I am getting export failures with the error message “Internal error (Connection failed on the Execute stage)” when trying to export some images that have AI masked areas, cropping, and DeepPrime processing with the PL9 trial. Once the errors start neither changing to Standard processing nor removing cropping allows them to export properly. Shutting down PL9 and restarting the program (not the computer) causes them to export correctly.
Searching here turns up lots of error issues, but they seem to have been generally controlled with the current Nvidia drivers and most recent PL9 trial release (mine is v9.2.1 build 542). I am not knowledgeable enough about PL9 to judge whether mine is a subset of an earlier error, or new (sorry).
Any thoughts as to the cause or a solution? Could I be doing something wrong? While my system is older, it is, I think, closer in capability to the preferred specs than the minimum specs.
System is a Ryzen 7 (8 core) with 32 Gb of ram and an Nvidia 2060 GPU containing 6 Gb of graphics memory and running the current Studio driver (581.57) under Windows 11 (23H2). The computer still has PL7 installed on the same drive as the PL9 trial.
Thank you, Wlodek. In the Event Viewer (Local) under Windows Logs/System around the time I was doing this work there are two sets of 3 entries as you describe. Each one reading:
Error (next to a red circle with an ! inside) in column 1 (column heading of “Level”), the date and time in column 2 (column heading of “Date and Time”), nvlddmkm in column 3 (column heading of “Source”), 153 in column 4 (heading of “Event ID”), and None in column 5 (heading of “Task Category”).
If I understand the link you provided, DXO is aware that this bug still exists with XX60 series Nvidia graphics cards even with v581.57 Nvidia drivers, and that DXO is attributing the problem to a bug in Nvidia’s drivers. Do I understand that correctly?
Also, that DXO is recommending that users with xx60 series graphics cards switch to CPU processing until the fix is available, or refrain from using AI mask keywords.
My problem with those recommendations is that I haven’t yet used AI mask keywords. Only generated AI masks by hovering and letting the software choose the area to mask. Is that a significant difference from AI mask keywords…as I still get the error?
I don’t have any insights into DxO, so I can’t answer your questions.
I got similar problem, i.e. together with nvlddmkm event 153 logged, not only with AI preset masks but also with hovering AI masks and exports (but they may be different problems under the hood). With PL9.2.x/581.57 they happen much more rarely than with the previous driver, but they still do happen. I’m on 4070, so guess they may happen more often on xx60 (?). There are also other outstanding serious issues, not (visibly) related to nvlddmkm event 153, e.g. related to VRAM or shared GPU memory allocation. Hopefully by April they’ll be worked out, while waiting for spring…
Andras, IIRC, it has happened with both denoise options. What I am less certain of is whether failures running DP3 were initial failures or later in the sequence after an initial failure running DP XD2.
Once PL ‘fail’ in this (or ‘Internal errors’ come up), you need to quit from PL (its ‘not recover’ itself). So, to switch instantly to Standard or DP3 → not helps, need to quit from PL.
I suggest to test in ‘baby steps’ to find out what can be ‘maxed out’ to not get issues with 6GB GPU VRAM.
Steps like:
Avoid to run other apps - they may use some GPU VRAM also (like web browser with Youtube playback may use like 0.4GB of GPU VRAM)
Try it with ‘Standard’ NR (what not use GPU VRAM) and see how its working. I know use DP NR is the target, but to see otherwise its play nice. No any DP export.
I’m pretty sure its play nice with Standard NR.
And now set-up to DP3
Use only ‘Manual’ AI masks (so NO ‘pre-defined’ AI mask like: Sky → seems for ‘pre-defined’ AI mask is need far more than 6GB GPU VRAM).
Avoid to use Loupe tool. Not turn on ‘Deeprime rendering’
Export with DP3.
If its works → step-up to DP XD2s and see.
May also try with ‘paralel export’ : 1 (and not 2)
Note: before changes please restart PL to be sure.
At general as i see, most of the problems (not all of course) in PL9 related the GPU VRAM amount. DP XD2s use more GPU VRAM the DP. ‘Pre-defined’ AI making (like: Sky) use more GPU VRAM than ‘Manual’ AI masks. Loupe tool also us some GPU VRAM amount (in the case of DP3 like 1.5GB GPU VRAM). Deepprime rendering: ON (clicked) also us GPU VRAM. And so on. And at general* PL9 if some GPU VRAM once used its not release → Once AI model loaded to GPU VRAM its stay → its just loaded once, but stay. Similar in the case of Loupe, once you use it, its ‘reserve his place’ in the GPU VRAM.
So, to ensure with 6GB GPU play nice, may some ‘puzzle/lego solving’ may not a bad idea.
May you fond out some ‘constellations’ (test):
Manual AI mask + NR: DP3 + Export → Play nice
Manual AI mask + NR: DP3 + Loupe + Export → May issues come up in Export.
Manual AI mask + NR: DP XD2s + Export → May issues come up
And so on.
Its can be great to see you finding → its may helps for PL user colleagues what works out in the case of 6GB GPU VRAM.
I cant say about nVidia drivers → colleagues in the forum can say about that / check other threads.
I think there are some complicated interactions happening. All requiring AI masking to happen.
