Looking for Benchmark performance of Mac Studio

Question

I'm considering a purchase of a new Mac Studio to replace my current Dell XPS Core I-7 system. Does anyone have a benchmark report for MMA v13 running under Mac OS 12.4? A report for the M1 Ultra configuration would be ideal, but even the M1 Max set-up would be very helpful.

Answer 1

The current Mac Studio has an Apple M1 chip with either:

• M1 Max: 10-core CPU (+ 24-core or 32-core GPU) or
• M1 Ultra: 20-core CPU (+ 48-core or 64-core GPU)

The 10-core M1 MAX CPU consists of 8 standard cores and 2 efficiency cores. The ULTRA version is basically two Max chips stuck together.

The benchmark report below includes results for the M1 Max chip used in the Mac Studio.

For comparison, I have carried out tests on the new Mac Mini M2 Pro just released, running Mma 13.2:

• M2 Pro: 12-core CPU (and 19-core GPU).

The 12-core M2 Pro CPU consists of 8 standard CPU cores and 4 efficiency cores.

Caveat re cores

I should mention, as caveat, that having more and more cores won't necessarily improve the performance of Mathematica on your computer. This is because Wolfram artificially restrict the number of kernels that can run on your computer, depending on the type of license you have.

• For a home license, your computer is restricted by Wolfram to using 4 kernels.
• For an education license, your computer is restricted by Wolfram to 8 kernels.

I believe it is possible to pay extra money to allow Mathematica to take advantage of further cores on your computer --- but suspect that the uptake of such licensing is very limited, and that most people end up with the standard restrictions. As such, even if you buy the 20-core Mac Studio Ultra, you unfortunately probably won't see any increase in performance from Mathematica than over the 10-core version --- which is disappointing. My personal view is that the deliberate crippling of hardware by a software package is unsavoury and unsatisfactory. In any event, given the proliferation of cores on even entry level computers in recent years, the current Wolfram structure appears outdated. This issue (of the restriction of kernels) has very real relevance to the question being posed by the OP (Ultra vs Max).

Mac Mini M2 Pro

The Mac Mini M2 Pro being tested has 12 cores. Of these, 8 are standard cores and 4 are efficiency cores. I don't know if Mathematica can run a kernel on (or take advantage of) the efficiency cores: my understanding is that the efficiency cores are intended to take care of low level threads in the background, keeping the high performance cores free to do the heavy lifting. In any event, Mathematica restricts my new Mac to 8 kernels anyway, and the new Mac M2 Pro has 8 standard CPU cores.

Since the M2 chip is supposed to be faster than the M1 chip (other things being equal), I would expect that the Mac Mini M2 Pro will produce better Mathematica benchmarks than the Mac Studio (for all the reasons above) -- even if the Studio has 20 cores. A related question is to what extent Mma takes advantage of GPU cores, and to what extent the Benchmark carries out tests that takes advantage of the GPUs.

Mac Mini M2 Pro Benchmark results

Starting from a fresh launch yields a Mma Benchmark of 5.17 in a total time of: 2.7 seconds.

Needs["Benchmarking`"]
Benchmark[]

{"MachineName" -> "mac-mini", "System" -> "Mac OS X ARM (64-bit)", "BenchmarkName" -> "WolframMark", "FullVersionNumber" -> "13.2.0", "Date" -> "February 13, 2023", "BenchmarkResult" -> 5.167, "TotalTime" -> 2.679, "Results" -> {{"Data Fitting", 0.138}, {"Digits of Pi", 0.144}, {"Discrete Fourier Transform", 0.222}, {"Eigenvalues of a Matrix", 0.223}, {"Elementary Functions", 0.315}, {"Gamma Function", 0.184}, {"Large Integer Multiplication", 0.157}, {"Matrix Arithmetic", 0.047}, {"Matrix Multiplication", 0.09}, {"Matrix Transpose", 0.079}, {"Numerical Integration", 0.263}, {"Polynomial Expansion", 0.042}, {"Random Number Sort", 0.351}, {"Singular Value Decomposition", 0.268}, {"Solving a Linear System", 0.156}}}

Here is the graphic comparison to other systems:

BenchmarkReport[]

Note that the above graphic includes a 10-core M1 Max which scores 4.5 and which would be presumably almost identical to the M1 Max version of the Mac Studio.

Final comment

A benchmark test that only takes 2 or 3 seconds to evaluate is not a very useful benchmark. It might have been fine 20 years ago ... but it is in need of an update. As others have noted in other threads, running LaunchKernels[] prior to running the benchmark produces much higher benchmark numbers, and much slower speeds, which sounds dubious, so it is not really very clear what the benchmark is testing in that scenario, and also why it is not automatically testing (and distinguishing between) single core AND multicore performance, like almost all other benchmarks do.

Nor is it clear how that benchmark package separates the performance of the machine (which a benchmark is meant to be testing) from the restrictions of the software, when the software may be artificially crippling the hardware on some computers simply because of the user's license type.

Answer 2

I just started looking at Mathematica and have a trial version which is limited to 4 kernels. I'm running it on AMD 3995WX CPU which has 64 cores/128 threads. This however should not matter because Mathematica will only use 4 threads, and this CPU frequency is lower than several others included in the benchmark report. Yet, it beats all those CPUs by substantial margin. Maybe much larger caches help? That would probably only happen if software purposely written to take advantage of this.