SPUMD Logic CPU Test - obvious M1 issue?

[SPUMD]

Hi friends,

 

2 years ago i set up an Logic CPU Test that simulated situations like having enough power but CPU spikes causing CPU dropouts stopping the playback

It took some time to find the right balance but it was pretty helpful to understand how to deal with CPU settings, diffrent output devices can cause different results aso...

 

I tested with:

 

---

MacOS: Mojave 10.14.6

Mac Model: Mac Pro (Early 2009)

Prozessor: 12 x 3.46 Intel Xeon

RAM: 48 GB

Logic Version: 10.4.8 - 10.5.1

---

Depending on the output device, I was able to (clean) playback

 

20 Tracks with MH2882 2D Output Device (MH had a driver Issue that caused cpu "crackles" at an early stage and only 20 tracks could be playbacke dean)

89 Tracks with Loopback Output Device

91 Tracks with Merging VAD Output Device

 

This Mac Pro setup Geebenches @ 559 (single core) 3905 (muli core)

Now I have an M1 MacBook Air 16 GB RAM too, and it Geekbenches @ 1705 (single core) 7415 (muli core)

 

Following the Geekbench scores everybody would assume, the M1 Air preforms better in this test, but it doesn't, it preforms way worse!

---

MacOS: Big Sur 11.6.1

Mac Model: MacBook Air (2020)

Prozessor: M1 8 core

RAM: 16 GB

Logic Version: 10.7.2

Output Device: Internal Speaker

 

clean playback: 40 tracks with Internal Speaker Output Device

---

 

Temp Summary: My M1 AirBook with a 3x higher single core and 1.9x higher multi core Geekbench performance clean playbacks 2.25x lesser tracks in this test compared to my old Mac Pro.

 

What is the issue? I don't know, that's why I opened this thread cause I want to replace my old Mac Pro with a new Mac Studio and this result confuses me.

I would appriciate if folks in this forum, would do the test to, especially those who own a M1 Max to see if there is something wrong with my MacBook Air of it's a general issue that should be reported to Apple.

 

1) Please use the following Logic Audio Settings:

 

I/O Buffersize: 128 Samples .... (if only playback, lower settings bring more preformance when you push the cpu to the limit)

Processing Threads: Automatic

Processor Buffer Range: Large

Multithreading: Playback & Live Tracks

Summing: Standard Precision (32 Bit)

Rewire Behavior: Off

 

2) Please do the test with different Output Devices if possible. I tested my M1 only with the Internal Speakers cause all my audio Interfaces now have a ethernet connection and I don't have a thunderbolt adapter.

 

3) When you open the test project, you will see only one software instrument track loaded with an alchemy setting and a Bach midifile on it. Send 1 is open and points to an AUX Track with 5 Chroma Verbs in a row. Make sure all verbs are turned on. Stay on the set playback loop cause that's the part where the CPU always drops out cause there are midi glissandos. (more data at once for the CPU). You will see also an Midi Track. Some people use it to push the limit cause if you select the Midi track and start the playback, Logic says: cool, no software instrument i have to calculate if the user decides to play on on the midi keyboard and releases more cpu power.

Is this a real world test? No, cause we are testing the CPU, but since I alway ran into CPU spikes in a production that caused CPU dropouts, stopping the playback, it simulates what happens when you run into real world limits. And no, don't let the test frustrate you, it is set up to start with an heavy cpu load.

 

Download Link SPUMD Logic CPU Test

 

Stop the playback, duplicate the Alchemy tracks (best in 10s), copy the midi region to the new alchemy tracks, playback again till the cpu drops out. Write the results in this thread providing the following informations: (p.s. close all other applications like safari while testing)

 

---

MacOS: xxx

Mac Model: xxx

Prozessor: xxx

RAM: xxx

Logic Version: xxx

 

clean playback: xxx tracks with xxx Output Device

---

 

 

Thanks for participating.

Greetings SPUMD

Edited March 16, 2022 by SPUMD

[des99]

I see you ran your test on 10.7.2 - have you tried running it on 10.7.3 yet?

 

Edit: Logic 10.7.3, 16GB M1Pro MBP, Monterey 12.3, with a USB audio interface, I got 84 tracks with those settings (and 86 on internal MBP speakers) - which is a bit over double your M1's result of 40 tracks, which is roughly what I'd expect, as the M1 Pro/Max has double the high performance cores.

 

The M1 Pro should give more or less the same performance as an M1 Max in this test.

 

Note that the performance you will see is heavily dependent on how optimised Alchemy is for native - I've got a feeling it was heavily optimised for Intel, but not yet quite as optimized for M1, but may be wrong on that. You might see different results with a different test plugin, for example.

