I am referring to Apple, not Windows, and who cares about Snapdragon? No professional audio engineer or studio is going to use Snapdragon. Your fascination with this hardware is puzzling, especially when the overwhelming majority of studios and engineers use Apple Macs, which are being consistently favoured over this folly.
Both DEMUCS-GUI and rebeat already utilise Apple GPUs. DEMUCS will unmix to six stems in a fraction of the time that SpectraLayers 12 takes. Against version 12.0.20, the difference was striking: DEMUCS completed the task in under a minute, compared with over 30 minutes in SpectraLayers.
So the question remains: when will you be supporting Apple GPUs?
@ondre Demucs GUI is not competing with SpectraLayers, itâs rather a GUI wrapper for a Pytorch model, with no other feature than that.
Demucs GUI itself is not doing any GPU acceleration, itâs entirely done by the Python+Pytorch environment they silently install under the hood, Pytorch doing the GPU acceleration. This environment needs to be downloaded and takes several GBs on your system, which is not removed when removing Demucs, and this is even before counting the AI models. Itâs not really a technical solution that would be suitable for a product like SpectraLayers.
Thereâs a reason why Demucs GUI is not capable to have such models running inside the Demucs GUI application itself. Itâs really not that simple to export, use and accelerate models outside the whole Pytorch environment.
as you say PyTorch uses the MPS/GPU, and I compared apples with apples, same track, 1/30th of the time. The download on my system is considerably smaller than the 5gb+ that SL12 takes.
Why wonât you answer the reasonable question âwhen will you be supporting Apple GPUs?â
Iâm not trying to stir confusing but may help here (and possibly clear up a misunderstanding). I donât believe with RX in-particularly that there is acceleration going on but believe it is more-or-so optimized.
Simply because I donât have any reliable answer to that question, it would be unreasonable to make predictions when thereâs no simple technological solution to it, without relying on an entire external Python/Pytorch environment that would need to be installed somewhere on your system, and would not be integrated directly within SL.
Thereâs a reason why the closest competitors to SpectraLayers, RX and Acoustica, donât provide any GPU acceleration at all. These hardware technologies are highly complex when you need to have them built in the software itself.
The question of GPU acceleration on mac with SL is not new. And itâs actually already there, as Iâve said several AI modules use GPU acceleration on mac already (not sure why you ignore this ?). For instance Unmix Components, Unmix Drums, DeClip, DeBleed, DeReverb, Unmix Transcription, all use some degree of GPU acceleration on mac.
But stems separation models have a degree of complexity that simply doesnât run on CoreML, the mac acceleration technology that SL primarily use. Apple did not design CoreML to handle tensors that have dimensions higher than 16384 in size, which several AI audio models use. Every year I talk to Apple in hope they fix their AI model converters for complex models so I could use other Apple technologies, but they simply donât. And Iâve been in talk with them about this for at least the past 3 years.
If you want better AI models support on mac, you can help by asking Apple for better AI model converter support.
I am not ignoring anything, all I use SL for is STEM unmixing, so I am genuinely puzzled as to why you are enabling it for lesser-requested features but not adding it to the part of SpectraLayers that needs it most. If you did not think it was beneficial or necessary, you would not have added it to the PC version several years ago. It seems that the Mac is at best an afterthought, while you allocate finite resources to the frankly ridiculous Snapdragon project.
Demucs-GUI is a very simple programme to use, so I am equally puzzled as to why you appear to blame Apple when the tools are already there. If you can use it for âUnmix Drumsâ, âDeClipâ, âDeBleedâ, âDeReverbâ, and âUnmix Transcriptionâ, then why can it not be used where it is needed mostâŚ..The âUnmix Songâ process ??
He answered that already, even though somewhat convoluted. For some reason they donât want to install Pytorch and a Python environment alongside SpectraLayers and want to run everything in SpectraLayers itself. Since they donât use Pytorch they need to implement it with an Apple ML-framework which doesnât support the necessary requirements.
Iâm not really sure why they donât want to include Pytorch and Python though. Maybe Steinbergâs management simply said no?
My Spectral layers has never crashed on PC. except that one time months ago i ran it out of Cubase and it became unresponsive, i think that was V11, not v12
i9 13900ks, 32 gig Ram, RTX 4060 Ti with latest Nvidia drivers
I used a four piece (sax, drums, bass, piano) Stanley Turrentine recording into SL 12.0.30 with my older PC, set the program to use the AMD card, selected the âBalancedâ setting, and selected âUnmix songâ.
The AMD card used 7.8G of the 8G available, 16.6G of my 64G DDR4 ram, and 1% to 12% of my Intel i7 CPU. Time was 11 minutes.
Ignoring the time it took, the audio quality was exceptional. For example, the sax and piano separations are the best Iâve ever heard from SL. For that matter all four of the instrument layers could be used for quality stems, no problem.
Second song, same settings, John Mayer, Bold As Love (from the Continuum album). Guitars, bass, drums, lead vocal with some doubled vocal sections and some ooh wahs. 5 minutes and the results were again, exceptional, clean stems all.
@ondre Iâm not sure why you think WinARM had higher priority than mac ? I only mentioned WinARM 2 times here since the release of SL12, yet you draw this strange conclusion ?
macOS is actually my main development platform, thatâs where 95% of SpectraLayers code is developed. So Iâm well aware of its contraints. I already explained the reasons why stems models are not yet accelerated on mac, so I invite you to read again my previous post.
For what itâs worth, ReStem (previously ReBeat) on PC does not use GPU.
ReBeat installs a local Python 310 instance.
As it doesnt use GPU itâs very slow, however results are fantastic.
I love being able to gate the separated results and use Drumagog samples if I want.
By far the best drum unmixer Iâve used, but thatâs all it does.
Hasnt crashed or misfired once since I bought it just after first release.
I dont mind slower unmixing if the results are excellent and the app is really stable.
You have repeatedly referenced Snapdragon in your replies in several (separate) posts, since V12 was announced, which is precisely why I conclude that you have wasted valuable resources on a folly side project instead of addressing a far more important platform. It feels as though you have taken the loyalty of your core MAC users entirely for granted. GPU unmixing has been available on Windows for more than two years. I know this, because I was a Windows user when it launched and it was nothing short of groundbreaking. Yet here we are, still waiting for parity. V12 was un-usable at launch, itâs slightly better now but still absurdly slow compared to other products that do a similar thing (and with similar results). The least you could do is give a firm undertaking of when it will be delivered on Mac AND without attempting to charge extra (waiting for V13/14 for example). The disparity between the two major platforms has been unacceptable for far too long. I feel less bad wasting 60s on an unmix thatâs unusable compared to 10-30 minutes then finding that SL12 did no better and I had to bin it anyway. IF your product was leaps ahead of the competition in unmixing Iâd be less critical but its not. Itâs just a lot slower to unmix a song and the stem results are no better than the competition so itâs just 10-30 minutes wasted.
As an aside, and out of curiosity, is there GPU or NPU unmixing on the Snapdragon platform?
For users with low video memory, maybe deliver a workflow which makes it easier to batch multiple unmixing passes in series, where each pass is a separate unmix model (ie pass 1 = vocals. pass 2 = drums etc etc). Between each pass, video memory is fully unloaded. Only 1 model occupies the video memory at a time. AFAIK, this seems to be what happens behind the scenes now, but all the models are pre-loaded at once, hence the long pause before GPU processing starts. The video memory fills up and stays that way. Maybe if the memory only ever contains a single model, that could be useful to users with low-ish video memory.
