I understand that rendering is faster with a GPU than with a CPU – if you have the appropriate GPU memory.
I only have an NVidida GeForce GT 1030 with 2 GB. Until now, DAWs have always said that graphics are not important and recommended a passive graphics card due to the volume.
What I don’t understand is why, with SL 12.0.10, a 5-minute song took 5 minutes to de-mix, and with the new versions (currently 12.0.40) it now takes almost 20 minutes? The CPU or GPU hasn’t slowed down. Why is that?
I have asked a few times where to download 12.0.10, but have not received a reply.
I have backups of most of the old versions of Steinberg applications. Not sure if I’m allowed to distribute them here on the forum. If not, mods please delete my comment.
A hint for future reference, if you have a spare USB drive, plug it in before opening the Steinberg Download Assistant, then in the settings, point it to the USB drive. That way you will have copies of all the installers, and avoid filling up your local drive.
I usually download from the website and bypass the Steinberg Download Assistant. I get the Mac and Win versions so I have both in case I want to switch to Mac later on. Always useful to have them all.
In “Edit | Preferences | System”, the option “Use Shared Memory” seems to be unchecked by default in 12.0.40
Can you try checking that box, and restarting SpectraLayers?
Thanks. And yes, only the current version is on the website. I check the downloads religiously and always back up if there’s a new one, so I have an archive of all previous versions.
I did another comparison with a 06:48 song. Here are the unmix times:
SL 11 = 03:17
SL 12.0.10 = 18:58
SL 12.0.40 (unchecked shared memory) = 15:06
SL 12.0.40 (checked Shared Memory) = 22:16
I’m no longer sure whether it was the 12.0.10 or 12.0.0 version that ran faster, or whether I confused it with SL11. Do you still have the 12.0.0 version?
The reason why rendering is faster with a GPU is because if the trained models can be loaded into GPU VRAM, access is much faster as data transfer rates are higher in VRAM than in standard computer RAM. And the GPU is designed do massive parallel calculations.
Yes, that has been true until neural networks and massive GPU render power became a thing. Now any advanced audio manipulation benefits from modern GPUs.
I think you have mixed that up with SL 11.
SL 12 is using several larger trained models in a cadence and this will need more RAM.
The only solution is to get a modern GPU with at least 12 GB of VRAM (RTX 5070Ti). The more the better if you plan to make productive use of SLP. Otherwise it will be a waiting fest.
That’s why I recently bought a 5070Ti 16GB with a 256bit memory buss. Was thinking about the 5060Ti but the memory buss is a lot narrower, so I opted for the better card. Works so well for AI tasks.
With version SL 12.0.0, the unmix process took 11:28. That’s significantly slower than with SL 11, but still significantly faster than with 12.0.10 and 12.0.40. So my memory wasn’t completely wrong.
It’s understandable that SL 12 takes longer than SL 11 with its new technology. But why are there such big differences between the 12.x versions?
12.0.0 = 11:28
12.0.10 = 18:58 (+ 7:30!)
12.0.40 = 15:06 (+ 3:38)
But there’s nothing we can do about it, we can’t change it and we have to live with it. In any case, I’m not going to buy a new graphics card for 700 €…
The initial 12.0.0 release did not have the full quality for Unmix Song and was not GPU accelerated. With 12.0.10 came improved models. With 12.0.20 and above came full GPU acceleration.
With my GPU Unmix Song in best quality with everything active is almost 3 times faster than realtime (without loading time for models into VRAM).
I wish I would have had to pay only 700 € for the GPU I need…
