I’ve been holding up upgrading from PL8 to PL9 (or even installing the trial) due to this GPU fiasco, but reading the comments here, I am wondering - is the issue limited to GPUs with less than 16GB VRAM?? Do people with 16GB RTX cards use PL9 smoothly (and with acceptable export times) even with those preset AI masks? I happen to have a beefy 4070 Ti Super with 16GB VRAM alongside 32GB of system RAM. Am I clear to go without rolling back Nvidia drivers?
thx
Seems with 16GB GPU you’re fine. In the case on NVidia GPU cards the driver version can be important.
Please check @BHAYT comments previously in this thread (and also various other threads) about the matter of GPU memory and also with nVidia drivers - he does a incredible work on that!
I’m sorry what fiasco are talking about? Fiasco is a very strong word so you’re obviously not talking about the bug with Nvidia drivers that Nvidia and DxO are working together on to resolve and for which an easy work-around exists that many of us have used successfully to allow us to use AI masking.
Virtually every implementation of AI tools from every post-processing software program had some initial issues with this technology when first released. I have been using AI masking almost daily with no issues since installing the older driver as recommended until this problem is rectified. It’s not a fiasco for me or many others
Mark
Sorry for the lack of responses yesterday but my wife and I decided to make the most of the sunny weather and go to High Beeches gardens for some exercise and Autumn Colour.
The journey is not that long except the road was closed as we approached the village near to the gardens and had to find a detour which had its own problems but we got there in the end!
@BoxBrownie Thank you for the update. I am typically logged into my account on all three of my Win10 machines so as I understand
“Yes, Windows 10 Extended Security Updates (ESU) are available now for eligible personal Windows 10 devices running version 22H2 through an enrollment tool in Settings, and for commercial organizations via a paid subscription. Consumers can access the ESU program at no additional cost by syncing their PC settings with Microsoft OneDrive or by redeeming 1,000 Microsoft Rewards points; a paid option is also available. Commercial organizations can purchase licenses through Microsoft Volume Licensing or a Cloud Solution Provider (CSP).”
So I either pay £30/system or take out a OneDrive option to back up regularly, with 5GB free, to secure 1 year of ESU updates. The suggestion I have seen is to back up only the minimum amount of system config data to stay within the 5GB limit, gets one year grace period before security updates cease to be delivered!
As for AI masks they cause problems with displaying the images (editing) as well as export and I suspect they are using different AI models for the two tasks.
I say this because “Sleep” causes VRAM to be cleared and immediately exporting (i.e. select images to be exported before putting the machine to sleep) when an AI export, including the “Sky” preset, can be achieved in a relatively “small” (smaller!?) amount of VRAM but as soon as an edit operation is undertaken the VRAM usage jumps, possibly doubles.
However doing another export seems to cause another jump in VRAM usage, is the export AI model now being loaded a second time?
As for DxO resolving the issue(s), the sooner the better but I am not convinced that the fix will have any great impact on the timings involved but rather (hopefully) they will reduce memory usage to a more realistic level and resolve the “random” failures both during both editing and/or export that some users are experiencing.
But I don’t believe that PL9 with any AI masks in use will ever be as quick and frugal as before (if it was ever considered particularly quick or frugal by users!?).
Export for some users is currently still possible on 6GB GPUs but whether that will be extended to a wider group of users I cannot say. If it can’t then there are going to be a lot of disgruntled users.
Part of the issue is that laptop users, in particular, potentially will have no alternative but to purchase a new laptop if they want to use PL9 with AI masking. Desktop users have the “dubious delight” of buying bigger (VRAM) (and faster?) GPUs.
I am not sure how much saving will actually be made by waiting and/or whether any deals may be forthcoming with packages including PL9 with FP8!?
@KenA What size (VRAM) 3060 and what drivers are you using, plus the issue seems (seemed) to be more prevalent when using the PL9 provided, named presets as you seem to have discovered
As for when is the right time to purchase I really can’t say, I am just a user and have no advance knowledge of when DxO will fix the issues, although I have to believe that they must fix them or the damage to their reputation will be considerable or rather even more than it is right now.
It’s a big old world but… is that the one up towards Crawley? If so, we’re not so far apart. I’m just down on the coast!
@noam Your 4070Ti Super(16GB) is more powerful than the 5060Ti(16GB) I bought recently but one thing that appears to be clear with PL9 is that nothing is clear with PL9.
Users of old GPUs with old drivers and only 6GB are using PL9 with AI successfully, but not necessarily using the named AI presets, e.g. “Sky”.