 

I'm not sure why you think your computer isn't working well - you say multicore performance is basically 2x your previous Mac, and you get basically 2x times the tracks (20 -> 40). My M1Pro is roughly 2x your Macbook Air, and I get roughly 2x the playback track counts (40 -> 80+). The maths seems to add up... or maybe I'm not understanding you correctly possibly... I don't really know what your other "loopback" drivers etc are, and it maybe be possible to get better performance by optimising settings too, of course.

[SPUMD]

No, did Apple optimise the performance so significantly? there is a difference of of clean playback tracks by -50%. That's huge.

[exmachina]

Doubt this is terribly helpful but for grins and historical perspective:

---

Mac Model: 27-inch, Late 2012)

Prozessor: 3.2 GHz Quad-Core Intel Core i5

RAM: 32 GB 1600 MHz DDR3

Logic Version: 10.6.3 (I'm stuck on Catalina)

 

clean playback: 18 tracks with M-Track 2x2 Output Device

clean playback: 20 tracks with Internal Output Device

---

[des99]

Doubt this is terribly helpful but for grins and historical perspective:---

 

I could run it on my 2011 MBP but I fear the results would be even more embarrassing and the fan noise would probably wake the household up...

[SPUMD]

I see you ran your test on 10.7.2 - have you tried running it on 10.7.3 yet?

 

Edit: Logic 10.7.3, 16GB M1Pro MBP, Monterey 12.3, with a USB audio interface, I got 84 tracks with those settings (and 86 on internal MBP speakers) - which is a bit over double your M1's result of 40 tracks, which is roughly what I'd expect, as the M1 Pro/Max has double the high performance cores.

 

Your Results: (It's easier to read if you use this format)

 

MacOS: Monterey 12.3

Mac Model: MacBook Pro

Prozessor: M1 Pro (How much cores?)

RAM: 16 GB

Logic Version: 10.7.3

 

clean playback: 84 tracks with "USB audio interface" Output Device

clean playback: 86 tracks with Internal Speakers Output Device

 

Your MBP M1 Pro Geekbenches @ 1746 (sc) & 12231 (mc) if you have 10 cores

My M1 Air Geekbenches @ 1705 (sc) 7415 (mc) with 8 cores (4 highperformance and 4 efficiency)

 

You really think you have "double the high performance cores"? Couldn't find infos about that. Can you. check your Systim Informations > Hardware > Total Number of Cores: ?

 

 

Note that the performance you will see is heavily dependent on how optimised Alchemy is for native - I've got a feeling it was heavily optimised for Intel, but not yet quite as optimized for M1, but may be wrong on that. You might see different results with a different test plugin, for example.

 

Thanks for the hint.

 

I'm not sure why you think your computer isn't working well - you say multicore performance is basically 2x your previous Mac, and you get basically 2x times the tracks (20 -> 40). My M1Pro is roughly 2x your Macbook Air, and I get roughly 2x the playback track counts (40 -> 80+). The maths seems to add up... or maybe I'm not understanding you correctly possibly... I don't really know what your other "loopback" drivers etc are, and it maybe be possible to get better performance by optimising settings too, of course.

 

Like I wrote, the Output Device is has an impact that's why listed different for my old intel Mac Pro

 

20 Tracks with MH2882 2D Output Device (MH had a driver Issue that caused cpu "crackles" at an early stage and only 20 tracks could be playbacke dean)

89 Tracks with Loopback Output Device

91 Tracks with Merging VAD Output Device

 

My old Intel 12 core Mac Pro outputs more clean playback tracks than your MBP M1 Pro while it Geeckbenches much lower. That's the issue.

Edited March 16, 2022 by SPUMD

[SPUMD]

Doubt this is terribly helpful but for grins and historical perspective:

---

Mac Model: 27-inch, Late 2012)

Prozessor: 3.2 GHz Quad-Core Intel Core i5

RAM: 32 GB 1600 MHz DDR3

Logic Version: 10.6.3 (I'm stuck on Catalina)

 

clean playback: 18 tracks with M-Track 2x2 Output Device

clean playback: 20 tracks with Internal Output Device

---

 

Thanks for participating. This test is not only for me, it's for everyone to compare performances.

[des99]

You really think you have "double the high performance cores"? Couldn't find infos about that. Can you. check your Systim Informations > Hardware > Total Number of Cores: ?

 

There is only one M1 Pro chip configuration of 2 E Cores and 8 P Cores - so yes, it has twice the P Cores of the M1. It's the same as the Max in this regard.

In "automatic" mode though, which is a test requirement of yours, I don't know how those cores are allocated, which is another test variable.

 

My old Intel 12 core Mac Pro outputs more clean playback tracks than your MBP M1 Pro while it Geeckbenches much lower. That's the issue.

