A third-party driver enabling Nvidia eGPUs on Apple Silicon Macs has passed Apple's notarization process. Developed by Sapling, the 'Miracle' driver is a DriverKit system extension, a user-space framework Apple promotes over deprecated kernel extensions. This technical distinction is important: Apple reviewed and approved the code, integrating it into macOS security. It's not an unsupported hack.

The driver supports CUDA compute workloads over Thunderbolt, targeting machine learning training, inference, and other headless tasks. It does not handle display output. For data engineers and ML practitioners, this addresses a specific gap: the ability to run CUDA-dependent workflows, like PyTorch or TensorFlow training jobs, on a local Mac without maintaining a separate Linux workstation.

Architecturally, this introduces a translation layer between macOS and the Nvidia GPU. Performance will hinge on Thunderbolt bandwidth constraints and the efficiency of this abstraction. While a PCIe 4.0 x16 slot offers 256 Gbps, Thunderbolt 4 provides 40 Gbps. For batch-oriented training where data transfer is intermittent, this may be acceptable. For bandwidth-sensitive operations, it will be a bottleneck.

The move is a pragmatic engineering concession. Apple's integrated M-series chips are optimized for performance per watt, but CUDA remains the industry standard for model development. By allowing this third-party bridge, Apple retains developers in its hardware ecosystem without diverting its own engineering resources to native Nvidia support. It's a calculated trade-off, maintaining the official Apple Silicon narrative while letting the community solve a specific integration problem.

Source: Webpronews