How to chose "best" graphics card for DxO hardware acceleration (in particular DP XD speed)

DxO PhotoLab Windows topics
21

I upgraded to an RTX 3050 and have not regretted it. My choice was dictated by the wattage of the power supply in my PC; there were faster cards, but I might have had to upgrade the power supply, and did not want to do that, because it would have meant effectively rebuilding the machine, given the number of leads I would have to remove and then reconnect. It is much faster than the old GTX1050 card. I have yet to find out what PL6 is like on an M2 Mac, though I am hoping it will be as good, if not better.

People are right about a fast SSD being very helpful. I need to store most of my photos on a large HDD, but I tried to work out a system to copy files that are currently being worked on to an SSD, and then sync them back to the HDD. I say tried, because I still haven’t got it tweaked as I would like it to be as there is too much manual syncing involved in the setup. Unlike Lightroom, where you just need the catalogue on a fast SSD, PL6 does like photos themselves to be on one.

22

I have yet to fit my RTX3060 but in regard to PSU wattage mine is 650W and AFAIK and am aware should have plenty of ‘headroom’. (Motherboard with i5 1st gen @95W, two SSDs, three HDDs, four fans).

I made sure when I built it to allow for excess capacity if for no other reason to avoid stressing the PSU & to allow for expansion and upgrade changes over time :slight_smile:

FWIW this PSU wattage calculator tells me I need 514W
Power Supply Calculator | Cooler Master

23

@BoxBrownie My system is an i7 4970K with 24GB, 4 HDDs, 2 x SATA SSDs and one NVME connected via a graphics card slot on a PCie card.

The power supply is a Seasonic Platinum and the most power consumption that a power meter has shown (so far) with my new RTX3060, when processing DP XD, is about 330W or thereabouts and the machine idles around 100W with most of the normal graphics routed via the onboard Intel chipset.

With respect to choosing the best graphics card I would suggest that it is a mixture of graphics card AND CPU. The problem is that its is easier to change graphics card than CPU, if I wanted to change CPU then I would need to

  1. Buy a new processor
  2. A board to run it
  3. New memory (DDR3 won’t run on the board I will need DDR4)

That all “just” costs money, then we come to all the software installed!

I should, hopefully be able to clone the Boot SSD (although I initially had trouble moving to the i7-4790K from a previous generation of processor and motherboard) and install the clone into the new machine but every piece of software that uses a machine footprint will need to be re-licensed!

The issue comes down to this, which I have been “banging on” about for weeks, replacing the GTX 1050 with the RTX 3060 has shifted the “problem” with exporting away from the GPU to the CPU.

It hasn’t completely eradicated GPU time for DP exports but the Noise Reduction component (mostly GPU with some CPU to control the process) now accounts for 96 seconds of the time of an export of 109 RAW images I took yesterday which took 595 seconds in total. This took 1181 seconds on an almost identical i7-4790K with a GTX 1050Ti GPU and the NR component was 738 seconds, i.e. 738 seconds for NR down to 96 seconds.

DP XD is a bigger problem with the 1050Ti card the whole export took 3358 seconds and the NR component was a “mere” 2915 seconds. The export time is now 744 seconds and the NR component is 245, and while it is possible that an even more expensive and faster graphics card could reduce that further, the biggest component now is the CPU only element of 499 seconds.

The techpower web site ranks the 1050Ti as 33% compared with the 3060 at 100% but looking at the performance ratios the older card is performing below that ratio I think and PhotoLabs clearly only uses part of the card and only for short periods of time.

The problem is that no-one from DxO comes near the forum @Musashi and we need real guidance to point to what features DP and DP XD actually exploit on the GTX and the RTX and the Radeon cards to better guide our use and purchases, but I won’t hold my breath waiting.

Putting testing etc. to one side we went for a walk along the paths around one of the two adjoining Golf courses that use the Southern slopes of the downs for their fairways yesterday (Sunday)

image
image1920×1052 234 KB

I then used the 109 pictures I took to run a variety of tests, and these are summarised in the following tables

image
image1390×202 12 KB

In the tables I classify the NR element as GPU, but it has to contain an element of CPU to control the GPU elements and consolidate the results into the image etc.

