M4 Mac mini reference

M4 Mac mini cross-OS benchmarks

Qualification status

The benchmark harness uses the same portable workloads on macOS and Aurora, three independent timed runs, and reports the median with the minimum and maximum. Tool and model versions are pinned, inputs are hashed, and each run records OS, kernel, compiler, CPU topology, thermal state, and available power telemetry.

The current comparison combines the committed 2026-11-07 macOS baseline with a 2026-11-13 Fedora run-59 pass. NixOS remains unmeasured: its one-shot was cleared after the mandatory gate refused admission. Missing cells are not estimates.

The J773g under test is a Mac16,10 with an Apple M4 (10 CPU cores: 4 performance and 6 efficiency; 10 GPU cores), 16 GB unified memory, and a 256 GB internal Apple SSD. A planned simultaneous A/B with a second macOS Mini is same-SKU evidence only after those CPU, memory, and storage characteristics are confirmed to match.

Workload macOS 2026-11-07 Fedora run 59 Fedora/macOS
STREAM Triad 93.97 [93.62, 94.35] GB/s 92.83 [90.51, 95.33] GB/s 0.988x
7-Zip, one thread 11,606 [11,406, 11,725] MIPS 4,317 [4,316, 4,323] MIPS 0.372x
7-Zip, ten threads 80,924 [79,644, 81,548] MIPS 41,681 [41,562, 41,947] MIPS 0.515x
Sequential read QD1 2,610.92 [2,608.54, 2,612.81] MB/s 336.53 [307.53, 1,160.75] MB/s 0.129x
Random 4 KiB read QD32 163,132 [161,919, 163,750] IOPS 8,735 [8,678, 8,948] IOPS 0.054x

STREAM is within 1.2% of macOS. The large 7-Zip and storage gaps are candidate findings: tool versions and fio scratch sizes still need exact matching before they become qualified invariants. NixOS, GPU, ML, boot-time, and power cells remain open.

ML next run

The bounded 16 GB matrix uses Qwen2.5 0.5B/1.5B/3B, Llama 3.2 1B/3B, and a Phi-mini-class model in Q4_K_M, plus Q8_0 and F16 for the smallest model. llama.cpp CPU and Vulkan measure pp512 and tg128 separately; Linux MLX repeats the same models where supported. Every model and prompt is SHA-256 recorded.

Microbenchmarks cover 512/1024/2048/4096 square GEMM in fp32/fp16/bf16 where supported, batch-1 GEMV, batched small tensors, and a ResNet-like conv2d. MobileNetV2, ResNet-50, and whisper tiny/base use fixed hashed inputs. The macOS counterparts are MLX/Metal, Core ML CPU/ANE/all policies, PyTorch MPS, and tensorflow-metal. omarchy-ANE results are accepted only with proven J773g device placement.

Candidate performance investigations

These are hypotheses to test, not measured NixOS regressions:

  1. CPU single-thread performance: verify CLPC/cpufreq residency and sustained performance-core frequency under 7-Zip and sysbench.
  2. CPU multi-thread performance: compare scheduler placement, capacity data, and thermal throttling across the ten cores.
  3. GPU and LLM throughput: compare the same llama.cpp revision, model, quantization, prompt, and token counts using Metal and Vulkan; glmark2 and a macOS-native graphics test are not treated as directly comparable.
  4. NVMe: compare scheduler, queue depth, direct-I/O support, and filesystem mount options with identical fio job files and scratch-file sizes.
  5. Boot time and idle power: separate firmware/loader time from userspace and capture Linux hwmon/SMC telemetry only when a supported sensor path exists.

Any verified gap greater than 20% becomes a candidate invariant or tracked performance ticket with the raw three-run evidence attached.

Reproduction contract

Use the repository’s portable runner and exact pinned inputs. Run macOS first, then select NixOS only through the one-shot trial gate, inhibit idle sleep for the session, and return to the saved Fedora baseline with an empty one-shot selector and no trial files. Never compare best-of-run values: report the median, range, and relative spread from all three runs.

Part of Techmore