I cannot get my “problem” image working on my old 2060(6GB) with any drivers, but can get it working on my 3060(12GB) with older drivers and on my 5060Ti(16GB) with the latest drivers. The recommended older driver might install (and work) on your 4070TiS(16GB) or not I cannot say, i.e. I could not get it to install with the 5060Ti(16GB) and had to use an earlier driver but not the one recommended by DxO.
You could wait until the “dust settles” or start now, i.e. take the trial and see what happens. PL9 appears to work if you steer clear of AI in general and AI presets in particular.
I you cannot roll back to earlier releases, because it would interfere with other software, then if PL9 takes exception to what you attempt then there are still other features in PL9 which make it a worthy release but ensure that there are no AI images in any folder you access or export if you have experienced problems.
So backups of DOPs from previous releases is always essential just in case you decide to delay any move to PL9 after using the Trial and, potentially, copy (isolate) images you want to experiment on with AI rather than leave them in the original folder.
If you have an earlier release then do not uninstall it until you are finally settled on PL9, arguably I don’t remove it at all so I have PL4, PL5, PL6, PL7 and PL9 on at least one of my machines!
Please remember DOPs and presets from an earlier release can typically be used on later releases but not vice versa, so beware “damaging” earlier DOPs by overwriting them with PL9 DOPs because then all edits will be lost if you try to go back to the earlier release.
If in doubt backup, if in no doubt backup, basically backup and potentially keep a history of edits where it is critical to be able to return to them at some later date.
Yes it is, actually at Handcross and I live on the coast South of there but not Brighton actually West of there!?
@andras.csore Thank you for the reference.
@mwsilvers Mark did DxO actually have to release the product when they did? Did they cut short the proper Beta Test and if so for what reason?
It is a fiasco and users who suggest that they are guinea pigs or, more politely, Beta Testers, would appear to be correct. “Excusing” the issue because it is between two companies and NVidia is actually big enough that it doesn’t really matter to them but it certainly matters to DxO, is not acceptable.
One of the real puzzles is exactly what circumstances trigger the problems? Users with old kit and old drivers seem to be immune to the problems (to an extent) while other users with fast, expensive kit are suffering failures. Although my 2060(6GB) certainly cannot handle my “problem” image at all!
@noam Sorry I missed the following off the response above. I originally wrote the responses in a single post and decided it was too long so I split them and wrote a new response to your query,
This is the original response I wrote
So far my 5900X(32GB) with a 5060Ti(16GB) has run successfully on the very latest and some earlier drivers with a batch of images (VCs) with the Preset “Sky” applied, i.e. my “problem” image plus multiple VCs of the same image.
One interesting issue is that while I expect some variance in export times, based on past experience, some tests have run faster on an earlier release than on later releases by 10 seconds per run of 20 images i.e. 1 min 10 seconds versus 1min 20 seconds versus 1 min 35 seconds for DP3 exports, taking about 15000MB of VRAM.
However, I may just have been “lucky” and with the wrong image and wrong edits my exports might start failing.
The 5600G(32GB) with 3060(12GB) has fared better than some other users but I have managed to crash it on more than one occasion.
By way of a comparison with the same images that took 1min 10 seconds on the 5060Ti(16GB) and took over 15,000MB of VRAM it took 2 min 20seconds and took over 12,000MB on the 3060(12MB).
The processors are about 2x for the 5900X versus 5600G and the the 5060Ti versus 3060 is 71.4% versus 42%, i.e. the 5060Ti is about 70% faster than the 3060 and the run timings are 2x faster for the 5900X versus 5600G.
The GPU performance ratio between your 4070TS(16GB) and my 5060Ti(16GB) is 105% to 71.4%, i.e. the 4070TS is about 47% faster than my 5060Ti but sadly that isn’t a guarantee that the 4070Ts will actually work successfully with PL9!!??
I wish it was that “easy” to predict and that is the real issue with PL9, the unpredictability of whether it will work or not.
I have 8GB VRAM and current driver is 581.15. I have no problems with AI masking in other software or video edits using DaVinci Resolve. I noticed that prior to my current driver the NVidia app lists 576.40 as the previous driver.
@BHAYT Thank you so much for the detailed response!
I guess I’ll just wait a little more, or until my curiosity takes over. I’m optimistic that they are working around the clock to fix it and will release an update in the near future. Cheers!
