Seems to me the “multiple models running in series behind the scenes” could be better handled in terms of resources. Loading them all takes ages, then some systems crash. Maybe the models should be loaded and run more dynamically in memory, even though it would be slower i.e. load vocal model, run it, unload it, load drums model, run it, unload it etc etc. That’s essentially what has to be done when using LaLaLAI (for comparative example)
You can add me to the list of crashes during unmix song; Ryzen 5950X, 64 gB RAM, RTX4070 with 12 gB VRAM, quality set to “HIGH”, running WIN 11 Pro. Initially SL reports approx 1.5 minutes to unmix a test track, but then process monitor shows VRAM use spiking to 100%, whereupon estimated time to complete spikes to about 6.5 minutes, then SL crashes a few seconds afterwards. Setting quality at “Balanced” does not crash, however quality of vocal and guitar track separations is noticeably poorer than what I could get with previous SL versions with quality set at “High”.
A GTX 1660 has only 6GB of VRAM. Anything with AI model will eat a lot of VRAM - the larger the model the more precise the results, but at the cost of memory.
Either get a GPU with much larger VRAM or use less precise and smaller models.
That your GPU is able to run all those other tools without issues unfortunately means nothing when it comes to AI compute.
I would think a new card should have at least 16GB of VRAM.
A Nvidia “RTX 5070Ti” would be the sweet spot spec-wise, it has fast memory bandwidth, is close in speed to a RTX 5080, but cheaper.
You can have a look at Pugetsystems, they do a lot of benchmark tests with different cards and different software packages.
A good alternative would be to buy a used “RTX 4070Ti Super” with 16GB VRAM, I don’t think there is much difference in speed to the 5080Ti.
16 GB will not be enough. In my latest tests, SpectraLayers will take more than 17 GB VRAM if you use all stems on an Unmix Song process. Of course, you can reduce the number of selected stems to reduce the VRAM amount and then run the Unmix Song multiple times on the resulting Other stems.
If you only select a single stem, you might even get it to fit into your GTX 1660 6 GB VRAM. Alternatively try pure CPU processing, which is much slower, but should work.
If I understand this well, I need to purchase Spectraleyers for $349.99 and then an $800.00 card only to see if it works? And if it doesn’t, then just keep buying more expensive cards until it works correct?
I’m not trying to be sarcastic. I’m just trying to get to the amount I need to put together to be able to use the software.
You dont need to spend that kind of money if you are only using Spectralayers for unmixing, instead of spectral editing.
You can do very high quality unmixing for very little money by using good online services such as Lalal.AI
If you want a spectral editor which can also do unmixing, then maybe the monetary investment in Spectralayers makes sense.
The recent 12.0.30 release still didn’t fix it for me.
I just ordered a 5070Ti 16GB. Will report back on Friday when it arrives.
P.S. If anyone is looking at upgrading, check memory bandwidth specs (which are important for AI processing). I currently have a 3070Ti and was about to pull the trigger on a 5060Ti 16GB. The memory bandwidth was less than with my 3070Ti so I bit the bullet and went for the 5070Ti instead.
OK, I have to talk about the updating of the GPU in an older PC. If you are running an 8G DDR5 PCIe x 16 3.0 for example, and you buy a 16G DDR6 PCIe X 16 4.0 (NOTE: 4.0) this will only deliver 8G to your PC. To update the PCIe x 16 3.0 to say 12G, this means you would have to buy a 24G PCIe 4.0 card. The 4.0 video card in a 3.0 bus equipped PC can only deliver 1/2 of the G that it has. (The 5.0 card with DDR7 - the new GPU cards - is the same story.)
Go to ‘Tom’s Hardware’ site (Tom’s Hardware: For The Hardcore PC Enthusiast and dig around a bit, you’ll learn about this. Unfortunately Intel and the other PC MB makers not only change the CPU processor Slot, but also change the bus speed of the PCIe slot. Don’t shoot the messenger.
Reporting back. 5070Ti 16GB works for unmixing everything on High setting, that is until it’s half way through then it crashes. 15.5GB/16GB usage roughly.
HIGH: Drums, Bass, Vocals:
Works (10.3GB/16.0GB VRAM usage)
HIGH: Drums, Bass, Vocals, Piano
Works (15.1GB/16.0GB VRAM usage)
HIGH: Drums, Bass, Vocals, Piano, Guitar
Works (15.1GB/16.0GB VRAM usage)