I know this has been discussed on the Forum. I gather there is a sort of consensus that pushing the number of “parallel” processing threads (in the final Preferences tab) past about ½ the number of CPU ‘cores’ makes not much difference. This is my experience, too.
What I did not fully appreciate is how the choice of “Prime” or “Deep Prime” noise reduction affects the export processing time. On my 2019 iMac (27", i9 with 48GB of RAM and the standard Vega 48 GPU) when I select Prime – that I expected to be less of a burden on the processor – the export of either TIFFs or JPEGs is woeful (4 mins to export 10 images).
But I have found that choosing the more intensive “Deep Prime” processing — that is unnecessary on 90+ of my full-frame raw images since I’m careful about exposure — is a LOT quicker. Export takes about HALF the time (2.25 mins for the same 10 images). I speculate this is because Deep Prime pushes the work off to the GPU and leaves the CPU to do the rest of the image processing.
I’m guess you all knew this and maybe someone mentioned it previously. But I must have missed the ‘memo’.
In my experience, DeepPRIME always gives better results than PRIME and even images shot at lower ISO’s with smaller amounts of noise greatly benefit from it. I suspect that the only reason that PRIME is still available as an option is because processing is so much faster than DeepPRIME on machines that don’t have a supported GPU.
Not that I have ever been inclined to do it much, but I don’t think I’ve ever come across a shot that DeepPRIME handles worse than PRIME.
But, as I have said elsewhere, export speed is of little concern to me because I usually leave exporting until I’m about to walk away from the computer. Certainly if I have more than a handful of images.
I can confirm Mark’s report that DeepPRIME never gives worse results than Prime NR and I always thought Prime was amazing before DxO created DeepPRIME. As someone who is often working with images up to 16000 ISO, I have occasion to notice differences between noise reduction tools.
Don’t entirely agree that small amounts of noise require or benefit from DeepPRIME but DeepPRIME will do minimal damage to an image’s colour or perceived resolution when wielded with a light touch.
What you say about. a “light touch” DeepPrime is interesting Alec. I have been wondering what the ‘auto/wand’ setting in the DeepPrime dialog means.
DP sets the “dead pixels” level (%? who knows?) to “40” or so on my relatively new CANON R5 that I am pretty sure has no dead or even ‘hot’ pixels for an ISO 100 shot, but “24” for an ISO 4000 shot. Huh?
I don’t understand, either, what the ‘luminance’ value measures. I don’t see any linear relationship of the value to e.g. the ISO of the images or to the quality of the output.
In generally low-key but well exposed images when I set either of the DP values (luminance and ‘dead pixels’) lower or even off, I can see a tiny overall luminance difference in the images when I export to full-size TIFF files (by switching quickly between them on screen) but with no real difference in the quality of the images. Maybe you have to print at some large scale to see that.
But there’s a big difference in export times. On my M1 Mac Air (my ‘iPad’ replacement) which uses the “built-in GPU” of the M1 chip, I can export a full-size 16-bit TIFF file in 5 seconds. ANY use of DP adds 35 seconds to that, and the different settings of DP add a further 1-3 s in the time-to-export.
On low ISO shots, I wouldn’t apply noise reduction at all in most cases (exception for extreme shadow lift for instance). DeepPrime/Prime NR have more to offer you as a Canon shooter than almost any other brand. Canon RAW images suffer from a lot of very ugly chroma noise. I moved from Canon (5DIII, 5DSR) to Nikon (Z6, D85) a few years ago, mostly over Canon’s cripple hammer on video and partly for better very high ISO images (12800 ISO+, night sports). Prime is still valuable on Nikon but instead of offering a full 1.5-2 stops improvement as on Canon, Prime is more like a .75-1 stop improvement on Nikon.
Why? Nikon manages to make most noise luminance noise, which looks like grain.
The point of the story is that it’s worth continuing to plug away at finding the right settings on Canon. Forget about the default of 40 – that creates an artificially smooth image, albeit not quite as waxy as most of the competition. The magic wand will give you reduction at similarly extreme values. Chose noise reduction manually. The area to play around is from about 5 to 20. I mostly use settings around 12 and the effect is plenty strong.
Export times don’t bother me. Interface lag bothers me (a lot) as I like to make quite a few manual adjustments and see the results right away as I work. Hence I turn on noise reduction as a last step when working on an image or even as a last step across all the images in in a large set of images shot on the same camera at roughly the same ISO range.
There is mention above of offloading to a (supported) GPU. The GPU is used for the screen of a computer (including both still and video). Does DeepPRIME take away some GPU threads for DeepPRIME to “share” the GPU with other uses? Or are these systems with two video “cards”, say an Intel and a Nvidia, with the latter being the GPU? If DeepPRIME uses a Nvidia GPU, was DeepPRIME for machines that support CUDA, coded/built using Nvidia CUDA?
