Problem: after editing in LR 1-2 hours dramatic slowdown
Source of problem: Nvidia 4090 memory fills up. 20/24gb is still fast. 23/24gb sloooooow. Only way to empty memory? Close Lightroom.
Note: initially running two monitors off of Nvidia 4090
Solution: unplug 5k editing monitor from Nvidia 4090, plug into motherboard Display Port
Reason for post: I want to understand why?
Observations: with 5K monitor plugged into motherboard, the igpu inside of the 13900k operates the monitor. But LR brushes etc run on 4090. Confirmed. 4090 memory now utilizes around 16-22gb but never approaches 24gb.
Note: smart pc builders told me none of this is even remotely possible and that I’m ignorant. Hey maybe they’re right. But I need a more civil platform to discuss and learn. I’d really like to understand why! Thanks!
Yeah I hear ya. Seems like an Nvidia driver update could resolve this. Or maybe a small revision to LR code. But maybe Nvidia/Lightroom users = very small use case
One PC builder (everyone is an expert) said there’s no problem, everything is working as intended, but LR workflows “simply need more vram. I think this might be true. And it explains why my solution is effective imo.
Perhaps this is another explanation why Apple prefers unified memory. You aren’t likely to run out of vram when your Mac has 96gb of shared memory.
But for now, my solution provided me 8 hours of continuous LR work with zero slowdowns. Hopefully this thread helps future PC users.
ohsnaphappy wrote:
One PC builder (everyone is an expert) said there’s no problem, everything is working as intended, but LR workflows “simply need more vram. I think this might be true. And it explains why my solution is effective imo.
Perhaps this is another explanation why Apple prefers unified memory. You aren’t likely to run out of vram when your Mac has 96gb of shared memory.
But for now, my solution provided me 8 hours of continuous LR work with zero slowdowns. Hopefully this thread helps future PC users.
This is EXACTLY why I left PC and went to Mac for LR use. I've never bought the "it just works" line from Apple, but in this one case, it simply does
One thing I'm curious about, even my old 1060 had a slew of output ports. Have you tried different combos of DP and HDMI to see if the performance issues persist? I'm not sure if all output ports are on the same bus on the card or not.
Also, if you have 128gb of system ram, you can increase your cache in Camera Raw/LR significantly past the 5gb default. Some might even tell you to change that to 50gb.
There seems to be a long-standing memory leak in Lightroom related to the GPU memory use. In my experience it's somehow related to some of the tools, in particular the healing brush tool. If you do a lot of touchup with that tool it will destroy your GPU memory over time. On an Apple Silicon Mac with 128GB of RAM it just never becomes an issue.
I have exactly the same experience. Had it on the 3090 since late 2020, and now on the 4090 for about a year. Nothing changes in the LR world apparently. It's extremely annoying to the point when I just go to the menu and disable GPU acceleration entirely.
There is one thing that I can recommend to reduce GPU-related glitches in Adobe applications: install the most recent Studio Drivers (the default drivers are "Game Ready"). Those are much more stable and geared towards gfx design/rendering apps. It did help me in the past to remove random flickering in LR.
docusync wrote:
I have exactly the same experience. Had it on the 3090 since late 2020, and now on the 4090 for about a year. Nothing changes in the LR world apparently. It's extremely annoying to the point when I just go to the menu and disable GPU acceleration entirely.
There is one thing that I can recommend to reduce GPU-related glitches in Adobe applications: install the most recent Studio Drivers (the default drivers are "Game Ready"). Those are much more stable and geared towards gfx design/rendering apps. It did help me in the past to remove random flickering in LR.
I certainly wouldn't turn the GPU acceleration off; it can massively speed up operations like jpeg exports and the new AI Denoise and masking functions. As you point out, you definitely want to use the Studio drivers; the "Game Ready" drivers can be flaky with most other applications outside of game.
The easy solution to the GPU problem on the Windows side is to just restart LR every hour or so. Annoying but it resolves the problem with taking the performance penalty of turning off GPU acceleration.
jhapeman wrote:
I certainly wouldn't turn the GPU acceleration off; it can massively speed up operations like jpeg exports and the new AI Denoise and masking functions. As you point out, you definitely want to use the Studio drivers; the "Game Ready" drivers can be flaky with most other applications outside of game.
The easy solution to the GPU problem on the Windows side is to just restart LR every hour or so. Annoying but it resolves the problem with taking the performance penalty of turning off GPU acceleration.
It's not that bad. The i14900 alone (GPU = off) can denoise a 50Mpx file in ~11-12 seconds (estimated time 5 s). I'm getting about the same time with the 4090 (and exactly the same 5s estimate). Scrolling and zooming is not smooth without GPU but I can live with that. I export into PNG and it's very slow regardless of the GPU (apparently it's not optimized at all). To me the most annoying part is when those slowdowns start happening and I'm in the middle of intense editing. I can't restart LR instantly because it usually takes ~5+ seconds to shutdown the current instance. It won't allow to start a second instance while the first one is still running. It is pain, but I still like this machine because unlike my MBP it got some serious storage (HDDs) where I keep all my files. Recently I upgraded my RAID card that would allow to connect 24 SAS drives (the chassis can only fit 16, but now it has some room to grow ).
