[Sunham Kim]

Some updated numbers for those who wonders how Stata works on the brand-new Apple Silicon.

M1 Mac-mini, base model (8GB ram)
Stata 16.1 SE. Not sure if it's compiled for Apple Rosetta but it's running smoothly.

variable	p50	mean	sd	min	max
replace	.012	.0123	.0008233	.011	.014
regress	.0825	.0827	.0012517	.081	.084
predict	.027	.0271	.0012867	.026	.03
correl	.061	.0646	.0118152	.059	.098
bootstrap	12.5705	12.6407	.217314	12.482	13.21
mvtest	.1675	.1688	.0084696	.153	.187
xtile	.751	.7464	.0284066	.683	.781
expand_drop	.053	.0541	.0026437	.052	.061
arfima	87.2375	88.9892	3.781558	85.517	94.98
eigenv	.229	.2299	.0041753	.225	.239

All commands are more or less oka, at least compared to OP's figures, but arfima is terribly slow. Don't know why.

[Sunham Kim]

I should have attached this table on my previous posting.
For comparison...

MacBook Pro 16" (2019) Base Model (i7 2.6Ghz 6-core, 16GB RAM)
Stata 16.1 SE

variable	p50	mean	sd	min	max
replace	.0135	.0143	.0024518	.013	.021
regress	.0835	.0836	.0012649	.082	.086
predict	.0285	.0287	.0014181	.027	.031
correl	.061	.066	.0165261	.06	.113
bootstrap	11.1815	11.2029	.0619255	11.141	11.344
mvtest	.2035	.208	.0150555	.196	.238
xtile	.6905	.7192	.0767736	.622	.88
expand_drop	.0685	.069	.0033665	.065	.074
arfima	138.294	134.2177	5.631907	127.072	138.865
eigenv	.3055	.3342	.0930971	.3	.599

- The M1 Mac mini is marginally better than the Intel MacBook Pro. MBPr price = 3.5x Mac mini price and Stata is based on Intel binaries. Well...
- arfima performance is even worse. It might be a Mac problem, rather than a M1 one.

It would be interesting to see how those are different in multicore environment... but I do not have a MP license.

[Matthew Mercurio]

Okay, well I know this thread is a bit outdated, so it may be an unfair comparison, but here is the results run on my Alienware R15, Intel i9-13900KF, 24 Cores, 64G RAM

Stata license: Single-user 24-core , expiring 20 Apr 2024

Variable | p50 Mean SD Min Max
-------------+--------------------------------------------------
replace | .001 .0013 .000483 .001 .002
regress | .011 .011 .0008165 .01 .012
predict | .0025 .0036 .0036878 .002 .014
correl | .005 .0062 .0034577 .005 .016
bootstrap | 2.7235 2.7603 .0992438 2.678 3.001
mvtest | .049 .0503 .0027909 .047 .055
xtile | .232 .2313 .0127284 .214 .257
expand_drop | .0355 .0363 .0042439 .028 .042
arfima | 2.187 2.2132 .0847726 2.139 2.442
eigenv | .184 .1847 .0042177 .179 .192
----------------------------------------------------------------

W00t!!!

[ericmelse]

Interesting Matthew,

On my system:
13th Gen Intel(R) Core(TM) i9-13900K 3.00 GHz
RAM 64.0 GB
Windows 11 Pro (installed on SSD)
Stata/MP 18.0 : Single-user 12-core

I have a near similar result:

    Variable |       p50      Mean        SD       Min       Max
-------------+--------------------------------------------------
     replace |      .001     .0013   .000483      .001      .002
     regress |      .015     .0155  .0015092      .014      .018
     predict |      .003     .0032  .0013984      .002      .007
      correl |      .006     .0067  .0020028      .005      .012
   bootstrap |    2.8745    2.8973  .0467239     2.847      2.98
      mvtest |     .0475     .0488  .0037653      .045      .056
       xtile |      .219     .2206  .0124383      .199      .247
 expand_drop |     .0415     .0409  .0025144      .037      .044
      arfima |    3.0005    3.0169  .0876755     2.892     3.208
      eigenv |     .1845     .1854  .0030258      .181      .191
----------------------------------------------------------------

[George Ford]

Dang, looks like I need more cores on both my CPU and my Stata. The 12 core ran the bootstrap in 40% of the time than did my 4-core Stata 18 (even with a higher clock speed on my Ryzen 7 2700x).

[Leonardo Guizzetti]

My computer i7-1260P (12 cores/16 logical processors), 32 GB DDR4 RAM, Stata 18 MP/4.

