Apple approves driver that lets Nvidia eGPUs work with Arm Macs

Original Article ↗ Hacker News Discussion ↗

Apple has approved a third-party driver that lets Nvidia and AMD GPUs in external Thunderbolt enclosures work for compute workloads on Arm-based Macs, enabling tools like tinygrad to run LLM inference on eGPUs. Commenters are excited about new options for local AI and scientific computing, but note major limitations: no CUDA, no graphics support, and bandwidth bottlenecks compared to native PCIe, making this a niche solution versus dedicated Linux PCs or cloud GPUs. The news also reignites broader arguments over Apple’s closed driver model, eGPU practicality, and whether Apple’s platform control raises antitrust or consumer-rights concerns.

Scope of the New Driver

  • Driver enables Nvidia (and some AMD) GPUs over Thunderbolt/USB4 on Apple Silicon Macs, but only for compute.
  • No CUDA, Vulkan, nvidia-smi, or graphics/AAA gaming support; it’s tied to tinygrad’s stack.
  • Some consider the headline “misleading” because “eGPU” implies graphics; others say compute is what many care about (e.g., local LLMs).
  • Implementation appears to be a userspace driver plus tooling, with Docker used mainly to get Linux-side toolchains for Nvidia.

Performance, Bandwidth, and Hardware Constraints

  • Thunderbolt/USB4 exposes only PCIe x4, far below a typical x16 slot; TB4 ≈ 4–5 GB/s, TB5 ≈ 8 GB/s effective.
  • This is a serious bottleneck for some workloads, especially large-model streaming or graphics, but less so for many LLM inference tasks.
  • External enclosures are another cost/complexity point (PSU, cages, cables, dust, reliability).

Use Cases and Practicality

  • Enthusiasm from people with spare high-end GPUs who want local LLM inference or scientific compute on Macs.
  • Skeptics argue it’s “90% useless”: limited bandwidth, fragile support, no CUDA/Vulkan/PyTorch integration, and risk of breakage in future macOS updates.
  • Alternatives raised: just buy a cheap PC for the GPU, rent cloud GPUs, or use LAN-based “remote GPU” solutions.

Apple, Nvidia, and Platform Control

  • Debate over Apple’s historic refusal to sign Nvidia drivers and the opportunity cost (e.g., for servers or pro Macs).
  • Disagreement on whether this should trigger antitrust scrutiny; some say Apple lacks market monopoly, others argue monopolies can be defined narrowly (e.g., App Store on iOS).
  • Broader concerns about driver signing, “walled garden” policies, and whether users should be able to run any drivers/software on their hardware.

eGPU and UX Experiences

  • Mixed reports on eGPUs: some long-term stable setups, others plagued by sleep/power/connection issues.
  • Many feel traditional internal PCIe slots on Macs are effectively gone, making high-bandwidth GPU use on macOS an ongoing pain point.