I started with exporting two images without masking. One was 100% crop on my 4k monitor and the other the full 20 MB. Each one exported fine in Standard, DP3, and DP XD2. I switched to an image with AI masking and it exported fine with Standard and DP3, and the Loupe worked ok with DP XD2. But when I tried to export with DP XD2 the export failed with the error in my original note. I had clicked off the Raw image that was being exported, and when I clicked back onto it after the export failure, PL9 locked up the whole computer. After a couple of minutes PL9 crashed and the computer appeared to be working normally otherwise. But when I restarted PL9 and tried to export, none worked (Standard or DP3).
I rebooted the computer and tried again with an AI masked image and things went well with Standard and DP3 (successful exports in 4 sec and 13 sec respectively. Then I tried an export with DP XD2 and it took 8 min 45 sec. But it worked. Then I tried another export with DP3 and it worked, but instead of 13 seconds, it took 1 min 26 sec. I then let the computer sit undisturbed for an hour (with PL9 still open) and did another DP3 export. It was successful in 1 min 28 sec (same as just before the hour). A subsequent export with DP XD2 also succeeded and took 8 min 48 sec (same as it did before the computer’s 1 hr “rest” interval).
Bottom line is that this sequence did not trigger the export error. Exporting with DP XD2 did cause subsequent DP3 exports to slow down by a factor of 6x compared to the speed of DP3 exporting prior to exporting with DP XD2. But all eventually ran successfully.
I suspect this means that other, uncontrolled, factors influence whether or not there is an export error, or even a program crash. Such as: which other programs are open, which other programs are being used, etc.
I did some testing with my 5 year old HP laptop. A Ryzen 7, 4 core CPU with integrated AMD Radeon ™ Vega 10 graphics using 32 GB shared memory. I got similar behaviors but the underlying data were different. Drivers and Windows 11 are all up to date. PL9 is up to date (v9.2.1 build 542). Under Preferences/Performance the AI graphics driver is set to AMD Radeon Vega 10 Graphics (this has worked the best without errors with PL7 for 2 years).
I edited an image with 2 AI masks created via hovering (i.e., no AI mask keywords). It exported successfully in 38 seconds using DP3. I then tried to export it again using DP XD2 and the export failed with the same DXO error message as my original post (Connection failed on the Execute stage). Next I tried an export using DP3, again, and the export failed with the same error message. I restarted PL9 and tried an export using DP XD2 and it succeeded in about 2 minutes. I then tried another export using DP3, and it failed with the same error message as the other failures.
When I opened the Windows Event Viewer there were no events listed for the time when these things were happening. I wouldn’t have expected the same event listing as with my Nvidia GPU, as that error had a name specific to Nvidia. But in the case of my laptop there were no AMD issues when the exports were failing with the same DXO error message.
This makes me think that the bug is not specific only to the Nvidia driver or any graphics driver. The common elements are only DXO PL9, its AI masking, and DeepPrime XD HD2. It seems to me like this is a DXO problem and that the likelihood of it being solved with the next generation of Nvidia drivers is low.
DXO at best appear to have chose to use a different AI approach than. Its possible using something that used an unusual use of Drivers such as NVIDIA and others. As such its clearly causing a lot of problems and if they chose a different approach to everyone else being late in the game and ether didn’t update there drivers or chose not to adapt to changing drivers its really a bit rich to blame these producing the drivers.
Well… here’s a happy accident. I’ve been testing On1 and Darktable neither of which is as good as PL8 in terms of speed, convenience and results. I figured I’d just use PureRaw and get a better editor, but…
Before giving up, I changed the AI acceleration in Preferences to "Intel(R) Arc™ Pro Graphics, restarted Pl9 and VOILA! Everything exports nicely and quickly even with more then one AI keyword mask, with DeepPrime XD/XD2s enabled, and with other effects and global adjustments like creative vignetting. The INTEL driver shares my system memory, and though I have 64 G it seems to reserve 32 G and leave the rest alone.
So for me, problem solved. I hope it works out for everyone who is dependent on NVIDIA graphics. For some reason I have BOTH on my laptop.
And I am still deeply disappointed in DXO for releasing what is clearly a poorly tested and obviously buggy upgrade. Not everyone has or can afford NVIDIA cards with more than 16 g of VRAM!
Seems 12GB PL works pretty good, even if everything enabled, used. Seems even with 10GB (Intel ARC) as i read feedback in here in threads works fine. So, more than 16GB VRAM definitely not required.
Each thing use some GPU VRAM amount. Seems DP XD2s use more than DP3. AI ‘pre-defined’ (keyword, like: Sky) mask definitely use a lot (actually that’s the most ‘killer’) → Manual AI masking use far less. Loupe/Deepprime rendering use some. Export use also some, etc. So, its can work with 6-8GB, if you use like: For AI masking: only ‘manual’ AI masks , only DP3 and no DP XD2s, no loupe, etc. It’s even work with 4GB of VRAM! (export works with a little workaround). DxO release note also speak about only ‘AI mask’ and not ‘pre-defined’ AI mask, talk only DP3, etc.
I write this a tons of time. All is about GPU VRAM amount - and of course GPU performance affect overall performance working with AI masks - newer chips faster, higher clock faster, etc.
Yep, now i see, we already talk about 8GB GPU in other thread. So, you see, GPU VRAM (or whatever its can use) is the thing.
My system doesn’t specify more about the Intel Arc other than Driver date is 2025-02-26 and the driver version is 32.0.101.6637.
As far as export speed goes I haven’t put a stopwatch to it but it is around 30 seconds for a complex mask. Doesn’t seem much different for less complicated exports
