Official Serum 2 thread!

[vanerio]

That's your case, not everyone's.
My test: x3 osc (unison x8) -> 24 dB/oct lowpass filter -> x4 notes chord
Results:
Serum 2: 18 % CPU
PhasePlant: 13 % CPU
MSoundFactory 1 % CPU or 2 % with oversampling (yes, 1%, 18 times less)

I tried your test, but slightly changed, on my M1 Pro Mac. 48khz, 128 sample buffer.

2 saw oscillators at 16x unison, into a clean 24 db/oct lowpass filter, playing 4 notes. I just eyeballed this on Live 12's CPU meter. I could take more accurate measurements using Bitwig's DSP graph, but I think this is good enough for a rough approximation.

Dune 3: 2%
Hive 2: 2%
Phase Plant (have to use 4x osc, because of 8x unison limit): 7%
Serum 1: 27%
Serum 2: 12%

Maybe Serum 2 added a smarter unison implementation?

Notes:
- Dune 3, Hive 2, and Phase Plant have dedicated saw oscillators. Serum 1 and Serum 2 don't have dedicated saw oscillators and are always in wavetable oscillator mode.
- Hive 2 is set to "Clean" mode to more closely match Dune 3's default behavior, which is also clean. "Normal" mode uses more CPU than Dune 3. (But choosing a more expensive filter in Dune 3 also makes its CPU go up.)
- Phase Plant's normal (non-non-linear) filter cramps and doesn't sound great, unlike the filters in all of the other synths here.
- Phase Plant's CPU starts going up as soon as you add an envelope modulator for the filter, which isn't the case for Hive and Dune when using the fixed-routing modulation envelopes. Using the mod matrix probably makes the usage go up, though it was too low to measure. Serum 1 and 2's usage goes up less per envelope modulator than Phase Plant. The mod matrix probably gets more expensive in Dune 3 if you set its modulation rate to audio rate.

That is really interesting. I replicate now your test with x2 oscillators and for me it is the opposite:
Serum 1: 20 %
Serum 2: 22 %

I am at 2048 samples buffer...so 128 would fry me up some potatos for dinner.

Edited: I update the CPU % values, as my previous test was with x2 oscillators (x8 voices), now they are (x16 voices). Just 1 instance is pretty hungry in my system even with huge buffer.

This is in line with my previous testing: all my Serum 1 presets require more CPU in Serum 2.

[tumface]

That is really interesting. I replicate now your test with x2 oscillators and for me it is the opposite:
Serum 1: 20 %
Serum 2: 26 %

Another theory is that the ARM (Apple Silicon) version of Serum 1 didn't receive the same level of SIMD hand-optimization than the Intel version did, since the M-series Apple processors arrived somewhat late into the life of Serum 1. But this is just a total guess and might be completely wrong! Either way, both are fine for me on both my M1 Mac and old Ryzen PC. I don't usually choose Serum for big unison sounds or lots of polyphony. Better performance on my old AMD processor would be nice, but it's still usable as it is already, if I'm careful.

This is in line with my previous testing: all my Serum 1 presets require more CPU in Serum 2.

The same for me on my Ryzen PC. But not on my M1 Mac, where Serum 2 is the same or better than Serum 1.

[vanerio]

That is really interesting. I replicate now your test with x2 oscillators and for me it is the opposite:
Serum 1: 20 %
Serum 2: 26 %

Another theory is that the ARM (Apple Silicon) version of Serum 1 didn't receive the same level of SIMD hand-optimization than the Intel version did, since the M-series Apple processors arrived somewhat late into the life of Serum 1. But this is just a total guess and might be completely wrong! Either way, both are fine for me on both my M1 Mac and old Ryzen PC. I don't usually choose Serum for big unison sounds or lots of polyphony. Better performance on my old AMD processor would be nice, but it's still usable as it is already, if I'm careful.

This is in line with my previous testing: all my Serum 1 presets require more CPU in Serum 2.

The same for me on my Ryzen PC. But not on my M1 Mac, where Serum 2 is the same or better than Serum 1.

Exactly, I also opt for other synthesizers for polyphonic or unison sounds. As in my first test, MSoundFactory always does the job magically. It seems that adding unison voices in MSoundFactory lowers CPU consumption for some reason haha!

[Gam456]

If you mean when you press stop on the transport then have you unchecked the 'Silence not + FX tails when host transport stops'?

The fx tail is still broken in Reaper
Basically, fx will stop at the end of the clip.

I don't stop anything, like I said at the end of the clip. Seum is the only vsti with this behaviour.

