How I Test Macs

Two Mac minis stacked on two Mac Studios next to a MacBook Air on Adam's test bench

Every Mac on my Mac Tests page runs the same tests on my desk, and that page pulls the results live from my test sheet. Here’s exactly how I run them, so you can see what the numbers mean and even run some of them yourself.

The rules every Mac follows

  • Clean start. Fresh restart, nothing else open, background apps quit. It sits a few minutes before the first test.
  • Cool start for the heavy stuff. Before Lightroom and the long tests, the Mac is shut down or idle for at least 15 minutes. Hot runs came in 10-20% slower in my testing, so the number you see is the cool run.
  • Warm-up first. Every app gets opened once before I log anything. First launches skew results, since After Effects, Resolve, Blender and Cyberpunk all set things up the first time they run.
  • Laptops stay plugged in.
  • One Mac per config. I test the Mac I have, not a stack of them. Two Macs with the same chip can land a few percent apart, so don’t read too much into tiny gaps.
  • More than one run on the big numbers. If two runs are more than 5% apart, I run a third and use the middle one.

Real-world tests

Final Cut Pro

  • 3 minute 4K timeline exported with the Social Media 4K setting.
  • 30 minute 4K timeline exported with the Apple Devices 4K setting. One export at a time, with a break in between.
  • Magnetic Mask on an 18 second clip.
  • Stacked 6K clips: 6K open gate footage from my Lumix S5IIX. I stack as many clips as will play for 30 seconds with no dropped frames (with Final Cut’s dropped frame warning on). Two tries, and I log the lower one.
  • Export times come from the exported file’s timestamps, not a stopwatch.

Lightroom Classic

  • 500 RAW photos from my Sony A7 IV and Lumix S5IIX, imported fresh every time (no saved catalog), then exported at 100% JPEG quality: 500, then 200, then 50.
  • AI Denoise on 1 RAW photo and on 50.
  • Newer Macs run Lightroom Classic 16 with no edits on the photos. Some older Macs on the page ran Lightroom 15 with exposure edits, so compare those carefully.

Puget Bench

These benchmarks are made by Puget Systems, and they run real projects inside the actual apps. Standard preset on every Mac.

  • Puget Bench for DaVinci Resolve (Resolve Studio)
  • Puget Bench for Premiere Pro
  • Puget Bench for After Effects (run twice, and I log the second because the first launch skews it)
  • Puget Bench for Photoshop
  • Puget Bench for Lightroom Classic (cool start)

If a Puget Bench for DaVinci Resolve run glitches (Super Scale under about 5 fps), I rerun it and drop the bad one. App and benchmark versions are in the notes for each Mac.

Xcode

  • My scripted version of the open-source XcodeBenchmark project: a clean build, run three times. The page shows the average of the three.

Blender 5.2

  • Classroom and Barbershop scenes on the GPU, three runs each. Classroom on the CPU, one run.
  • Animation test: the Red Autumn Forest demo file from blender.org, rendered in Cycles on the GPU at 128 samples, frames 1 to 25. This is the long one that shows if a Mac slows down as it heats up. Timed from the output files.

Logic Pro

  • A 1 minute all-MIDI project bounced to uncompressed AIFF, at 64 tracks and at 512 tracks.
  • Track count uses the Gearspace Logic Pro multicore benchmark, a project music producers have used to compare Macs for years. I keep enabling tracks until playback skips or stutters when I start it or click around the timeline. The last count that plays clean is the number.
  • A heavier test with a real song, sampled instruments and plugins is coming.

Local AI (LM Studio)

  • Gemma 3 4B, with a new chat for each prompt. I log time to first token, tokens per second and total tokens.
  • Writing speed: “Write a detailed, 50-paragraph essay explaining the historical significance of the Roman Empire on modern law, philosophy, and architecture. Be verbose and do not stop until you run out of tokens.”
  • Reading speed: a long section of Wikipedia text about Apple’s environmental work, pasted in whole, with a request for a two-sentence summary. Same file every time.
  • Bigger models (Qwen 3.5 9B and Qwen3.8 27B) are being tested now and will show up on the page.

Image generation

  • Diffusion Bee with its stock model, and the prompt “a cat working on a laptop with nachos on the side.”

Gaming: Cyberpunk 2077

  • The game’s built-in benchmark. High preset, texture quality High, VSync off, frame cap off, ray tracing off.
  • Two passes: native, and MetalFX Quality (sharpness 0.50, no frame generation).
  • One throwaway run first on each Mac, since the first run compiles shaders.
  • Mac minis and desktops: 1080p fullscreen on the same external monitor. Laptops: their own screen at the default fullscreen resolution, 50% brightness, plugged in. On the 14″ MacBook Pro that’s 1800×1125, within 2% of 1080p’s pixel count.

Synthetic benchmarks

  • Geekbench 7: single core, multi core, OpenCL and Metal.
  • Geekbench AI: CPU, GPU and Neural Engine.
  • Blackmagic RAW Speed Test: CPU and Metal.
  • Internal SSD speed: Blackmagic Disk Speed Test and AJA System Test Lite (16GB, 4K).
  • Cinebench 2026 doesn’t run reliably on every Mac yet, so it only shows up where I got clean runs.
  • Older tests (Geekbench 6, Cinebench 2024, Xcode 26 and older Blender runs) are hidden by default since newer versions replaced them.

Why I test this way

This testing method combines scientific reasoning with real world use cases to give you a variety of information on the differences between the models.

See the results

Pick the Macs you care about on the Mac Tests page. Still stuck between two? Tell me what you do and your budget and I’ll give you my pick.