@KenA I reread you original posts and it seems as if it worked acceptably well except with the DxO AI presets. However, it then appeared to start to go wrong. Non keyword masking may be less VRAM consumptive but you have “only” 8GB, quite how much you need to “guarantee” success I do know.
Last night I did some editing with the photos I took yesterday and exported with ‘No NR’, ‘Standard’, ‘DP3’ and ‘XD2s’ and the maximum VRAM used was 2994MB.
No AI was used on any image and VRAM usage like that is what users have come to expect and then suddenly we have this “voracious” monster that seems to gobble up as much VRAM as it can.
However, I was running the 5600G 3060(12GB) on 511.09 and decided to try a “Sky” preset on one image and it failed, i.e. I got
Odd, given that the sky is easy to spot! The memory at that point was 6889MB!
So I installed 572.83 and things improved, or so I thought! The image I had been trying before worked and exported successfully. Another image was selected and the ‘Sky’ preset was fine but on a third image it fell over with 11933MB of VRAM in use.
It is as flaky as a flaky thing on a flaky day that is flaky, sorry for the technical description, i.e. somewhat unreliable!!
Switching to the 5060Ti(16GB) (V576.02) I was able to set more “sky” masks and memory remained at 8873MB and it didn’t fall over.
Exporting all the images with XD2s took 5min 51 seconds and memory increased to 13466MB certainly an improvement from the 15min 40seconds for the 5600G with 3060(12TGB) with no AI images but that only used 2886MB of VRAM. System memory was also half on the 5600G than on the 5900X, 11384MB versus 23311MB!?
@KenA My head hurts!?
@noam you are welcome.
Regards
Bryan
@KenA You also requested details of whether this worked at PL 9 not really ready for release! - #365 by KenA so I installed that driver on my 5600G with 3060(12GB) and it installed successfully and started applying the “Sky” preset to some of the images from that visit.
I marked the images I had chosen with a “Green” label and exported 39 images successfully but slowly. The images are located on another machine and are being accessed via a gigabit LAN but that has been the case for all tests with those images thus far.
9min 35 seconds to export 39 images and the the VRAM usage was 12205MB and main memory usage was 25058MB.
While attempting to set the “Sky” preset I encountered two strange incidents
With one portrait image it selected part of a tree as the sky!? The red is the mask not Autumn colour.
It failed to find the “Sky” at all in this image. I thought that signalled a failure but it just cannot find the sky when a similar image adjacent to it worked fine, or as fine as “Sky” selections manage, i.e. they really don’t like branches!
It completed an export of my “problem” image.
I will do some more tests later but that set of drivers seems to be O.K.
Yes, seems Sky AI tends to be ‘upwards’ , i guess the AI prompt written in this way…
Example: Original:
@andras.csore I understand what you are showing but I also tested a number of scenarios and it works more often than it doesn’t!
All the following were processed on my 5600G with 3060(12GB) i.e.
By the way I am actually disappointed that given where some other products are now with “Sky” masking that DxO introduce a “Sky” model that is way off target!
The Tree example with the misplaced “Sky” mask:-
We have the image with the “Sky” mask, the image with “Sky” mask turned off and its lansdcape image version that worked fine, well as fine as the PL9 “Sky” mask can manage!
The mistaken “Sky”:-
There isn’t really any real sky in this image but PL9.1 insists that that it has found one.
Where is the “Sky” problem:-
Hovering finds bits of sky easily but attempting to select using the “Sky” preset fails on this image but the next image in the sequence is fine, well as fine as …
The top of the “Sky” problem:-
This is similar to your image but with holes that PL9.1 “sort” of finds “successfully”
Additional:-
Would it have been good or bad if PL9.1 thought the reflection was “Sky”
and with this one it has decided that the tree leaves make a better “Sky” than the bits of the "Sky itself.
Many thanks for the tests you carried out with the 576.40 driver and the 3060 GPU. You seemed to have had success without the AI not locking up. I did look up the difference between the 3060 and 3060 TI. It seems the Ti version is faster and better for gaming but only 8GB compared to the 12GB of the non TI. When using the non keyword masks I did not come across any VRAM problems in editing or saving. So it looks like this may solve my problem. I just wish now I’d discovered that my previous drive was the 576.40 before my trial ran out. Many thanks for all your informatiom and testing, it’s been much appreciated.
I loooked up the requirement spec for installing PL9 it states for minimum RTX with 6GB but recommends 8GB with the latest drivers. But clearly having the latest drivers has problems for the keyword AI masking.
Regards, Ken
