RTX 5090 and M4 MacBook Air: Can It Game?

Original Article ↗ Hacker News Discussion ↗

eGPU passthrough on Apple Silicon

  • Thread is impressed by the technical depth: custom DriverKit PCI driver, QEMU patches, DMA handling, and making an RTX 5090 work through Thunderbolt into a Linux VM on an M4 Mac.
  • Multiple comments clarify this is not macOS using the eGPU directly; the GPU is passed through to a guest OS, with macOS just acting as host.
  • Discussion notes Apple’s strict PCI entitlements and that this project relies on special permissions; some doubt Apple will broadly approve a generic “VFIO-like” driver that effectively exposes raw PCI to user space.
  • Others point out Apple already ships internal frameworks for GPU/PCI passthrough and paravirtualized graphics, but much of it is private or incomplete for public use.

AI inference performance

  • The most practically exciting result for many is LLM inference: massive speedup in prompt “prefill” on the 5090 vs Apple GPUs, while token generation on Apple Silicon is already strong due to high memory bandwidth.
  • Comments explain: prefill is compute-bound (matrix math), generation is bandwidth-bound; dedicated GPUs win hard on compute.
  • M5 gains from added tensor cores, but high-end Nvidia remains “another league,” especially for long-context TTFT.
  • Some think future Apple chips (e.g., M6) may add dedicated silicon to close the prefill gap, making eGPUs less relevant.

Mac gaming prospects

  • People enjoy the “mad science” of getting modern Windows games running, but many emphasize that Mac gaming’s main issue is compatibility, not raw GPU performance.
  • eGPU + VM is seen as too complex for most gamers; a regular PC with native PCIe and Windows remains simpler and more reliable.
  • Some lament Apple’s poor OpenGL support and Rosetta’s eventual end, worrying about long-term game preservation on macOS.

Apple’s platform strategy and missed opportunities

  • Strong criticism of Apple’s hostility to third-party GPUs (especially Nvidia) and discontinuation of the Mac Pro; seen as ceding workstation and server markets to Windows/Linux.
  • Others argue the Mac Pro market is tiny and modular workstations are economically unattractive compared to laptops and integrated designs.

LLMs as tools and their limits

  • Several comments debate LLM usefulness: they’re great as an initial “gut check” and code assistant, but unreliable on up-to-date or niche hardware topics.
  • Experiences shared of models confidently giving outdated or wrong hardware/software info, even after correction, reinforcing that they’re helpful tools but not authoritative sources.