[User3]

How to make the panning automatically follow only the release curve of the ADSR envelope in Serum 2 (but not the other parts of the ADSR envelope)?

Set the ADSR envelope to 'BMP'. Pick an LFO to modulate the pan and set the delay time (the tine it takes for the LFO to kick in) to match the beat division of the decay stage of the ADSR envelope.

EDIT: maybe this screenshot will help clarify..

Thank you. This is not quite what I need.

[jrwaltb]

That is really interesting. I replicate now your test with x2 oscillators and for me it is the opposite:
Serum 1: 20 %
Serum 2: 26 %

Another theory is that the ARM (Apple Silicon) version of Serum 1 didn't receive the same level of SIMD hand-optimization than the Intel version did, since the M-series Apple processors arrived somewhat late into the life of Serum 1. But this is just a total guess and might be completely wrong! Either way, both are fine for me on both my M1 Mac and old Ryzen PC. I don't usually choose Serum for big unison sounds or lots of polyphony. Better performance on my old AMD processor would be nice, but it's still usable as it is already, if I'm careful.

This is in line with my previous testing: all my Serum 1 presets require more CPU in Serum 2.

The same for me on my Ryzen PC. But not on my M1 Mac, where Serum 2 is the same or better than Serum 1.

Exactly, I also opt for other synthesizers for polyphonic or unison sounds. As in my first test, MSoundFactory always does the job magically. It seems that adding unison voices in MSoundFactory lowers CPU consumption for some reason haha!

I am wondering if this is because of what a previous poster said about it not reporting the CPU to the DAW in the same way as the others. I have it and I guess I can test on my M1, can you share the preset? (If you want, not a big deal either way.) Actually should we all standardize the test on a preset for each synth, I have them all except Dune 3?

[vanerio]

That is really interesting. I replicate now your test with x2 oscillators and for me it is the opposite:
Serum 1: 20 %
Serum 2: 26 %

Another theory is that the ARM (Apple Silicon) version of Serum 1 didn't receive the same level of SIMD hand-optimization than the Intel version did, since the M-series Apple processors arrived somewhat late into the life of Serum 1. But this is just a total guess and might be completely wrong! Either way, both are fine for me on both my M1 Mac and old Ryzen PC. I don't usually choose Serum for big unison sounds or lots of polyphony. Better performance on my old AMD processor would be nice, but it's still usable as it is already, if I'm careful.

This is in line with my previous testing: all my Serum 1 presets require more CPU in Serum 2.

The same for me on my Ryzen PC. But not on my M1 Mac, where Serum 2 is the same or better than Serum 1.

Exactly, I also opt for other synthesizers for polyphonic or unison sounds. As in my first test, MSoundFactory always does the job magically. It seems that adding unison voices in MSoundFactory lowers CPU consumption for some reason haha!

I am wondering if this is because of what a previous poster said about it not reporting the CPU to the DAW in the same way as the others. I have it and I guess I can test on my M1, can you share the preset? (If you want, not a big deal either way.) Actually should we all standardize the test on a preset for each synth, I have them all except Dune 3?

As I said, no, it is not that CPU consumption is unreported.

Preset (I used a 4 note infinite chord):