The %age column is the comparison of the GPU (NR) time per image for the 3060 divided by the same figure for the 1050Ti. There may be better ways of expressing this but clearly the improvement for DP is better than the improvement for DP XD on my systems, with my RAW images and my edits, other users may do better or worse than this!

The Google spreadsheet is useful but it is “simply” comparing the overall time to complete the whole job and is actually rolling the CPU usage (governed by the CPU power and the ability of DxPL to use that power productively) together with the GPU usage (governed by the GPU power and the ability of DxPL to use that power productively).

A faster CPU will distort the overall score and may make a slightly slower GPU look better than it actually is.

1 Like
24

I have yet to find out what PL6 is like on an M2 Mac, though I am hoping it will be as good, if not better.

Unfortunately it is quite bad with the base M2 models with small number of GPUs, as the neural engine acceleration is having issue with putting a purple/red cast on photos processed with DPXD.
While using GPU is fine, it is much slower comparing to what the neural engine was offering.
I should also say this happens only on the lastest Mac OS version (13.x Venturna), but Mac OS 12 is fine afaik.

25

I recently went from a Zotac GTX 1660 Ti to MSI RTX 3070 and have noticed a (HUGE) difference in processing times from my Panasonic G9 raw files. My system is Ryzen 9 5950X, 32G ram, Win 11 and 1T SSD, using DXO Photolab 6.2 and DeepPrime XD either exporting to disk/another application or Nik Effects etc nothing takes longer than 7 or 8 secs. If you can afford a decent graphics card then go for it.

2 Likes
27

Thanks Andy,
Your configuration is almost identical to mine, then adding a state-of-art-graphics card such as the RTX 3070 that you mention will have the same positive effect!
Thanks for mentioning!

1 Like
28

Hi leecd, I recently replaced my aging RX480, which worked with DeepPRIME XD, with an ARC A750 running the latest Beta drivers and sadly at the moment this does not work with DeepPRIME XD. The DxO preview window when XD is selected displays either a highly pixelated area or just black and when exported produces an image that is 10% colours and 90% black. I currently have trouble tickets out with both Intel and DxO. Suggestions welcome!

29

get a supported card ( consider it as the price paid for not checking matters before buying intel dgpu ) or wait until it is fixed ( put RX480 back in use for now ) …

30

How to chose “best” graphics card for DxO hardware acceleration (in particular DP XD speed)

If you REALLY want the BEST one, take the last serie and more powerfull one.
Which is in NVidia world RTX4090. (I’m not sure they will produce a RTX4090Ti - Ti are about 15/30 % faster than no Ti ones).

But think about form factor (size of 3 pci slots plus some space for cooling) and about power supply needed.

If not the BEST one, choose AT LEAST a 20xx serie. No less.

31

I found you have to look at the power needed. The 3 series needed a lot more than the 4 and to be safe would probably need a change on power supply. You also have the problem of changing power connections. Many of the newer cards need diffrent one to the ‘old’ 6/8 one again probably needing more changes. Many of the new cards are also much longer and even wider than older cards so check space. So its not just which card is better but a range of things you need to look at.
The lessens I learnt, as it was many years since building my own PC’s, were there a clips that hold the card down. Many CPU coolers make access interesting, my original card being an old one was no problem but the new one, will need a plastic ruler to press it to get the card out. A magnetic screw driver is invaluable for not losing the retaining screws as the new cards are much bigger in height as well. Cables will also often be tied up by plastic ties to shorten them (to minimize restriction to airflow), new cards being bigger need a longer cable length to get to there socket without an amazingly tight bend, not a good thing. The card I ended with minimized the need for higher power and just fitted in my case with out to much problem (ASUS Dual GeForce RTX™ 4070 OC Edition 12GB GDDR6X | Graphics Card | ASUS UK).

32

In my humble opinion, the 4090 is complete overkill. The CPU is just as important as most processing is actually done by the CPU

For instance, I upgraded my PC. I came from a Intel i5 6400 and a AMD RX480 and started by replacing the motherboard/CPU/memory. I switched to AMD Ryzen 5 7600X. Still using the RX480 though. This already more than halved the processing time: from 55s on avg to 25s on avg using DeepPrime XD

Next I replace the RX480 with the RTX 4070: average processing time with DeepPrime XD now is 4 seconds.

