it.

Why This Matters Beyond the Benchmarks

The bigger story isn’t the performance numbers. It’s that Apple signed the driver at all.

Apple has kept NVIDIA off Mac for over a decade. The feud between the two companies is legendary in tech circles. For Apple to approve a driver that enables NVIDIA hardware on their platform — even through a third-party intermediary, even limited to compute workloads — signals something shifted.

Maybe it’s pragmatism. The AI development community overwhelmingly uses Mac laptops, and those developers need GPU compute. Maybe Apple decided that a compute-only driver with no display output isn’t a competitive threat to their own silicon. Whatever the reason, the door is now cracked open.

George Hotz and Tiny Corp turned what started as a USB hack into an Apple-sanctioned product. Whether TinyGPU becomes a mainstream tool depends on tinygrad’s ecosystem growth. But the precedent? That’s already set.

You Might Also Like

• Ollama mlx on Apple Silicon 1810 Tokens sec Prefill and the end of Llama cpp on mac
• George Hotz Wants to Sell you a 12000 ai Supercomputer and 221 Hacker News Comments Cant Stop Arguing About it
• Cern Hls4ml how the Worlds Largest Physics lab Burns Tiny ai Models Directly Into Silicon Chips
• 700 Github Stars in a Week Apfel Exposes the Free llm Apple Locked Behind Siri
• Popair Just Gave my mac a Superpower Upgrade