When exporting with HQ denoising and Prime denoising, DxO manage to take full advantage of my CPU (a i7 with Quad-Cores ).
But when using DeepPrime, the software seams to use only half of them… Which is pretty curious / sad given the time required to treat each pictures.
So I was wondering if :
- I am missing a setting ?
- I am the only one experiencing that? If maybe it’s due to my CPU, or if you all experience the same ?
See attached a screen shot of the monitor of activity. On the left, exporting with Prime. On the right, exporting with DeepPrime.
DeepPRIME leverages your GPU, instead of CPU … Does that explain your observation, Alex ?
If you have a compatible GPU that`s a normal behavior.
The really compute Heavy stuff is run on your GPU then, with some additional processing on the CPU.
What i can tell from my observations on quite slow machines is, that during the frist half to two thirds of the processing time most work is done on the GPU with some additional low to medium load on the CPU.
While in the last third it consumes most of the CPU resources for a short time.
However the faster the whole process runs the less you will notice it on individual files.
You can also force a “CPU only” mode btw by selecting it in the menu and restarting PL4.
But that will run considerably longer and hotter and even saturate at least an 8C/16T CPU.
I should add that even when DeepPRIME is running on CPU (either because you forced it or because you don’t have a compatible GPU), not all logical cores will be used. All physical cores are used though. This optimization is managed by Apple’s technology « Core ML » and the reason looks to be because Core ML makes use of advanced instructions (vectorization) that is basically already using all the compute capacity of your CPU. Note that we tested processing two DeepPRIME at once on CPU to actually use all the logical cores and the total time to process the two images was almost the same.
Thanks Lucas for those explanation…
Do you think in a new release the soft would be able to use 100% of both the CPU and the GPU?
Or is that not a option regarding what you explained with the “Core ML” tech ?
It could be, but a CPU is so slow to process DeepPRIME compared to a GPU of « similar rank » that the benefit would be low. Additionally, while GPU is busy for the sub part of the export that uses AI, other exports that are at a different step in the export process can continue using the CPU.
It would be more interesting to be able to use several GPUs if the user happens to have more than one.
I invite you to create a feature request if that’s something you’re interested in .