Code: Select all

$eNrtWMtuGzcUvXYD9LHLHwjIokARYDQj+bFpHVqPRKhlTzyOPe2OmrmWCFOkQHJk2YugP9BVV+2qQFf5k3bZoEB3RT+hBQp0UyAF5yGPHMuWp12GC2F47zm85OHrUtDfDcN+IBMRd2lkpLrQaAwTQw3kEbhr5NFTLgeU6yM6SKuk3caJGUEMafl9xyJYjD6qg8jQKYIL5BiVZlJAEx48BPLs+YuJpuMJZ2JovZ2YGakCdokdqi9Au4164-G2u132gHabde-x9kYTiM+okMcMz1uSczrRGNtW+nS21z3Y5TI6SwlujfhUUc6Rs0tqbHh3jewdBhPOjMG0VTfv9o-fPLnm8YoB-bFDfJ4MmQC9IAuEQTLIPHNLVg3mggXIMTIYFw3UH9dzFd-ubE7WRStAQjuqL+zPPh0j6AMdMc6pkQpIN+H8urWWiqDPegJcCHepOKuD81afskl8IZiWoiPogONbhgVZc08bTSIQ4u++teWnncLelzGC3pe1AdVYmzATjeYcf0Q1gj6kIpbj2jmLce56xoajvEkb3T-pZw21mbY9iCG1FXMAN81uGIzoBIFMKU+wQH7yhARHHd8-7ASdIyvDl1KOwXEcpyzr2r1k9Uqypl8HiQF3-Z3Ctyi8fi+FGzcq-OCdwrconB4Na7nCfTbDBXFTgz0QemKSmJZMhAH3vbnK7+dfgUHKbTWF9USMszoUrNRZt0GuvO68mnrdRa9nJ-7K61nvQWJK5Dflugfum9vnuN9r93wlh4qOWyMqhljMNpAFc4saHErFLlHdfBXcoN16rl2XcbMoXmap3XmIkvEk60+2yJ5D-NlHv-4w+fT1Dnl0dDFB0Ht+zWvadWDXVLY6WlJoQ4WBqoNrJUqhML5CjabXBt2mUxa-bOMpTbipHaKWgooIb1xvRFB+cYnKbQAp9T2QiYrQahKMpTQjgVoX++TVDjmkTAzk+ZywRk6YiOV5NshnVAggZMZ0S3KpQNfzAuSpYvF1Y3gVoiOmQDr7x8VSX3dCX+FpKr-NDUixM9WYcgh9yaxuM4i-+u3jXx78-NcO6XI61Lvgfu2UvN---eEGaT58UvI6IfHrpSGvER-pmbYfwUiep308tpvQmpyslC748H+6222rxBganblA9lAMzcjeJB0xRS7tQUC0ZsP5MliQpmB6lZmNysxmZeZGZeZmZeZWZeZ2BaavZIRau1AilCAfXEG8uyGNuyHNuyEbd0M274Zs3Q3ZvgXSEVMXwiOFY8llfs-dfK9ZqLc6tLE6tLk6dGN16Obq0K3VodsrQrP3V3p0Zou2lGEcIkebr8T-vPrzdX-g75A2RvSihAgSbSgTV6+b1YLudQ-c6-tgAVraNznJcrwKnEYFTrMCZ6MCZ7MCZ6sCZ-u+nCw-TfcbFfrUXpjpi3f50fXBnOVVYjUqsZr3ZH2OFzqiHN2KvPuO7Ri5jJipFnSBXDXyNcdVIzbh04nC5QCbZO6iiMOlkL6MT0aIfCngiI0HCreaSwG7CqkZ1b2lgM5sYjvKpHDdpaDiFFrm35cGD05Pr-vTP5byg2rR05LCKMk5KpuaMTHkaJNngRxccEISGTbFLGXO-lWyiVobpyzCPhpUuvReXCNF5Uihnaowz8AvegbHPXEqwZmbAkMNQvgUhUJwwkw-G1EZFiU8m3knJNxm8GY0BifsIjWJQpsNhnkyX7yUYWno-xInHXD+Sk4zUD9Xwn6rCZBD27a7HvapUWx25cyk8SA8YbEZ9SHM3wlkIUt3wXGcl-8C7TlL+g==

Just to avoid confusion. Unison in 'No base pitch' mode works symmetrically, so 4 voices means 8 voices (see image)

x8 voice unison.png

But you know what's even creepier? Just turn unison voices to x32 in each oscillator (that would be 192 voices)...CPU will only raise +1-2%

[tumface]

I have it and I guess I can test on my M1, can you share the preset? (If you want, not a big deal either way.) Actually should we all standardize the test on a preset for each synth, I have them all except Dune 3?

OK. here are all of the presets I used. I saved them as VST3 .vstpreset for easier loading/compatibility:

2x 16x saw 4pole lp presets.zip

Dune 3, Hive, Phase Plant, Serum 1, Serum 2.

I didn't match them all exactly sonically. But the settings should be roughly equivalent, except for Phase Plant, because Phase Plant doesn't allow unison to go beyond 8x. So, it has to use 4 oscs instead of 2.

Also, for the Serum 2 preset, I remade it from a blank init preset instead of loading it from a Serum 1 .fxp, to see if it made any difference, in case there were any additional compatibility flags that get set or something. It didn't make a difference.

This is more of a unison voice test than a normal performance test. Usually I would use more synthesizer features and fewer unison voices, even in Dune or Hive. So, I don't think it's a great test of overall performance. Only unison.

[tumface]

Just turn unison voices to x32 in each oscillator (that would be 192 voices)...CPU will only raise +1-2%

This is similar for Hive and Dune. And probably a bunch of other synths. They have dedicated algorithm implementations for detuned oscillator unison. Raising Dune 3 to 32x for both of its oscs only adds 1% CPU usage for me in that preset I just uploaded.

