This is a user forum, so nobody will list and summarise things.
Common understanding is that you should get a decent video card, specially if you plan to use DeepPrime and DeepPrime XD. One of the threads opened a Google spreadsheet where real world measurements were added by users. Try to find the respective post.
What constitutes a decent video card for this purpose: memory ? cores ? speed ? cost ?
The same questions can be asked of CPU.
Some applications do not take advantage of the latest features, so it might not be prudent to merely “throw money at the problem” and buy the most expensive configuration we can afford… or is it ?
To put it another way, is there a way to prioritize the requirements ? For example, back in the day with Photoshop, the more RAM we had, the better. With insufficient RAM and scratch-disk space, whatever CPU we had was I/O bound. What are the analogous factors here ?
To ask the question a different way: what are the bottlenecks or constraints to efficiency, where spending money will be most effective ?
(Canon, PL5 on desktop GTX 1050ti, Win 10 & on laptop GTX 4050, Win 11)
@noname has asked the key question. I would add, that while you can get away with GPUs that are below the minimum specified by DxO, it’s not a good idea. In other words your minimum budget for a GPU has to be the cost of the minimum as specified by DxO. Beyond that, the higher the spec the faster PL will export images using DeepPRIME or DeepPRIME XD noise reduction.
Also bear in mind that the amount you go over the minimum spec should be driven by how important outright speed is to you.
For example, if you edit 10 photos and export them in a batch when done, then speed may be unimportant, as you can go off and do something else while they export.
If, on the other hand, you edit 500 photos in one sitting and want to export them for a client before moving onto the next job, then you’ll want to invest in a lot of power to churn through the exports.
I recently bought a new computer. Since then I have been trialling new software, such as Topaz and DxO.
Now I realise that I need a decent gpu. Not cheap, I’m looking at GeForce RTX 3050 8gb, which is around £220.
DxO is slow on PureRaw, on 5DS size files it takes around 6 minutes each image, even a 20mp image takes over 3 minutes. Topaz is even slower on Sharpen AI. Nik Efex only takes a few seconds to complete.
Careful with interpreting export times when using the same image(s)…as DPL’s cache can make a difference. Exporting one image with the sequence of auto/manual/auto selecting GPU got me times of 10/7/7 seconds respectively - with edits set to NoCorrection + DeepPRIME at its default values.
ETA is wrong in the beginning. After a while, estimated times more closely match what you’ll get.
Well I tried it on different but similar images, times were exactly the same.
So gpu was 3x faster than the the auto default setting which used cpu only.
This has been the same on other images, I did a batch of 14 images and they took around 84 mins. The times are consistent 6 mins on cpu (auto), 2 mins on gpu. Why is auto going to the cpu only setting? Results look identical to me.
Times near the end can wrong as well jumping from 2 mins to only around 10 second’s sometimes. Or saying 10 seconds when it takes over 60 seconds.
I recently really pushed the boat out and changed my 2012 i7-960 (12GB) with unsupported GPU for a Core i9 13900KF + 32GB and RTX 4070 + 12GB. What a difference on using Deep Prime AI, 30 minutes down to under 3 seconds for a Nikon D810 RAW file.
I have a NUC with embedded graphics and Core i9 9th generation mobile (power limited) and that took 4 minutes for the same image. I also tried it on a gaming laptop (i7 + 8GB) with supported graphics card (4GB), vintage 2017 and that took around 65s.
My results back up your statement about graphics card and associated memory. The CPU doesn’t count for that much. I started off looking at the RTX 30XX series but thought if I’m not in the market for another PC for 10 years or so what the heck. I did place a limit on spend though so juggled parameters within that limit.