A RTX 4090 obviously is a lot more powerful. But will it be worth investing +€300-400 over the 4070 just to save maybe 1 or 2 seconds?

33

To use it solely for denoising I think so too.
But the OP asked what was the BEST one in the title. And this is this one.

However a test that could be done (not by me) is using this card with a very recent and fast motherboard and cpu, very fast SSD and ram, and see how many images can be processed at the same time (du to high amount of memory on this card, this when reducing all other bottlenecks).

34

Luckily someone already tested that. There’s a spreadsheet available with defined set of images to benchmark PhotoLab.

There is a RTX 4090 entry with a very potent Intel i9 13900. That user processes the batch of 5 D850 images in 13 seconds when processing with DP XD. The same user also benchmarked with the AMD 7900XTX (15s) and RTX 4080 (17s). This should give an idea.

My own machine processes these images in 25s.

So, in a nutshell:

  • 4090: 13s (2,6s per image)
  • 7900XTX: 15s (3s per image)
  • 4080: 17s(3,4s per image)
  • 4070 (with less potent CPU: 6 cores vs. 24 cores): 25s (5s per image)

Now, the last entry is not a very good comparison. Single core performance of the i9 13900k is already quite a bit higher but due to the sheer number of cores of the i9 13800k the multicore performance is 2,5 times as high.

It’s up to the user to determine if the 4090 is worth the extra money :slight_smile:

35

How many maximum simultaneous images where set in this test ? (in preferences/performances/display_and_process tab).
This was my question : Is a setup like this one able to run more simultaneous images and therefore use the advantage of large memory capacity of 24Gb cards (if all possible bottlenecks are reduced to minimum : drive speed, cpu speed, ram speed, pciexpress speed, etc…).
I think a test of 5 image isn’t enough if a combo like this take advantage of processing more simultaneous images…

36

More threads/simultaneous images just because you have more cores isn’t always faster. It’s not a linear thing.

Others on the forum already tried that. Maybe not with a beast such as the i9 13900k but with a 10 core processor. This proofs that just because you have more cores you can’t just raise the number of simultaneous images

37

I know all this.
My question is are there tests done with preference setting changed ?
The one which deals with this value.

38

If you read the link you’ll see that at least someone did it.

The purpose of the sheet was not to test the performance of PL to that extend. The purpose of the benchmark was to get a general idea of performance of a CPU/GPU combination with a somewhat predefined setting: at least the same images with the same processing settings.

If everyone used the same number of threads is unknown to me, but apparently, according to the thread I linked to, that is rather irrelevant too as more threads =/= faster processing times.

39

I don’t see exactly what are specs of the station used in the link.
It seems it is a mac ? And an other people talk about a 8 core PC ?

Anyway, for few seconds computations, pciexpress (or something related to speed to feed graphic card) is probably the botleneck. Even with a top workstation.
And 4090 probably can’t run 100% for denoising.

40

George,

Thanks for the update and am sorry to hear about the ARC A750 GPU card issues as I’m sure they are not fun.

I ended up replacing my graphics card with an AMD RX 6650 XT. Processing time on the Olympus 20 megapixel files has gone from 24 to 6 seconds. Not too bad for the price.

I could probably get faster speeds by also upgrading the motherboard, going from PCIe 3 to PCIe 4 and also from SATA SSDs to NVMe SSDs as well. However, for what I do, I’m happy with the current processing speeds.

Lee

41

After much reading (in this forum, dedicated PC component sites, computer magazines) and fierce budget fights I decided for a Geforce RTX 3060TI card. Not fully state of art however good price/cost ratio and obviously good experience of DxO PL users in this forum, too. There was enough reserve in my 500W power supply and it has 2x8 pin connectors for graphic cards, hence no problem with that.
Measurements for PL DP XD (using Studio drivers 536.67 as of 07/18/2023): 10 Canon CR3 35MB photos on SSD: 54 seconds
~5% of the “CPU only” time 1.100 seconds - highly impressing … almost too good to be true …
I was a bit sceptic about the GC fan noise but during normal operation it’s not audible, during processing DP XD development fan goes up but not too noisy, totally ok.
So more than happy with that …

1 Like