Both standalone and as extension in Nuendo the process “unmix song” crashes when system = GPU.
The fix with DirectML was applied.
What could be the case?
Nuendo 18.104.22.1684 64bit 2023.7.25 22.214.171.1245.dmp (1.3 MB)
System = Windows 10 latest
CPU = i9-10900X 3.70GHz
RAM = 128 GB
@karel_broekhuis can you screenshot your System Informations in the NVIDIA Control Panel ?
You are faster then light…
NVIDIA System Information 07-25-2023 22-46-05.pdf (15.1 KB)
OK, see attachment.
It might be due to the limited Dedicated Video Memory of this card (2048MB).
Loading the AI model for unmix song and running it uses 2GB of dedicated GPU memory.
Even idle, a card already uses 1GB or more of that memory.
I’d recommend a GPU that has at least 4GB of memory. The cheapest GPUs these days usually have 6GB.
Thank you, Robin.
I’ll keep that in mind.
@karel_broekhuis what happens if you try a simpler process with the GTX 1050, such as Unmix > Components or Unmix > Drums ?
Stand alone and with ARA inside Nuendo: same results:
Pop song 3’32"
Unmix drums: OK in about 15 secs
Unmix components: OK in about 14 secs
Unmix song (even with a 10 secs fragment and “fast”): crash.
Comparison with Sytem = CPU: Unmix components about 20 secs
Whole track unmix song mode “best”: about 3mins30secs.
Thank you for your help,
Thanks, ok that seems to confirm the crash is indeed due to too much VRAM use with unmix song (unmix drums and unmix components only use a couple hundred MB).
It is amazing how quickly the developments in AI software have pushed older but still reasonably respectable GPUs into virtual obsolescence. The workstation laptop that I’ve used for a lot of audio, photo and video work over the past five years has an nVidia Quadro P1000 - one generation earlier than @karel_broekhuis’s GTX 1050, but it does have 4GB of VRAM. It will unmix songs OK in SpectraLayers 10.0.10, but its performance on some of the more heavyweight photo processing algorithms is atrocious (the GPU has no tensor cores) and it doesn’t have enough VRAM to accelerate Adobe Lightroom Classic fully. The video decoding and coding features in this old chip are not good enough to work with modern cameras in Premiere Pro without considerable pain. Really, my venerable old machine has reached the end of the road: the GPU is inadequate in every department and the six-core CPU isn’t really adequate either.
I’m coming to the point of replacing that laptop and am considering a workstation laptop with a very expensive nVidia Ada Lovelace chip with 8GB+ of VRAM. For future proofing, I feel I ought to be aiming for 12GB VRAM minimum, but that will result in a staggeringly expensive laptop.
I realise that GPU-intensive work is really better done using desktop hardware, but that doesn’t fit my usage patterns. It is a case of buying the most powerful and best-cooled laptop (or should that be “vaguely portable”?) that I can.
Have you any advice about the sweet spot in the current nVidia lineup for audio processing, Robin?
Don’t spend too much on the GPU if you’re not going to do some gaming or VR with it - any RTX 3xxx or 4xxx is future-proof enough for audio and video production, whatever model you choose
Just to confirm: today installed an AMD Radeon RX6600 8GB SWFT210 1792 cores.
€269 including montage. Result 3½ mins popsong unmix in about 10 secs.