 

In this one test - like I say, Alchemy was a third-party product originally, and it was highly optimised for Intel, and some of those optimisations aren't possible in the same way on the quite different ARM chips. So rather than taking this result as a representative comparison of the power of each machine, it might be an idea setting up a few different tests with different plugins, and then comparing those results too. Maybe your test is indeed a representative test, but it also might not be - which is hinted at by the more general Geekbench scores of the machines.

 

I don't know why you get such massive different results with different audio drivers, that seems strange to me.

[SPUMD]

Ok, I will test it with another plugin and what happens when changing Processing Threads in the audio preferences. I picked Alchemy cause it's so heavy on CPU.

There are two M1 Pro Prozessors: 8 & 10 Cores.

 

Metric Halo had an driver issue with the 2D card. Basically I could run over 80 Tracks with it but only 20 with a clean playback in this. Adding more caused CPU "crackels" in the sound like digital noise or dirt. This problem was the reason why i created this test and that's how i found the problem.

[des99]

Ok, I will test it with another plugin and what happens when changing Processing Threads in the audio preferences. I picked Alchemy cause it's so heavy on CPU.

 

I'd suggest trying another instrument, and also try an audio plugin test (other people have setup tests for this stuff already you can try, if you search around).

 

There are two M1 Pro Prozessors: 8 & 10 Cores.

 

Ah, you're right, there's a binned option on the low end 14" that disables two P cores. It's basically a way to sell off broken M1 Pro's during manufacturing - they just disable up to two cores that don't work, and slap them in the low end 14". I had forgotten about that, thanks...

[SPUMD]

I did the test using the Logics Sculpture plugin. No big improvement only 49 clean playback tracks, cause the plugin preset has a lower cpu load by default. So the plugin can't be the issue.

 

Seems like the M1 ain't the kiler in this test compared to my old intel 12 core Mac Pro.

[des99]

No big improvement only 49 clean playback tracks

 

Compared to what?

 

Are you using the same test, just replacing Sculpture for Alchemy?

[SPUMD]

Compared to the previous test. Yes, the same test Alchemy replaced by Sculpture.

It's not the plugin. If it would be I would be able to clean playback way, way more tracks.

[des99]

I get that but you haven’t said how many tracks you got on the other machine in this new test, so I have nothing to compare it to… 49 tracks on one machine, no idea how many tracks on the first machine…

[SPUMD]

Right, i did this test on the other machine a 1000 times. it would increase by 9 clean playback tracks too cause the Sculpture eats lesser CPU. The relation wouldn't change.

[des99]

Ok, so you're saying in essence you're still seeing double the tracks on the Xeon machine with half the multicore Geekbench score, compared to the M1. Which would seem to run counter to expectations given on available CPU power alone.

 

This leads me to think that probably the biggest factor in the case of synths at least, is optimisations. For example, there are very good Intel optimisations where with clever coding you can, for instance, calculate four voices in the same CPU time as one, and coding and optimisation strategies for Intel are quite mature - many instruments took advantage of those optimisation techniques.

 

I know from what other developers have said that that particular Intel optimisation/technology isn't available on the M1 chips, which can initially impact performance on code that heavily leverages it. Now, there might be additional M1 optimisations found over time as people get better at coding for these new architectures, but probably devs are just trying to get their codebases to work in the first case, and probably haven't spent a lot of time optimising yet.

 

It would be interesting to see, as I said a few times above, comparisons using audio tracks and audio plugins, rather than synths (again, there are well known benchmarking Logic projects already for these) to see if you still get the performance differences you see with synths, or whether it shifts a bit, because from what I remember audio plugins aren't quite as affected by those optimisations in the same way.

 

Because *something* has to explain the relative differences your tests show, compared to the overall processor power, and this is really the only thing I can think of that could lie at the root of it - other than flawed tests, or perhaps the audio settings you are using are not the best settings for the M1 and are throttling it down because of it. I believe the Geekbench scores to generally be a fairly robust overall measure of CPU performance, and if they are not reflected in testing in this case, that would indicate that code differences between architectures could well be at the heart of it.

 

I don't really have much more to add really, as without ninja level DSP knowledge I can't get more into the weeds of what might be going on (and this is only conjecture on my part of course, based on a bit of coding and testing experience).

 

But it's interesting though. Hopefully we'll see M1 performance actually improve over time as developers spend more time coding for it, once they've got past the pain points of the actual transition (which we're still in the middle of).

[exmachina]

SPUMD, might it be a better test if you yanked memory out of the Pro until both testbeds had 16GB?

[SPUMD]

It's not the RAM cause on the M1 Air it takes only 20% in this test.

A test with audio plugins and tracks is missing the point cause it's a CPU test. I never ran into CPU issues in such case, but with highload CPU Software instruments and reverb.

