External GPUs Get Apple Approval
Tiny Corp says Apple approved AMD/NVIDIA eGPU drivers over Thunderbolt/USB4 for Macs, enabling cheaper local inference by plugging in consumer‑grade GPUs (e.g., a $900 3090 versus a $5k Mac Studio). The move could materially lower costs for on‑prem and edge ML experiments that need more accel than Apple Silicon alone. (x.com)
Tinygrad’s account on X posted that Apple “finally approved our driver” for AMD and NVIDIA eGPUs on April 1, 2026. The TinyGPU documentation lists minimum requirements as macOS 12.1+, a USB4/Thunderbolt port, and supported cards (AMD RDNA3+ or NVIDIA Ampere+). Installation in the repo is automated via a curl script that downloads TinyGPU.app and prompts a system toggle for a driver extension in System Settings. Reporting and the project notes indicate TinyCorp implemented Python userspace drivers that expose external cards as raw compute accelerators rather than as macOS display devices. Demos used an ADT‑Link Thunderbolt→PCIe adapter and cited roughly 40 Gbps (≈5 GB/s) bidirectional bandwidth, a figure positioned as adequate for inference workloads but constrained compared with native PCIe slots. Apple’s Mac Studio starts at $1,999 for current configurations, while eBay listings for used RTX 3090 cards in early April 2026 showed prices commonly between about $700 and $1,400, illustrating the hardware cost delta for local acceleration. TinyCorp’s coverage and reporting say the team intends to ship a purpose-built eGPU board with integrated power management and reset controls in Q2 2026, and the tinygrad repo already exposes setup scripts for early test deployments.
