About
MacStadium offers a wide selection of Bare Metal Mac hosts for a variety of applications.When selecting host machine varieties, most MacStadium customers seek to strike a balance between the volume of computation that they need to do, the speed at which a given job (like a CI build) should be completed, and the cost involved.
Overview
Benchmarking When evaluating Macs, the goal is to determine the optimal performance of a given piece of Mac hardware, including CPU performance, GPU performance, and real-world performance for CI/CD applications. To achieve this, refer to these synthetic benchmarks:- XcodeBenchmark
- Geekbench 6 Multi-core
- Geekbench 6 Compute
- Cinebench R24
Not every benchmark compares across toolchains. Geekbench 6 scores stay stable when macOS and Xcode change, so they can be compared between tables. Xcode build times and Cinebench scores cannot.For the same M2.M machine, the Xcode build time is 106.259s in the full results table and 190.33s in the standard models table. That 79% difference comes entirely from the change in macOS and Xcode version, not the hardware. Its Geekbench 6 CPU score moved 1.1% across the same change.Cinebench behaves the same way: Maxon reset the scoring baseline between releases, so R23, R24 and 2026 scores are on different scales and cannot be compared to each other.Always compare within a table, and check the toolchain stated for that table before reading across rows.
Benchmarking Baseline
This machine is used as the baseline when referencing multipliers. In other words, a machine with a multiplier of 1.4x is 40% more performant than the baseline machine for the specified test. Multipliers are shown below the raw score in each cell.Current standard models
This table includes Standard models currently available as Bare Metal from MacStadium.
† Measured on macOS 27 with Xcode 27 and Cinebench 2026. These Xcode and Cinebench figures are not comparable to the other rows, so no multiplier is shown. Geekbench figures are comparable across all rows. See macOS 27 test round.
‡ M6.M ships with a 1TB SSD. The unit tested had 2TB. See the note on M6.M storage in macOS 27 test round.
macOS 27 test round (M5 Pro and M6)
Tested on macOS 27.0 (build 26A428) with Xcode 27.0 (27A266a) and Geekbench 6.7.1. Best of three runs per machine, with ten minutes of low CPU utilization between runs. The M5 Pro and M6 results are listed in Current standard models. Their Geekbench columns use the same M2.M baseline as the rest of that table. Their Xcode and Cinebench figures establish the baseline for future macOS 27 rounds and aren’t comparable to the other rows, for the reasons in the note at the top of this page. The M5 Pro has 18 cores (6 Super, 12 performance) in the M5Pro.XL and 15 cores (5 Super, 10 performance) in the M5Pro.L. The M6 has 12 cores (2 Super, 4 performance, 6 efficiency). The results diverge depending on how much work each benchmark asks for, and that divergence is the most useful thing in this table. The M6 leads both M5 Pro configurations on Geekbench single-core by roughly 9% (4743 against 4353). It then loses the Xcode build by 29% (114.76s against 81.89s), and loses sustained multi-threaded rendering in Cinebench by 67% (5698 against 9526). The M5 Pro carries 18 performance-class cores with no efficiency cores, while the M6 pairs 6 performance-class cores with 6 efficiency ones, so the longer and more parallel the workload, the further ahead the M5 Pro gets. Sizing a CI fleet on single-core scores alone would point at the wrong machine. Memory capacity made no measurable difference to any score. Both M6 machines were measured under identical conditions, and the 32GB unit finished fractionally ahead of the 16GB unit on every benchmark, by 1.5% at the widest and by 0.23% on Cinebench, despite carrying twice the RAM. The gap between the two M5 Pro machines tracks their core counts rather than their memory. Size memory for the workload, not for the benchmark.On the Cinebench figures. Geekbench and XcodeBenchmark were each collected in a single session per machine. The Cinebench runs were not: the first run was taken on 25 September 2026 and the remaining runs the following Monday. Machine configuration was verified unchanged between the two, and the gap between runs therefore exceeded the ten minute cooldown the method requires rather than falling short of it. The M6.M Cinebench figures are an exception: that unit was re-measured in full on 29 September 2026, in a single session.
On measurement conditions. Every figure here was collected with a console session logged in on the machine and no screen sharing viewer attached. Both conditions change the result measurably on macOS 27. A machine left at the login window with no console session was observed running
audiomxd at roughly 70% of a core and configd at roughly 42%, continuously, across all four units tested; logging in returned both to idle. An attached screen sharing viewer costs roughly 10% on multi-threaded scores. Reproducing these numbers requires the same conditions.On the M6.M storage. The M6.M SKU ships with a 1TB SSD. The unit tested here had 2TB, so its benchmark figures come from that configuration, and the table lists the shipping 1TB. This has no bearing on the Geekbench or Cinebench figures, and any effect on the Xcode build time would be small, but a shipping M6.M could differ slightly on disk-bound work.
On the Xcode 27 figures. XcodeBenchmark does not build under Xcode 27 as published: the project targets iOS 14.5 and Xcode 27 requires a minimum deployment target of 15.0. These runs apply pull request #675, which raises the deployment target to 15.0 and changes no source code. Results are therefore not comparable to the public XcodeBenchmark result tables.