Mac Mini M6: CPU Productivity Benchmarks
The M6 Mac Mini builds LLVM in 314 seconds against 505 seconds on the base M4, completing the build in 38% less time.
Timed LLVM Compilation
For a practical CPU workload I am using Timed LLVM Compilation from the Phoronix Test Suite. Building LLVM is a long, highly parallel C++ workload that tests sustained CPU performance, memory behaviour, and is indicative of the kind of workload a software engineer might actually hit.
Similarly to my Neo review, I used pts/build-llvm 1.5.0 (LLVM 16.0, Ninja). The result is elapsed time for the build step only, so lower is better. LLVM 16 would not build on this machine without a one-line patch to tools/sancov/sancov.cpp, this is because copy-list-init no longer compiles under the newer Clang and libc++. No logic changed so it's not going to move the time in any way that matters.
The M6 averaged 314 seconds across three runs (313.2s, 313.4s, and 314.0s), with 0.14% deviation. That is 44% less time than the binned M3 Pro at 565s, and 50% less time than the 64GB M1 Max at 626s.
The comparison this page is actually for is the base M4 Mac Mini, on the same recipe. It completes the build in 505 seconds, against the M6’s 314 seconds. That is 38% less time, saving 3 minutes and 11 seconds per build.
Timed LLVM Compilation
Lower is better. M6 is the average of three runs (313.5s, shown as 314s, 0.14% deviation) on pts/build-llvm 1.5.0. M3 Pro is n=1 on this bench. M1 Max and Neo are published Neo-review figures, not this sit. M4 is measured at 505s; run count and deviation are not recorded here.
Against Commodity CPUs
OpenBenchmarking publishes aggregated results for this same profile, pts/build-llvm 1.5.x with Ninja. These are public results from different machines and software configurations, so the spread on a given chip can be wide.
The current desktop flagships finish just ahead of the M6. The Ryzen 9 9950X averages 286 seconds and the Core Ultra 9 285K 287 seconds, so the M6 takes about 10% longer. The M6 lands next to the last-generation Ryzen 9 7950X3D, at 313 seconds, and a step ahead of the Core i9-14900K at 323. The 8-core Ryzen 7 9800X3D, the chip a lot of gaming PCs are built around, takes 474 seconds. Given this is the base M6 chip, this is a fascinating result.
The other Apple chips make this particularly striking. The same source puts the M3 Max at about 300s and the M4 Pro at about 399s, with AMD's Ryzen AI Max+ 395 at about 349s. Our 314s puts the base M6 close to the M3 Max's reported outcome, taking about 21% less time than the M4 Pro and 10% less than the Ryzen AI Max+ 395. For reference, the Ryzen AI Max+ 395 is quite a popular chip for small form factor systems. The framework desktop features that chip, but list prices are in Mac Studio territory.
Timed LLVM Compilation, commodity CPUs
Lower is better. Ordered quickest to slowest. M6 is this bench (macOS, 314s). Every other bar is an OpenBenchmarking.org aggregate for pts/build-llvm 1.5.x, Ninja, 9950X 286s ±17, Core Ultra 9 285K 287s ±6. Additional approximate anchors: M3 Max 300s ±42, M4 Pro 399s ±59, Ryzen AI Max+ 395 349s ±1. Apple M4 is ±22s. The ± figures are retained as reported, without assuming a confidence interval; Apple core-count variants are unspecified here. Different OS, compiler, and memory from the Mac Mini.
What the Telemetry Shows
The elapsed times above come from Phoronix, not from this trace. The logger stopped while the compile is running, but you can see the pattern, this chip is fully utilised.
Across that window the M6 held 98-99% CPU utilisation with P-cores averaging 4.66GHz and E-cores climbing to around 2.93GHz as the build saturated every core. There is no visible ramp down, which matches what Cinebench showed, this chassis does not appear to run out of cooling on an all-core load.
Mac Mini M6 - LLVM compile telemetry
Captured via powermetrics at ~1s intervals. The capture ends before the build does. CPU package power is reported as zero on this M6 build, so only clocks and utilisation are plotted.
What the Result Says
A five-minute LLVM build is long enough to expose a chip that only looks fast in a short benchmark. The M6 does not fade across it. It also clears a recent laptop-class Pro chip by a wide margin, which hints towards the architectural improvements Apple's M6 generation brings to the table.