    Variable |       p50      Mean        SD       Min       Max
-------------+--------------------------------------------------
     replace |      .003     .0032  .0004216      .003      .004
     regress |      .025     .0279  .0095621      .024      .055
     predict |      .006     .0084  .0042479      .005      .018
      correl |     .0135     .0141  .0018529      .013      .019
   bootstrap |    4.6275    4.7257  .3023861     4.552     5.552
      mvtest |     .0895     .0882   .004638      .077      .093
       xtile |     .3595     .3521  .0185739      .305      .365
 expand_drop |      .058     .0568  .0045412      .044      .059
      arfima |    3.1775    3.1712  .0551338      3.05      3.26
      eigenv |      .237     .2367  .0012517      .234      .238
----------------------------------------------------------------

[alejoforero]

My results on a:
CPU: 13700k, running at 5ghz
RAM: DDR 4, 128GB at 3200ghz, dual channel setup.
Storage: gen 4 NVME (4GBS+ read and write)
Stata 18, revision 30 aug 2023, MP4.

Variable	p50	Mean	SD	Min	Max
replace	.002	.0024	.0005164	.002	.003
regress	.0175	.0179	.0020248	.016	.023
predict	.0055	.0056	.0016465	.004	.008
correl	.013	.0137	.0043218	.01	.025
bootstrap	3.2785	3.2828	.0422263	3.226	3.362
mvtest	.064	.0637	.0047152	.053	.071
xtile	.249	.2468	.0078429	.234	.259
expand_drop	.035	.0356	.0030258	.031	.043
arfima	2.2175	2.2363	.059498	2.184	2.345
eigenv	.187	.1877	.0023594	.186	.194

Stata SE (set processors 1). (the cpu speed automatically increases to 5.2ghz):

Variable	p50	Mean	SD	Min	Max
replace	.008	.0085	.0007071	.008	.01
regress	.0615	.0609	.0024244	.058	.064
predict	.016	.0168	.0014757	.015	.02
correl	.0465	.0469	.0011005	.046	.049
bootstrap	5.1195	5.1298	.0573543	5.049	5.212
mvtest	.1275	.1285	.004223	.122	.136
xtile	.293	.2917	.010133	.267	.303
expand_drop	.0375	.0372	.0030478	.032	.042
arfima	2.7505	2.7471	.0865094	2.638	2.912
eigenv	.1895	.1903	.0025841	.188	.196

Similar configuration with a Ryzen 3900x 4ghz, MP4:

Variable	p50	Mean	SD	Min	Max
replace	.003	.0031	.0007379	.002	.005
regress	.0275	.0286	.0037476	.026	.039
predict	.012	.0127	.0040838	.007	.021
correl	.016	.018	.0051854	.015	.031
bootstrap	5.487	5.5256	.1562919	5.404	5.943
mvtest	.1065	.1061	.0027264	.1	.111
xtile	.426	.4233	.0128327	.392	.44
expand_drop	.0595	.0598	.0028597	.054	.065
arfima	4.985	5.0383	.1636324	4.964	5.501
eigenv	.2865	.2875	.0028771	.284	.293

My conclusion: better to invest first in a fresh CPU (350 USD, lifetime) than in an expensive MP license beyond 4 cores.
CPU vintage and clock speeds are king.

[George Ford]

Hmm. The 13700k smoked the 3900x, but has a much higher clock speed. From what I've read, it seems like the clock speed is the most important factor.

[alejoforero]

Clock speed is certainly key, but not everything: I reran the Intel13700k limiting the clock speed to 4ghz, comparable to the AMD Ryzen (everything else, RAM, etc, constant), and these are the results:

Variable	p50	Mean	SD	Min	Max
replace	.003	.0032	.0004216	.003	.004
regress	.023	.0233	.0010594	.022	.026
predict	.0075	.0069	.001792	.005	.01
correl	.016	.0151	.0015239	.012	.017
bootstrap	4.1	4.1079	.0278786	4.084	4.168
mvtest	.073	.075	.0048762	.072	.088
xtile	.282	.2832	.0057696	.277	.297
expand_drop	.042	.0427	.0020575	.04	.048
arfima	2.5775	2.5911	.0671126	2.528	2.773
eigenv	.2545	.255	.0023094	.252	.259

So the Intel is faster than the AMD even at the same clock speeds. There are two possible explanations:

1. Vintage: the 13700k (and 13900k, etc) is a 2022 cpu whereas the 3900x is a 2019 cpu. This impacts optimization, instructions per clock, etc. Both cpus have avx2, so thats not the reason.

2. Intel optimization: Stata uses Intel Math Kernel Library (MKL) https://www.stata.com/stata17/math-kernel-library-mkl/, which, while runs on AMD, is better optimized for Intel CPUs.

