Quite impressive on Apple M2 the boost is almost X2 , on some task X6
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Macos Sequoia 15.1 ML miracle performance boost
- Thread starter testato
- Start date
- Sort by reaction score
The boost might be just due to the amounts of RAM, as AI is quite memory intensive. Thus I don't think the linked comparison is particularly useful.
Tested also with two M3 with same amount of ram , also tested 2 different M3 with 8 vs 16 on same macos version , same results , also geekbench is not constrained by Ram
Same M3 24 GB vs 8 GB both on macos 14.6
Same M3 16GB (macos 15.0) vs 24 GB (macos 14.6)
Details macos 15.0 vs 14.6 cpu gpu ne
Hmm I checked the results and I still find them to be inconclusive:
So looking at all test only one appears to be valid for this comparison. As this one is such an outlier I'm still not convinced.
Does indeed illustrate well that 8GB memory does not to appear to be bottleneck for Geekbench.
Theres something strange here. The machine running macOS Sonoma is apparently clocked at just 2.17GHz rather than at the M3s default 4.05GHz. Could this be running in battery mode?
Now it's the 8GB version again which is significantly slower. Even though the previous test did not illustrate a difference such stark differences are somewhat suspicious.
This on has mostly a marginal difference, in favour of macOS Sequoia. Somewhat within fluctuations.
Stange as well. The slower machine is clocked slower again for this one. Though this time at 3.72GHz
So looking at all test only one appears to be valid for this comparison. As this one is such an outlier I'm still not convinced.
I understand your skepticism, however I would like to point out the following points:
- All machines running Sonoma have consistent results regardless of the clock value reported by Geekbench
- Only some tests report exaggeratedly better values, if the difference between Sequoia and Sonoma was due to the different clock all tests should report scaled values (reduced for Sonoma)
As the weeks go by we will have more data with machines with the same amount of RAM, at the moment many of those using Sequoia are 16 GB configurations.
A first hypothesis is that they have improved the ML framework in some scenarios or that they have changed how the kernel manages clock power scenarios. To confirm this, it would be necessary to match the powermetric data during the bench.
I have performed some of those done on Sonoma personally and they give the same results regardless of whether they are performed in battery or wall plug mode
- All machines running Sonoma have consistent results regardless of the clock value reported by Geekbench
- Only some tests report exaggeratedly better values, if the difference between Sequoia and Sonoma was due to the different clock all tests should report scaled values (reduced for Sonoma)
As the weeks go by we will have more data with machines with the same amount of RAM, at the moment many of those using Sequoia are 16 GB configurations.
A first hypothesis is that they have improved the ML framework in some scenarios or that they have changed how the kernel manages clock power scenarios. To confirm this, it would be necessary to match the powermetric data during the bench.
I have performed some of those done on Sonoma personally and they give the same results regardless of whether they are performed in battery or wall plug mode