How fast is a macOS VM, and how small could it be?
Apple Silicon vs Intel & macOS VM Performance
- Multiple comments report Apple Silicon Airs “stomping” comparable Intel MacBooks on general workloads, with big wins in performance and thermals (silent vs “hairdryer” fans).
- A Geekbench comparison is cited showing ~2.5× CPU score improvement for a newer MacBook over a 2020 Intel Air.
- Some users still find specific tools (e.g., iOS Safari via Xcode simulator on Intel) sluggish and are hopeful Apple Silicon will resolve this.
Memory Usage, Ballooning & Stability
- Discussion around a macOS VM using less RAM than allocated: some assumed no ballooning; others link to Apple’s virtualization docs showing a balloon device exists (though not for macOS guests).
- Several argue macOS manages memory well under desktop multitasking; others counter with examples of system apps leaking memory and causing OOM, especially in recent macOS versions.
- There’s disagreement whether these represent widespread systemic issues or isolated bugs.
Core Count, RAM & Build/Compute Workloads
- Observations that macOS VM memory usage shrinks as both RAM and vCPUs are reduced; most attribute this to the OS adapting cache/buffer usage, not per-CPU overhead.
- Multiple comments stress that physical RAM should scale with hardware threads, especially for parallel builds (Chrome, Vivado, FlashAttention) that can require gigabytes per thread and force lowering
-jwhen RAM is constrained.
Containers & VM Tooling on macOS
- Experiences with colima/docker range from “usable but painful/inefficient” to acceptable; limactl and podman are also mentioned.
- Apple’s
containerCLI receives praise for speed and integration, but lacks Docker CLI compatibility and Compose support, which some find limiting or “typical Apple.” - OrbStack is widely praised for performance and energy efficiency via a single-VM-plus-containers model, but its closed-source, per-seat licensing is a blocker for some.
Security, Signing & Notarization Automation
- One user’s 200-line Bash script for unattended signing/notarization is criticized as overcomplicated; others note keychain unlocking and avoiding GUI prompts add real complexity in CI contexts.
- There’s disagreement on how minimal such automation can be, but consensus that non-interactive setups are fragile.
GPU & ML in VMs/Containers
- A new M5 Air user struggles to get PyTorch with GPU acceleration inside isolated environments; virtio-gpu exposes graphics but not compute.
- Some mention emerging approaches (e.g., Docker Model Runner, podman+libkrun) but characterize Vulkan-based workarounds as experimental and slow.
- At least one person reports getting PyTorch’s MPS backend working inside a macOS VM using a Tart image, though only minimal tests were run.
Minimal macOS / Darwin / XNU
- Curiosity about stripping macOS to a bare, possibly non-GUI system leads to discussion of Darwin/XNU layers and the long-running PureDarwin project (currently in need of new leadership).
- Apple’s XNU source has moved to a different GitHub org and is still updated, but rebuilding/booting it is seen as a “fun but costly” side project.
- Instructions are shared for booting modern Macs directly into a root shell (single-user mode) via reduced security settings and
nvramflags.
Networking, Tailscale & Identity in Containers
- An example project shows using Apple’s container tooling plus Tailscale SSH so each container gets its own Tailnet IP/identity, separate from the host.
- This is positioned as useful for letting others in the Tailnet access a containerized app without exposing the whole Mac or resorting to port forwarding.
macOS on PCs & Remote Dev
- Running macOS on PC is described as possible via QEMU or “Hackintosh,” but with missing features (e.g., hardware-accelerated graphics, Secure Enclave–dependent services like Apple Pay, iMessage, FaceTime).
- One brief question about enrolling a macOS VM into Intune gets the answer that it’s likely only straightforward as BYOD; company-owned enrollment ties into Apple Business Manager.