Benchmarks are interesting to look at. But real life performance is what really matters. They don't always match.
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.
Apple M1 CPU & GPU speed is very disappointing
- Thread starter Appletoni
- Start date
- Sort by reaction score
- Status
Yes, thanks. Your score is quite a bit lower than the 3024 kN/s reported on the forum I linked to.
Supposedly, the executable is M1 native since it's installed by homebrew. But your M1 is only about 35% faster than my iMac, while it should be more like 65% faster.
Can you run
Code:
file $(which stockfish)Stockfish is scaling very well with number of CPU cores. The actual performance matters very little, more cores = better score. Your 10 year old CPU is 6c/12t, the M1 is 4hp/4he. Makes a comparison useless as it's not the actual performance that's compared here. It's also the reason why AMD 32c/64c CPUs are on top, sheer number of cores, independent of performance.
bigmcguire@BorgCube /Applications % file $(which stockfish)
/opt/homebrew/bin/stockfish: Mach-O 64-bit executable arm64
Interesting indeed
I rebooted, unplugged 4k monitor, waited for OS fresh boot to calm down and got similar results.
Thanks. The difference with the numbers posted online may reflect different versions of Stockfish. I think they used the "classical" version while we've installed the current (NNUE) version, which currently runs slower on M1 due to missing/insufficient SIMD optimisation. However, the NNUE version has the potential to be much faster on the M1 thanks to its Neural Engine (which would require someone to implement the algorithm in coreML, which may not happen anytime soon).
That would be relevant if Stockfish used a Neural Engine (or its equivalent) on x86 CPUs but does it?
Just got back and looked into this more.
Homebrew and stockfish formula are native apps.
https://brew.sh/2021/02/05/homebrew-3.0.0/
https://formulae.brew.sh/formula/stockfish#default
To compare with results from openbenchmarking.org the same parameters should be used except for thread count.
On MBA M1 results are in-line with OP's but his is slightly higher since he probably has a MBP M1 vs my results for MBA M1.
4 threads
stockfish bench 128 4 24 default depth
===========================
Total time (ms) : 55575
Nodes searched : 453871217
Nodes/second : 8166823
8 threads
stockfish bench 128 8 24 default depth
===========================
Total time (ms) : 56470
Nodes searched : 661228390
Nodes/second : 11709374
Increasing the thread count beyond 8 just makes it worse but no surprise since M1 doesn't support SMT.
For comparison, on older low end AMD 3800xt 8-core 16-thread CPU the results are:
8 threads
stockfish_13_win_x64_avx2.exe bench 128 8 24 default depth
===========================
Total time (ms) : 36246
Nodes searched : 653634657
Nodes/second : 18033290
16 threads
stockfish_13_win_x64_avx2.exe bench 128 16 24 default depth
===========================
Total time (ms) : 38215
Nodes searched : 999371451
Nodes/second : 26151287
Conclusion is stockfish benefits from SMT and also optimized for AVX2 vector instruction and BMI2 bit manipulation instruction on Zen 3 (AMD 5000x line of CPUs) which should be even faster. M1 doesn't support SMT but does have SVE vector and BITIMTR bit manipulation instructions so it's to be seen if stockfish can be further optimized for M1.
OP is too advanced and should be hanging out on Phoronix instead.
Homebrew and stockfish formula are native apps.
https://brew.sh/2021/02/05/homebrew-3.0.0/
https://formulae.brew.sh/formula/stockfish#default
To compare with results from openbenchmarking.org the same parameters should be used except for thread count.
On MBA M1 results are in-line with OP's but his is slightly higher since he probably has a MBP M1 vs my results for MBA M1.
4 threads
stockfish bench 128 4 24 default depth
===========================
Total time (ms) : 55575
Nodes searched : 453871217
Nodes/second : 8166823
8 threads
stockfish bench 128 8 24 default depth
===========================
Total time (ms) : 56470
Nodes searched : 661228390
Nodes/second : 11709374
Increasing the thread count beyond 8 just makes it worse but no surprise since M1 doesn't support SMT.
For comparison, on older low end AMD 3800xt 8-core 16-thread CPU the results are:
8 threads
stockfish_13_win_x64_avx2.exe bench 128 8 24 default depth
===========================
Total time (ms) : 36246
Nodes searched : 653634657
Nodes/second : 18033290
16 threads
stockfish_13_win_x64_avx2.exe bench 128 16 24 default depth
===========================
Total time (ms) : 38215
Nodes searched : 999371451
Nodes/second : 26151287
Conclusion is stockfish benefits from SMT and also optimized for AVX2 vector instruction and BMI2 bit manipulation instruction on Zen 3 (AMD 5000x line of CPUs) which should be even faster. M1 doesn't support SMT but does have SVE vector and BITIMTR bit manipulation instructions so it's to be seen if stockfish can be further optimized for M1.