 

I posted this to check how this test performs on other Macs espacially M1s cause i have only two macs to run this test. If there is an issue, there is an issue I don't expect someone in this forum to solve processor issues.

 

But good news .. by changing the processor threads to 8 in Logics audio settings i had an improvement.

67 clean playback tracks with internal speakers as output device and Alchemy as softwre instrument with the Air M1

 

Compared to the 12 core intel Mac Pro it's still about -23 clean playback tracks (-25.5% less performance) but it's a significant improvement!

I contacted Apple and wrote them about the issue. Maybe it helps. But for now it earths my hopes for CPU heaven with a mac studio M1 max.

[des99]

A test with audio plugins and tracks is missing the point cause it's a CPU test.

 

Well, it's not missing the point, audio plugins are DSP processes, which all require calculation on the CPU, and it *might* bypass the optimisation issue and give some insight as to whether that might be having an affect. It's all part of learning the behaviours of these relatively new systems.

 

But good news .. by changing the processor threads to 8 in Logics audio settings i had an improvement.

67 clean playback tracks with internal speakers as output device and Alchemy as softwre instrument with the Air M1

 

This is one of the things I was alluding to in terms of changing the settings from your initial constraints. I see that one setting alone has made quite a difference - I've said elsewhere, we're not exactly sure how "automatic" allocates threads and cores, but a bit of manual help has improved performance for many people. I haven't seen a huge difference, but then I haven't done a huge amount of testing or playing with it yet either - but certainly other people have reported improved performance by leaving automatic.

 

Compared to the 12 core intel Mac Pro it's still about -23 clean playback tracks (-25.5% less performance) but it's a significant improvement!

I contacted Apple and wrote them about the issue. Maybe it helps. But for now it earths my hopes for CPU heaven with a mac studio M1 max.

 

Thanks for sending in the feedback. Maybe you should be thinking about the M1 Ultra then...

[SPUMD]

When testing i find it more helpful to seperate things CPU, RAM, SSD Stream. This gives more clear results and in the aftermath you can sum the impressions.

The M1 Ultra is to expensive for me. Basic setting is twice the money.. for that extra I would rather by a analog synth.

 

Mate, how about the fan noise of your MBP, do you come into situations where it bothers?

[des99]

I’ve never heard the fans once. Fans are off (0 rpm) since I bought it.

 

Even under high loads, internal cpu temps of 86 degrees or so for half an hour at a time, the case is barely warm and the fans are still off.

 

Compared to what I came from, it’s glorious!

 

(And I’m running in clamshell mode almost all the time.)

[SPUMD]

cool

[wonshu]

MacOS: Monterey 12.3

Mac Model: MacBookAir

Prozessor: M1 (8-core / 7-core GPU)

RAM: 16 GB

Logic Version: 10.7.3

 

clean playback: 45 tracks (64 before measure 38) with interal speakers Output Device

[des99]

Mate, how about the fan noise of your MBP, do you come into situations where it bothers?

 

I was just watching an interview with the Apple hardware guys, regarding the new Macbook Pros (M1 Pro+). One point was interesting - he said:

 

"With all the CPUs maxed out, the chip uses around 30 watts. Well, 30 watts is not going to make the fans move. You can beat it up all you want with the CPUs, it's not gonna do anything."

 

I can verify this, as I've done workloads with all the chips maxing out for an hour or two processing files, and the thing was silent, and barely warm to the touch.

 

Of course, there are other things on the chip that also draw power, particularly the GPUs, so if you're maxing out all your GPUs, and you have a lot of them, and you're using ProRes engines etc - then the chip might then require some cooling and so the fans might kick in there. (This is another reason why I specifically choose an M1 Pro - I don't need the extra GPUs, and the chip is physically smaller, so runs less chance of spinning up the fans, which I'm particularly sensitive to.)

 

So for typical audio use (and therefore minimal GPU use), a MBP is not going to spin up the fans, even if you're maxing out the CPU load in a heavy session.

[exmachina]

But good news .. by changing the processor threads to 8 in Logics audio settings i had an improvement.

67 clean playback tracks with internal speakers as output device and Alchemy as softwre instrument with the Air M1

 

Compared to the 12 core intel Mac Pro it's still about -23 clean playback tracks (-25.5% less performance) but it's a significant improvement!

I contacted Apple and wrote them about the issue. Maybe it helps. But for now it earths my hopes for CPU heaven with a mac studio M1 max.

 

Great news! I ran across that tip recently and forwarded it to a friend who had been experiencing performance issues after updating Logic -- I suspect Apple broke the Automatic thread feature a while back and it still appears to be broken. Thanks for reporting it.

[SPUMD]

@des99 thanks for the info

@exmachina you're welcome