Apple Approves Tiny Corp Driver Enabling NVIDIA eGPUs on Arm Macs for LLM Inference

In a surprising move, Apple has approved a driver from Tiny Corp that allows NVIDIA external GPUs to work with Apple's Arm-based Macs — no longer requiring users to disable System Integrity Protection (SIP).

What Happened

Tiny Corp, the company behind the tinygrad ML framework, announced that Apple is allowing their driver to be signed, meaning:

• No SIP disable needed — previous workaround required compromising Mac security
• Driver extension approved — installable through System Settings
• Docker-based NVIDIA compiler — uses containerized compilation for NV support
• Native AMD support — RDNA3+ GPUs work with HIP compiler on macOS

The Technology

TinyGPU

TinyGPU is an app that enables using AMD and NVIDIA GPUs on macOS over USB4/Thunderbolt:

• Supported NVIDIA: Ampere and newer (RTX 30/40/50 series)
• Supported AMD: RDNA3 and newer
• Requires: macOS 12.1+, USB4/Thunderbolt port

How It Works

1. Connect external GPU via Thunderbolt
2. Install TinyGPU driver extension (Apple-signed)
3. For NVIDIA: compile using Docker-based NVCC
4. For AMD: use HIP compiler on macOS
5. Run LLM inference via tinygrad

Why This Matters

For Mac Users

• NVIDIA GPUs on Mac — the long-requested capability for AI/ML workloads
• No compromise on security — SIP stays enabled
• LLM inference — run large language models on external NVIDIA GPUs

For Apple

• Signals a policy shift — Apple is opening up to third-party GPU drivers
• AI workloads are driving demand for NVIDIA hardware on Macs
• Could be a stepping stone to more open GPU support

For the Industry

• Tiny Corp gains legitimacy with Apple's approval
• tinygrad becomes more viable as a cross-platform ML framework
• Blurs the line between Mac and PC GPU ecosystems

Limitations

• Requires Docker Desktop for NVIDIA compilation (not plug-and-play)
• Designed for LLM inference, not gaming
• Apple's own Silicon still the primary GPU for most Mac tasks
• Performance depends on Thunderbolt bandwidth