First Geekbench for MacBook Pro M1 Max with 2x Faster Multi-Core Performance
macrumors.com
macrumors.com
Fairly disappointing, I would have expected the higher memory bandwidth to contribute something. I've found the STREAM benchmark correlates very well with the kind of performance I need for databases and the like, and in the past upgrading RAM from DDR4-2700 to DDR4-3200 gave a ~10% improvement on Geekbench scores under Linux.
The only reason to get one of the new MBPs is the higher RAM capacity (and the amenities of the new chassis, of course), not CPU performance.
M1 (with active cooling) https://browser.geekbench.com/v5/cpu/10480705
So seems to the big core itself isn't updated to A15's Avalanche, maybe obvious from the name.
I'm impressed anyhow.
Single-core scores around 1850 are normal for a 5950X with default settings, i.e. no overclocking.
That is better than Apple M1, but not by much.
But I wouldn't trust Geekbench across platforms.
The scores of many benchmarks, especially the scores of single-thread benchmarks, are lower and also much more variable on Windows than either on Linux or on MacOS.
My assumption is that this is due to the fact that almost all Windows computers run continuously many junk programs, e.g. antiviruses and the like, in parallel with whatever you are doing, and those interfere with the benchmarks.
Running the same benchmark on Linux or maybe on a Windows completely cleaned of garbage and with the useless services stopped, should give results well above 1800 in non-overclocked single-thread.
Ryzen 9 5900X or 5950X will have a few fast cores, maybe 2 fast cores, which will run a ST task at 4.8 GHz, respectively at 4.9 GHz. If you are lucky, and more likely if you use PBO, the fast cores of your CPU might run at a slightly higher frequency, e.g. by 100 MHz, than guaranteed, especially when you have a good cooler that keeps the CPU at a low temperature.
However the rest of the cores will have a lower maximum frequency than the nominal frequency, e.g by 100 MHz or by 200 MHz.
The Linux scheduler, at least the default one, does not schedule the threads preferentially on the fast cores, so when running a ST benchmark there are more chances to run it on a slow core than on a fast core.
You can determine manually which are your fast cores, by running some single-threaded benchmark pinned successively on each core, e.g. using taskset, while monitoring the clock frequency, e.g. with turbostat.
The 1736 value looks like a normal value for a slow core. Rerunning the benchmark, but pinned on a fast core with taskset, should give on a 5950X a value not far from 1850.
I have a slower 5900X and with a memory slower than normal (only 2667) and no PBO, but I still get over 1800 on a fast core.
Eventually the rest of the market will get access to TSMC’s 5nm capacity & we’ll see what they can do on the same process.