I am on PL5 and compared performance like for like with PL6.
I exported 36 images (30MP each) all with DeepPrime enabled in PL5. Duration 2:33 Minutes.
Then I repeated the same in PL6 which took only 1:31 Minutes.
Then I repeated the same in PL6 but with DeepPrime XD which took 3:02 Minutes.
Is this consistent with what others observe? Is PL6 faster for DeepPrime?
Do you have the same number of files to be simultaneously processed set in both PL5 and PL6?
Yes, the settings are the same. I checked before.
For me it sounds logic, because PL 6 DPXD has to do more calculating for the higher output quality.
It’s a little bit like in Excel Solver Calculating, the more Constraints you set or like to have the longer the time till you get the result
I would say it depends on the HW and primarily the GPU.
It’s hard to judge without information on that.
On my systems, DeepPrime performance is basically unchanged with PL6 and xD happens to be (since the DNN is 3x larger) about 3 times slower in most cases or needs about 3 times the (processing) power in the cases where DeepPrime has some HW utilization issues.
Correct, but I don’t think that lighthunter changes his system from test to test
Yes, but it’s pretty hard to judge whether this is normal or not without someone being able to compare to similar HW.
The architecture used seems to heavily influence the result.
This was already significant when comparing PL4 vs. PL5, when they switched to fp16.
AMD Zen3 gets about 3x slower when using DPXD on the CPU compared to DP.
The Intel Haswell I tested got about 5x slower and on top of that I’ve already seen some Intel TigerLake that were up to 9x(!) slower with DPXD than with DP.
On the GPU side, this is probably even more interesting since ML accelerators are now a thing.
For me, when exporting the same set of files using the same machine (Mac Mini M1 with 8Gm memory), export times using PL5 & 6 using DeepPrime have been essentially identical.
Export time of the same set of files when using DeepPrime XD (in PL6 obviously) has taken twice as long.
I don´t know if PL6 is faster than PL5 for DP but processing speed and the differences between DP and DPXD is similar for your and my system.
I clocked a couple of example files (24MP)
DPXD 7 sek
DP 3 sek
DPXD 8 sek
DP 4 sek
DPXD in my case is around two times slower.
I have an i7 with 16 GB, SSD and Geforce RTX 3060 TI plus OpenCL active.
I wonder if somebody with a fast computer have tried if performance will increase if you increase the number of parallel processes from default 2 to something else.
Also take in account that there might be some problems with artefacts using XD. XD might not be the obvious choice of that reason. Please look at this link and the screen captures in it:
DxO PL5 vs DxO PL6 NR algorithm - DxO PhotoLab / DxO PhotoLab Mac - DxO Forums
Maybe DXO has released XD a little bit too early, because it might need some tuning before it´s fool proof.
Interesting feedback, thanks.
I run a Ryzen 5900x, Radeon 6700xt (16Gb vram), 32GB ram, and a 4th gen SSD. During export, my CPU is not fully loaded and most happens in my GPU.