Zebra 3 Public Beta (final beta)

[DCrown]

If all the people who take this craft somewhat seriously just got a decent processor, the developers wouldn’t have to cut corners for sound quality so much anymore. It’s 2025, processors can take it.

If every plugin works without cpu issues but Zebra 3, what does it tell you?
Maybe no issues for you, but I doubt Zebra 3 will be released just 4 U!
As the last u-he video shows, Urs is offering solutions concerning high cpu usage and going to improve, so CPU obviously is a serious topic.
What a snobbish post.

[werter318]

If all the people who take this craft somewhat seriously just got a decent processor, the developers wouldn’t have to cut corners for sound quality so much anymore. It’s 2025, processors can take it.

If every plugin works without cpu issues but Zebra 3, what does it tell you?

That those developers make sure to take people’s shitty processors into account more.

[DCrown]

If all the people who take this craft somewhat seriously just got a decent processor, the developers wouldn’t have to cut corners for sound quality so much anymore. It’s 2025, processors can take it.

If every plugin works without cpu issues but Zebra 3, what does it tell you?

That those developers make sure to take people’s shitty processors into account more.

Ok, snob.
Nevertheless I will buy, even though I probably won't be able to use all presets.

[werter318]

If all the people who take this craft somewhat seriously just got a decent processor, the developers wouldn’t have to cut corners for sound quality so much anymore. It’s 2025, processors can take it.

If every plugin works without cpu issues but Zebra 3, what does it tell you?

That those developers make sure to take people’s shitty processors into account more.

Ok, snob.

CPU matters, obviously. But designing everything around the lowest common denominator limits progress. Better sound often does cost more CPU that’s just reality.

[PieBerger]

If all the people who take this craft somewhat seriously just got a decent processor, the developers wouldn’t have to cut corners for sound quality so much anymore. It’s 2025, processors can take it.

If every plugin works without cpu issues but Zebra 3, what does it tell you?

That it's new and awaiting optimisation. When Diva was first released it melted CPUs and people complained about it, then 3 years later when Serum 1 came on the scene, it melted CPUs and people complained about that too. Both of these synths are very resource efficient in 2025, and allow plenty of instances even at high/max quality settings, so it stands to reason that after a reasonable period of time, Zebra3 will also become more efficient and less demanding on *older* hardware therefore.

If you want to be on the bleeding edge, you have to keep your hardware up to date, that's just how it is. If you can't do this for any reason, you have to suffer a bit in the short term, compromising quality and or workflow until the inevitable optimisation comes or you can afford/justify upgrading your hardware.

[whassup]

Yeah. We had that with Diva. Quite successful anyways.
If I had to choose between new sounds with better quality or more voices/instances...
And this is coming from a guy who used Hive for several years almost exclusively because of CPU restrictions.
So this is my current preference backed up by a rather modern CPU.
YMMV.

[Garrincha]

Wasn't that the night we talked about phasers with 40 stages?

Exactely

I kept my Boss SE-70 for that reason although I never actually use it anymore.

Next thing I'd like to explore in Zebra is a combination of additive and FM. That should really be fun. To me Zebra always was a bit like Synclavier and Fairlight. In one moment the dream maschines for signature sounds and sprinkling fairy dust over a project and in the next moment it delivers the complete track all on it's own.

[werter318]

If all the people who take this craft somewhat seriously just got a decent processor, the developers wouldn’t have to cut corners for sound quality so much anymore. It’s 2025, processors can take it.

If every plugin works without cpu issues but Zebra 3, what does it tell you?

That it's new and awaiting optimisation. When Diva was first released it melted CPUs and people complained about it, then 3 years later when Serum 1 came on the scene, it melted CPUs and people complained about that too. Both of these synths are very resource efficient in 2025, and allow plenty of instances even at high/max quality settings, so it stands to reason that after a reasonable period of time, Zebra3 will also become more efficient and less demanding on hardware therefore.

If you want to be on the bleeding edge, you have to keep your hardware up to date, that's just how it is. If you can't do this for any reason, you have to suffer a bit in the short term, compromising quality and or workflow until the inevitable optimisation comes or you can afford/justify upgrading your hardware.

You’re right, but you have to realize many of those optimizations don’t come for free either. Even things like more aggressive inlining or vectorization can change numerical behavior, summation order, or timing in subtle ways. Most users won’t notice or care, but if you’re chasing maximum sonic stability and feel, those details can matter. Again I go way too technical for a forum but I am just all for pushing digital, that's all. Do with it what you want.

[Benjamin923]

Can anyone point to me some good examples of CPU heavy/melting patches that showcase the quality? I want to get a good sense of how Z3 fares on my CPU and also what kind of quality comes with the increased load

[Urs]

You’re right, but you have to realize many of those optimizations don’t come for free either. Even things like more aggressive inlining or vectorization can change numerical behavior, summation order, or timing in subtle ways. Most users won’t notice or care, but if you’re chasing maximum sonic stability and feel, those details can matter. Again I go way too technical for a forum but I am just all for pushing digital, that's all. Do with it what you want.