There are tradeoffs with doing this. For example, in Hive and Dune[1], you can't adjust the wavetable position per unison voice. But you can in Serum.

1. Dune has a second layer of unison features which let you do that, but it's an entire full voice path (with filters) instead of just stacked oscillators summed before the filter.

[swilow11]

The fx tail is still broken in Reaper
Basically, fx will stop at the end of the clip.

I believe there is a global setting to cut FX tails when playback has stopped....

[Gam456]

The fx tail is still broken in Reaper
Basically, fx will stop at the end of the clip.

I believe there is a global setting to cut FX tails when playback has stopped....

Ohhh. My bad. I didn't know that. Thank You.

[SirKen]

Am I understanding things right that Intel implementation was more through than the AMD implementation? I wouldn't mind grabbing it at the intro price but I don't want to gamble on a future update that might fix the CPU issue, might being the key word.

[tumface]

Am I understanding things right that Intel implementation was more through than the AMD implementation? I wouldn't mind grabbing it at the intro price but I don't want to gamble on a future update that might fix the CPU issue, might being the key word.

Intel and AMD use the same instruction set architecture. In other words, they're compatible with each other and use the same machine code.

The differences lie in the different generations and models of the chips produced by both companies. Different hardware microarchitectures, which are compatible with the same code as each other, might have different characteristics. For example, on some microarchitectures (generations of chip) if you have instruction sequence XXX then YYY then ZZZ, it's slower than on a different microarchitecture. But then on that one, a different combination might be slower. And some microarchitectures only support some subset of extended instructions with extra capabilities.

This table shows a bunch of instructions and their latencies (latency in nanoseconds, as in data dependency chains, not audio latency) https://asmjit.com/asmgrid/ I sometimes use it when working on SIMD code.

By the way, I have no knowledge of the internals of Serum 2. My guesses are all just guesses. I could be totally wrong.

I wouldn't mind grabbing it at the intro price but I don't want to gamble on a future update that might fix the CPU issue, might being the key word.

Even on my old Threadripper PC, where Serum 2 performs significantly worse than Serum 1, Serum 2 is still totally usable. There are some presets that are unusable, but I wasn't going to use those presets, anyway. Serum is one of my favorite synths to *make* presets in, so that's what I'm usually doing.

Also, Serum 2 has no problem idling when it's not in use, so as long as you aren't using a bunch of instances all at the same time, no problem.

[SirKen]

Am I understanding things right that Intel implementation was more through than the AMD implementation? I wouldn't mind grabbing it at the intro price but I don't want to gamble on a future update that might fix the CPU issue, might being the key word.

Intel and AMD use the same instruction set architecture. In other words, they're compatible with each other and use the same machine code.

The differences lie in the different generations and models of the chips produced by both companies. Different hardware microarchitectures, which are compatible with the same code as each other, might have different characteristics. For example, on some microarchitectures (generations of chip) if you have instruction sequence XXX then YYY then ZZZ, it's slower than on a different microarchitecture. But then on that one, a different combination might be slower. And some microarchitectures only support some subset of extended instructions with extra capabilities.

This table shows a bunch of instructions and their latencies (latency in nanoseconds, as in data dependency chains, not audio latency) https://asmjit.com/asmgrid/ I sometimes use it when working on SIMD code.

By the way, I have no knowledge of the internals of Serum 2. My guesses are all just guesses. I could be totally wrong.

I read on another forum that they have not tested it properly on AMD chips at all. Don't know how true that is.

I wouldn't mind grabbing it at the intro price but I don't want to gamble on a future update that might fix the CPU issue, might being the key word.

Even on my old Threadripper PC, where Serum 2 performs significantly worse than Serum 1, Serum 2 is still totally usable. There are some presets that are unusable, but I wasn't going to use those presets, anyway. Serum is one of my favorite synths to *make* presets in, so that's what I'm usually doing.

Also, Serum 2 has no problem idling when it's not in use, so as long as you aren't using a bunch of instances all at the same time, no problem.

Honestly, that just sounds like coping to me. Especially if there really is a difference in performance between different chip vendors. They need to do better than that if they want to sell a synth for near 200 and without a license transfer.

[tumface]

I read on another forum that they have not tested it properly on AMD chips at all. Don't know how true that is.

That sounds made-up to me. It works fine on AMD.

Honestly, that just sounds like coping to me.

I'm able to use it without any significant problems. The plugin loads fast and works fine. I've reported some bugs and they've fixed all of them so far. I don't know how that's supposed to be coping.

And on my Mac, where I do most of my audio stuff, it runs even better than Serum 1.