Setting for RTX3060

[mozart]

Hi all. Just installed an RTX3060 (windows 11) with the hope of gaining some benefits when doing unmix song in spectralayers. 13

IN the task manager, it’s maxing out the cores on my i9 285k as well as showing some 100 spikes on the GPU.

I’m wondering if there are any settings - either on the RTX3060 driver itself or within spectralayers that I need to set other than setting the processing to the RTX3060 rather than “CPU”.

I’ve been reading on this forum about an “extreme” setting but not sure where that is…

Can anyone advise?

[Alfo]

To optimize SpectraLayers for an NVIDIA RTX 3060, go to Preferences > System (or General/Audio Device, depending on the version) and ensure that GPU acceleration is enabled. Your RTX 3060 should be set to use CUDA, as this significantly reduces unmixing and processing times.

https://forums.steinberg.net/t/unmix-song-with-extreme-settings-slow-zzzzzz/919134/35?page=3

[mozart]

Thanks for your reply Under preferences / system I don’t have an option for “GPU acceleration” I have chosen RTX 3060 in “AI processing device” which I presume is the same thing?

Unfortunately I don’t know how to set my 3060 to use cuda. I can’t see anything in the nvidia app for turning it on of off and google hasn’t turned up anything

[Jari_Junttila]

Check nvidia control panel for power management and that it´s set for prefer maximum performance. I have 3050 8Gb, it´s definetely not a fastest one, but it can manage, though with all maxed out, it has to use shared memory, something like 24 Gb, but I have 64 Gb on system to share..

[Alfo]

On an NVIDIA GeForce RTX 3060, CUDA is always enabled by default once the correct graphics driver is installed. There is no switch in Windows to manually turn CUDA on or off. Any current official NVIDIA graphics driver (Game Ready Driver or Studio Driver) from version 461.92 onwards fully supports the CUDA capabilities of the GeForce RTX 3060. There is no need to look for a separate legacy driver; the core CUDA runtime environment is built directly into the standard display driver. Have you actually installed the CUDA Toolkit?

[MrSoundman]

SpectraLayers doesn’t require the CUDA Toolkit.

[Shortstop]

@mozart To the right of setting AI processing device, tick the box for “Use shared memory”.

This will speed things up when unixing.

Download the Studio driver from Nvidia and install graphics driver, express settings. No need to install the Nvidia app.

That’s all you have to do.

@Alfo No need to complicate things.

[mozart]

Thanks for your reply. I did try the “shared memory” option but with that “on” I get crashes half way through an “unmix”

[mozart]

@Alfo

To accelerate an NVIDIA GeForce RTX 3060 in Steinberg SpectraLayers for AI processes such as unmixing (source separation), you need to install the latest NVIDIA Studio driverNVIDIA Studio driver, set up CUDA and TensorRT on your system, and enable GPU acceleration in the SpectraLayers program settings.

Ensure that the latest CUDA Toolkit is installed, as SpectraLayers uses the RTX 3060’s Tensor Cores for AI calculations.

Then open SpectraLayers and go to the Preferences, under the System or Advanced tab. Select your RTX 3060 as the preferred device for processing (enable GPU selection). Finally, restart the program so that the AI ​​models can access the graphics card’s VRAM.

Are you just getting these reponses from AI?

[Scab_Pickens]

@mozart

Are you just getting these reponses from AI?

Hello, Alfo!

@Alfo, are you a human? I’m thinking not. Every post I have read seems as if it was ripped from an AI search. Multiple categories with the same knowledge of all. Please make a comment that doesn’t sound robotic.

Sadly, it might be true.

[Shortstop]

Did you restart SpectraLayers after you checked use shared memory?

[Sophus]

@mozart

Hi all. Just installed an RTX3060 (windows 11) with the hope of gaining some benefits when doing unmix song in spectralayers. 13

I recommend you only split two stems at the beginning, for example “vocals” and “other”. Then split an additional stem from the “other” track and so and on. Experiment with selecting more stems, if it works fine.

In my experience, if you select all stems in one go, you need minimum 17 GB of VRAM, which the 3060 doesn’t have. This either leads to crashes or the process uses system RAM, which slows it down an order of magnitude.

[Shortstop]

I do unmixing regularly with a RTX3060 with 12GB VRAM.
Drums, Bass, Gtrs and Vocals in one go and then unmixing drums.
Never had a crash in SL 13 or previous versions.

@Sophus I don’t know where you get the minimum 17GB VRAM from.

[Sophus]

@Shortstop

I don’t know where you get the minimum 17GB VRAM from.

If I select all stems in the unmixing panel the VRAM gets filled up to around 17 GBs.

[mozart]

I did yes. It gets about 20 seconds in and then just closes

[mozart]

When I installed the studio driver, I chose not to install “Physx”. Was that the correct choice?

[Shortstop]

Yes. SL doesn’t use Physx or Cuda. It uses GPU to run AI models for unmixing ao.
How Much VRAM does your RTX3060 have and who much RAM do you have.

[mozart]

Thanks for confirming the 3060 has 12GB Vram and I have 96GB ram in the PC. Funnily enough, with the shared ram option ticked, it doesn’t crash every time I do an unmix song. Maybe it depends on the source material and how complicated it is?

[Sophus]

@mozart Could you please check the VRAM graph for your RTX 3060 in the Windows Task Manager during unmixing and whether it’s going over the 12 GB limit? If you have a display connected to your RTX 3060 Windows might already reserve 0,3-1,2 GB VRAM, at least in my experience. This amount is then not available for processing.

[Shortstop]

@Sophus

The point of use shared memory is to extend available memory and offload from GPU when needed.
I have unmixed 10 minutes of a recording consisting of drums, bass, guitar, keys, vocals and sax in one go without problems.

100% GPU 100% 12GB VRAM, 2% CPU and 14GB RAM used (That’s the point of using shared memory.)