There are a few things left that should run faster, probably before the 3.0 release.

I usually vectorise and inline early on. An unoptimised build runs with maybe 1 or 2 voices of a simple patch, but nothing elaborate. I.e. we can't really work on sonic aspects without optimising from the start.

There's a minimal chance that hand-coded AVX2 can speed up the filters. Strangely, in the additive renderer it had the opposite effect while in the curve renderer it helped big time. So I might end up trying AVX2 for the filters at some point, but as the chances may be minimal, I won't prioritise it just yet. So this is maybe something coming down the road.

Another option is multithreaded rendering. This is bit tricky as stuff like oscillators and modal resonators need access to cache lines of shared data, and it's difficult to assess when stuff is copied or written to. So multicore generally works, but there are some issues that made it unusable for now. We might revisit that later as well.

So yeah, quite a bit of potential before the initial release, and then some hope for possible improvements later on.

[Urs]

Can anyone point to me some good examples of CPU heavy/melting patches that showcase the quality? I want to get a good sense of how Z3 fares on my CPU and also what kind of quality comes with the increased load

Just use the Yellow filter on one or two oscillators. I think it sounds amazing, but the price in terms of CPU is steep.

If you want a reference, compare modules of Zebra 3 to modules of Zebra 2. Every single comparable module in Zebra 3 is at least twice as expensive on CPU, but offers objectively better sound quality, sometimes higher temporal resolution etc. Filters from Zebra 2 that can't do audio rate modulation can do so in Zebra 3. Envelopes can be modulated dynamically in Zebra 3 while in Zebra 2 the modulation was always "frozen" during stages. The ModMatrix in Zebra 3 is a completely different thing with its modifiers, and so are FMOs which are now dual oscillators.

The main culprit of excessive CPU usage in Z3 are heavy oscillator effects when run at high resolution, and Diva-quality filters, which are stereo in Z3. People wanted Diva-quality filters, and they got them. So here we are

[werter318]

You’re right, but you have to realize many of those optimizations don’t come for free either. Even things like more aggressive inlining or vectorization can change numerical behavior, summation order, or timing in subtle ways. Most users won’t notice or care, but if you’re chasing maximum sonic stability and feel, those details can matter. Again I go way too technical for a forum but I am just all for pushing digital, that's all. Do with it what you want.

There are a few things left that should run faster, probably before the 3.0 release.

I usually vectorise and inline early on. An unoptimised build runs with maybe 1 or 2 voices of a simple patch, but nothing elaborate. I.e. we can't really work on sonic aspects without optimising from the start.

There's a minimal chance that hand-coded AVX2 can speed up the filters. Strangely, in the additive renderer it had the opposite effect while in the curve renderer it helped big time. So I might end up trying AVX2 for the filters at some point, but as the chances may be minimal, I won't prioritise it just yet. So this is maybe something coming down the road.

Another option is multithreaded rendering. This is bit tricky as stuff like oscillators and modal resonators need access to cache lines of shared data, and it's difficult to assess when stuff is copied or written to. So multicore generally works, but there are some issues that made it unusable for now. We might revisit that later as well.

So yeah, quite a bit of potential before the initial release, and then some hope for possible improvements later on.

AVX2 is usually exactly where things can start going downhill sonically, depending on the algorithm, mainly due to changes in summation order, contraction, and reduced numerical determinism compared to narrower paths like SSE. I know you're fully aware of these things and have to deliver a somewhat CPU-friendly product though.

[Urs]

AVX2 is usually exactly where things can start going downhill sonically, depending on the algorithm, mainly due to changes in summation order, contraction, and reduced numerical determinism compared to narrower paths like SSE. I know you're fully aware of these things and have to deliver a somewhat CPU-friendly product though.

As we're getting nerdy here: We have unit tests to make sure that the results are within narrow boundaries of error. We also commonly don't use estimates, so the results of most operations should be identical.

[werter318]

AVX2 is usually exactly where things can start going downhill sonically, depending on the algorithm, mainly due to changes in summation order, contraction, and reduced numerical determinism compared to narrower paths like SSE. I know you're fully aware of these things and have to deliver a somewhat CPU-friendly product though.

As we're getting nerdy here: We have unit tests to make sure that the results are within narrow boundaries of error. We also commonly don't use estimates, so the results of most operations should be identical.

Yes and sadly that within narrow boundaries is purely subjective. Obviously hand-coded AVX2 is nothing like standard compiler AVX2 I'll give you that, but for sensitive ears it will still make a difference.

[Urs]

for sensitive ears it will still make a difference.

You can write

!INSTRUCTION_SET=SSE4.2

in a text file and save it as "default.h2p" (with h2p being the actual file suffix) into the root preset folder.

Then, wherever we switch instruction sets dynamically, it'll use SSE instead of AVX, AVX2 or AVX512.

Let me know if you find audible differences.