RTX 2060(6GB) have been something of a mixed bag and my conclusion is that AI, exporting and a 6GB GPU don’t mix,
I think its in-line (6GB) what i guess (think) a few days back it this thread:
PL 9 not really ready for release!
I copy the relevant part (from my comment), and add highlight with BOLD:
- CPU only everything works fine, AI subject detection, etc. Off course performance as it is.
- PL use for AI masking modell about 3-4GB of GPU memory
- For DP3 the AI noise reduction model may use 1-3GB (seems like 2GB), seems XD2s may use less GPU memory.
- With GPU - IF - you have only 4GB or 6GB VRAM, seems problems started
- If you have even 1 (one) enabled AI mask, the ‘AI mask’ model loaded to GPU memory and STAY in GPU VRAM there. Seems AI model memory size is 4GB.
- I guess 4GB VRAM its enough for AI mask quality and performance. Photos has limited sizes (pixel count), so i guess, its not hundred billion parameter (at least not for now)
- In few photos i see whole AI model runs in the backgrund, as its just select ‘people head’, its some cases also select hands, etc. So its runs subject detection, everything.
- I guess 'Subject", ‘Sky’ and similar pre-sets takes more GPU VRAM memory then if you select the subject with selection - as its need to scan the whole image. May that’s the reason why with my 4GB VRAM its never find any subject, sky, with one exception: hair (may its easier to find, may works only in a partially bald people)
- AI models far-far-far faster if its runs in GPU VRAM (and NOT in computer main memory). So, i guess performance is the reason, why only GPU VRAM used (and not GPU shared memory, as it can be far-far slower, and may not all model engine / GPU driver support shared - at least in not encounter with this in my computer)
- If you last selected photo has AI mask, when PL starts (in this last photo), AI mask model automatically load to GPU VRAM, So automatically use 4GB GPU VRAM (as the photo has at least one enabled AI mask)
- If you last selected photo Dont has AI mask when PL starts (in this last photo), AI mask model NOT loaded to GPU VRAM
- If you has 4GB GPU memory, seems AI masks itself its on the limit on works.
- But if you want (with 4GB VRAM) AI mask AND also export the photos ->Export crash. I guess, its just simply runs out of GPU VRAM - AI mask (4GB) + Export (2GB) > 4GB.
- I guess, GPU with 6GB VRAM its a minimum for AI mask (4GB) + Export (2GB). I guess, even its on the edge.
Summary: seems the 4GB of GPU VRAM in the edge of AI mask model, and not enough for export if you has even 1 (one) AI mask - as AI mask modell loads up approx 4B of GPU VRAM, and stay in there, and don’t leave any free GPU VRAM. May, just may GPU with 6GB of VRAM is near to edge for AI mask + Export.
Add to that (in the meantime) a few guess/observation:
- Please write the GPU VRAM memory size for the issue. As “i have a XYZ GPU” may not enough to get this info. It’s can help the observations.
- As far as i see, issues not happen with GPU with 12GB. Also not with 8GB.
- Issues definitely raised with 4GB GPU
- Actually i guess its works fine with 6GB. May you need to ensure GPU VRAM free memory is maximized as possible. May not a bad idea if you close another applications as possible - eg. not run LR or similar in parallel, or don’t use like YouTube playback (where browsers usually use HW/GPU acceleration, etc.). So i guess 6GB is works, but its on the edge.
- I don’t think the GPU is ‘new’ or ‘old’, only VRAM amount is the important. If its older, the things performance is off course slower, but just works.
- Regarding Nvidia driver (you need to back older driver version) - I guess, AI model (used by PL9) don’t find entry point / fail on initialization in the Nvidia lastest GPU driver. I guess the reason not exactly in DxO testing method, personally i think they tested fine with new (latest as possible) drivers. But may (just may) the AI model engine has some last minute change in the Release (9.0) versus Beta versions (may not the same as in the Beta versions). Anyhow, it’s an issue but seems workaround (with driver versin downgrade)
- I guess re-validate the AI quality and stability can be time consuming.