OP is too advanced and should be hanging out on Phoronix instead.
OP posted results without specifying anything and attributed scores of 230000 kN/s to desktop CPUs. OP doesn't appear very advanced.
I was able to replicate his results and just an average user and started using MacOS since 2/2021. Prior to that it's Mac SE.
Wow, you are really not doing a good job selling the phoronix crowd ?
And prior to your post, we compiled and ran stockfish on our Macs. Maybe you were inspired.
We were not interested in reproducing the OP's number as 8/16-thread runs are not very relevant to compare a ~20W CPU with 4 performance cores to a 105-W CPU (AMD 3800XT) with 8 SMT cores.
Anyway, it doesn't change the fact that the OP isn't much "advanced", as they provided no way to reproduce the results they posted. One cannot reproduce an experiment by guesswork.
I think Chess is something to do with red-headed orphan girls...I saw a documentary on Netflix about this...
That was a great series. As a chess player myself, it's been a long time since a good chess movie/series came out.
I quit looking at these "benchmarks" a while back. How it performs doing the tasks we want is the real test.
My observation between my new MBA 7core/256 vs my iMac with Core I9 8 core @ 3.g GHz . Using LibreOffice, same version on both, is that LibreOffice on the MBA starts up as fast as the iMac. Keep in mind that LibreOffice on the MBA has to run through Rosetta 2.
That tells me the new M1 chips are more than enough for anything I will ever want to do.
My observation between my new MBA 7core/256 vs my iMac with Core I9 8 core @ 3.g GHz . Using LibreOffice, same version on both, is that LibreOffice on the MBA starts up as fast as the iMac. Keep in mind that LibreOffice on the MBA has to run through Rosetta 2.
That tells me the new M1 chips are more than enough for anything I will ever want to do.
For comparison, results from $500 Lenovo Yoga 6 with 7nm AMD 4650U 6-core 12-thread comparable to MBP M1 since they both have fan but only spin up under heavy load. Now need to find 4800U and 5800U 8-core 16-thread.
One thing I noticed is running a few tests on the MBA M1 tanks the battery but not on the AMD. Have to do a controlled test to verify.
12-thread
stockfish_13_win_x64_avx2.exe bench 128 12 24 default depth
===========================
Total time (ms) : 68529
Nodes searched : 830608234
Nodes/second : 12120536
One thing I noticed is running a few tests on the MBA M1 tanks the battery but not on the AMD. Have to do a controlled test to verify.
12-thread
stockfish_13_win_x64_avx2.exe bench 128 12 24 default depth
===========================
Total time (ms) : 68529
Nodes searched : 830608234
Nodes/second : 12120536
Last edited:
I don't see how it is a relevant benchmark if it is an Intel binary running via Rosetta 2.
Also, relatively few people actually play chess. I guess those that do should stick with Intel for now.
So what? What possible meaning can you get out of this "benchmark"?
This isn't a (ahem) measuring contest.
No, it's time to throw in the towel and admit that M1 is overhyped. I mean, if it can't dominate in that one benchmark, it just isn't worth buying. We should all abandon our Macs and switch. Surely Apple was foolish if they thought they could take on AMD and Intel.
Where? Your 3800x is score is not even close to the 230000 kn/s desktop CPU line, almost 90% lower. Either OP is using an 80 core 160 thread x86 CPU, or the OP is using a different configuration.
Oh, please don't. Do you expect a Phoronix thread comes with only a result but no testing configuration?
M1 has an 8-wide decoder, which is by fart the widest commercialized design in the industry. Intel Skylake and AMD Zen(1,2,3) only has a 4-wide decoder, mainly due the the complexity of x86 instructions. To utilize such a wide decoder (and the excecution units in the backend) we have to rely on some super-scaler technique, but combination, used in this chess engine, usually has a very sequential instruction dependency which makes instruction level parallelism hard. This particular case is where SMT providing us thread level parallelism shines.
We don't downplay unfavorable benchmarks, but this particular workload is one of many cases that M1 is not good at(for now). If this particular case is very relevant to the user, that user should not buy an M1 equipped Mac. A lot of disagreement here only want to say 'M1 is fast for me' because OP comes with a scary title.
Last edited:
- Status