This was so insane I had to share
This is a one minute song being unmixed with upcoming patch 1:
Other goodies coming as well with this patch.
This was so insane I had to share
This is a one minute song being unmixed with upcoming patch 1:
Other goodies coming as well with this patch.
Ohhhh , i noticed you are using an Rtx 3090 , would i see any benefit with a lesser Gtx1650 oc v1 (2020 model )?
Not as brutal as the RTX 3090 but most definitely yes. You can probably expect at least 3 times faster than the CPU.
Holy smolly , i look forward to trying this , that seems pretty incredible
Wow!
.
⦠this ⦠changes ⦠everything ā¦
Especially your graphics card
Could you show Demos using, AMD and Intel GPUs. Very impressive Demo by the way.
Unfortunately I donāt have AMD GPUs to test with (hopefully some beta tester will have eventually one).
Regarding Intel GPUs donāt expect miracles with on-board (integrated) GPUs, at least what Iāve tested on my laptop did not led to significant speed gain. A dedicated Intel GPU such as Iris Xe should be a different story though.
amazing. I have an AMD rx6700xt I would like to do the test if I am provided with the file and the update
Works great here for AMD gpuās, much more improved performance and surprised how much more quickly it processed it. Good job.
Do Apple Silicon (M1) users benefit anything from this patch?
I have Nvidia 1080ti. This looks like an excellent way to speed up separations. Thanks Robin.
Great stuff, Will there be any aditional options to adjust what in demucs processing we usually call āShiftsā. This is where multiple predictions with random shifts are made. so e.g. if you chose 10 Shifts, this would essentially be x10 processing (basically not even possible without a GPU and lots of VRAM as thatās essentially ten times longer to process than no shifts). Also we would typically play with something called āOverlapā which controls the amount of overlap between prediction windows. (Lower ovelap = faster processing subject to GPU perofrmance) and finally āSegmentā which changes the size of each split.
That could be a possibility (which was not even a question when only CPU computing was available). Let me think about itā¦
Note though that thereās already a (secret) form of multipass computing happening under the hood when selection āBalancedā and āBestā quality, which enables much better extraction of instruments such as guitar and piano.
It works very well with the amd rx6700xt. much faster than my ryzen 9 7950x3d
great job robin
I donāt seem to have the option using GPU acceleration. Somebody any idea?
Hopefully M2 Max too. My shiny new MacBook Pro is waiting for it.
That is really impressive, Robin! But the cost of the RTX 3090 is insane, more than an ordinary pc. What more cost effective solutions can you suggest? Preferably fanless ones, too, since it is important - at least to me - that a gpu is quiet.
Best,
Magnus