The lesson so far seems to be: get the latest, fastest (clock speed), Intel CPU you can. Then get the highest core count license you can afford.


One missing variable in this analysis is the number of observations. It is entirely possible that the ability of Stata to scale across many cores is related to the number of observations. All these benchmarks have just 500k observations, which is rather small and parallelization savings might be overrun by parallelization overhead. In the Stata MP report N varies by command, but can be 30 million or higher.

[George Ford]

Interesting. Thanks for the additional work. Looks like I'll be moving to an Intel chip next round.

[Martin Schroeder]

Let me chime in there. I tested an i9-13900h notebook and an Apple M3 Pro, also because I wanted to know whether the best Apple notebooks can keep up with the best Intel ones. Three things to learn:

First, it seems that for Intel, setting your CPU to Maximum Energy ("beste Leistung" in German, don't know what it is in English), is important (see first and second test).

Second, the M3 is almost always faster then the i9, despite the i9 reaching 5.4 Ghz and the M3 only 4 Ghz. But there is one weird exception to the fast M3 speed: arfima is about 20 (!) times slower on the Mac! (This seems to be a general Apple problem, users above that had an Apple had similarly slow speeds in arfima). This suggests that Stata has to optimize some commands for Apple, while overall Apple is pretty fast in Stata.

Third, since the M3 pro only has 6 power cores (and 6 efficiency cores), only using 6 rather than 8 cores can actually speed up the calculations, as the 8-core setup makes Stata use 6 fast cores plus 2 slow cores, which (with diminishing returns with each additional core) actually rather slows the calculations down, rather than speeding them up. So for M3 users, I would suggest setting the number of cores to your computer's power cores, leaving the efficiency cores to handle the OS.


See the difference:

i9 Balanced 8 cores
Variable | p50 Mean SD Min Max
-------------+--------------------------------------------------
replace | .002 .0022 .0004216 .002 .003
regress | .02 .0206 .0011738 .019 .023
predict | .005 .0054 .0012649 .004 .007
correl | .01 .0103 .0014181 .009 .014
bootstrap | 4.0085 4.0014 .0341344 3.935 4.04
mvtest | .0685 .0689 .0031073 .065 .075
xtile | .3195 .3153 .0166203 .28 .334
expand_drop | .04 .0412 .0024404 .039 .045
arfima | 2.579 2.7908 .6079641 2.518 4.509
eigenv | .1645 .1648 .0042895 .156 .171
----------------------------------------------------------------

i9 Maximum energy 8 cores
Variable | p50 Mean SD Min Max
-------------+--------------------------------------------------
replace | .002 .002 .0004714 .001 .003
regress | .018 .0188 .002974 .017 .027
predict | .004 .0042 .0004216 .004 .005
correl | .008 .0082 .0004216 .008 .009
bootstrap | 3.4995 3.5048 .0378294 3.452 3.577
mvtest | .067 .0667 .0022632 .062 .07
xtile | .269 .2665 .0153858 .237 .286
expand_drop | .04 .0392 .0034254 .032 .044
arfima | 2.4985 2.4619 .0907346 2.3 2.552
eigenv | .142 .1411 .0026013 .137 .144
----------------------------------------------------------------

M3 pro 8 cores
Variable | p50 Mean SD Min Max
-------------+--------------------------------------------------
replace | .002 .0018 .0004216 .001 .002
regress | .013 .0129 .0003162 .012 .013
predict | .004 .0037 .0006749 .003 .005
correl | .007 .0069 .0003162 .006 .007
bootstrap | 2.9405 2.9445 .0200181 2.922 2.989
mvtest | .043 .0427 .0018288 .038 .045
xtile | .266 .2615 .0135831 .223 .267
expand_drop | .032 .0323 .0006749 .031 .033
arfima | 61.8575 62.078 .8496164 61.307 64.351
eigenv | .15 .1499 .0017288 .146 .152
----------------------------------------------------------------


M3 pro 6 cores
Variable | p50 Mean SD Min Max
-------------+--------------------------------------------------
replace | .0015 .0015 .000527 .001 .002
regress | .0125 .0124 .0006992 .011 .013
predict | .003 .0036 .0009661 .003 .006
correl | .007 .0068 .0006325 .006 .008
bootstrap | 2.85 2.8473 .0164725 2.818 2.867
mvtest | .042 .0437 .003401 .041 .05
xtile | .2605 .2578 .0150466 .217 .272
expand_drop | .0325 .0324 .0006992 .031 .033
arfima | 69.6715 68.7322 2.183425 63.51 70.001
eigenv | .1495 .1498 .0022509 .146 .153
----------------------------------------------------------------