Long story short - nowadays I'm trying to avoid LR for anything but culling, denoising and exporting while other "intense" stuff is done in PS.
Interesting thread. I do not see the problems that you all seem to be experiencing. I am running a 3090/24GB and don't see the GPU memory being fully consumed. I set up to continuously de-noise a few hundred A1 files and the GPU memory consumed is stable at ~5.4/24 GB of dedicated GPU memory and 0.4/32GB shared memory. It is processing files at about 11-12 sec per file. I am using the latest Studio driver and an older AMD 3900X CPU, 64GB memory.
I think the memory consumption is due to heavy use of brushes and ai. Aka local adjustments. So in each photo you’ve detected lips, teeth, wedding dress, skin, sky, background. Then on top of that you’ve used the heal brush on some fly away hairs & some random people in the background. K. One photo done, hundreds to go.
Although Apple does have an advantage with unified memory, I’m very skeptical about an M2 Ultra outperforming the 4090. I can’t even begin to explain how fast the heal brush is with the 4090. But to be fair, I’ve never used Apple silicon.
ohsnaphappy wrote:
I think the memory consumption is due to heavy use of brushes and ai. Aka local adjustments. So in each photo you’ve detected lips, teeth, wedding dress, skin, sky, background. Then on top of that you’ve used the heal brush on some fly away hairs & some random people in the background. K. One photo done, hundreds to go.
Although Apple does have an advantage with unified memory, I’m very skeptical about an M2 Ultra outperforming the 4090. I can’t even begin to explain how fast the heal brush is with the 4090. But to be fair, I’ve never used Apple silicon. ...Show more →
Note I said the M3 Ultra, not M2 and I was only referring to DeNoise as its one of the only tasks you can time and benchmark. The M2 Ultra on DeNoise is close the 4090 but doesn't beat it. That said, most of the horsepower in the Nvidia GPUs is totally useless for the majority of what we do; that's why for daily use an older 3090 or even a lowly 3060 is perfectly fine.
I have a mix of machines in my office and we do a ton of photo editing here; I have PCs with AMD and Intel processors, and a whole suite of Nvidia graphics cards including RTX 4080s and RTX 4090s as well as every iteration of the M1 Ultra, M2 Ultra and a newer M3 Max Apple silicon computers. I can tell you that in regular daily use like browsing the library and doing minor edits there is no noticeable performance difference between the M2 Ultra and the PC has the current top of the line Intel CPU and RTX 4090 graphics card. There are some significant differences in performance for tasks like preview rendering and jpeg exports, however, where even the older M1 Ultra outperforms the latest and best from Intel and Nvidia.
I know a lot of PC folks have a hard time believing this, but the numbers don't lie. Now, if you wanted to play a game on your Mac vs the PC, well that's a different story, but right now if you want the absolute maximum performance from a computer for Lightroom use, even the older M1 Ultra will outperform a new top of the line PC. I will refer folks to the testing thread I have shared here where I directly compare these machines editing the exact same batch of files.
ohsnaphappy wrote:
I think the memory consumption is due to heavy use of brushes and ai. Aka local adjustments. So in each photo you’ve detected lips, teeth, wedding dress, skin, sky, background. Then on top of that you’ve used the heal brush on some fly away hairs & some random people in the background. K. One photo done, hundreds to go.
This is 100% where the problem lies--the local adjustment brushes. It has to do with how LR saves all of the edits so that you can do non-destructive editing. It keeps them loaded up in memory while you're working on any given image and also seems to have a memory leak so if you edit say 100 images and do a ton of the spot cleaning of dust (we do a ton of this with product photography) at some point it will just reach the point where the GPU runs out of memory and everything starts to choke. The more memory the GPU has the longer you can go before this happens. If you don't do this type of editing you may never notice the problem at all, as it won't affect you.
For my business this was a huge problem as we are taking macro images of products and no matter how careful you are about keeping dust off of the products you still end up having to edit out a lot of little dust spots. Before Apple Silicon came along I had a 2019 Mac Pro with an AMD Radeon Pro W6800X Duo GPU; while that was basically total overkill, the 64GB of GPU memory meant that we could edit almost all day without having LR hang due to the GPU memory leak.
I've never had the problem since with Apple Silicon, probably because I suspect the code is totally different, because monitoring usage of RAM over a day I don't see any of the creeping consumption of memory that one would associate with a memory leak.
docusync wrote:
It's not that bad. The i14900 alone (GPU = off) can denoise a 50Mpx file in ~11-12 seconds (estimated time 5 s). I'm getting about the same time with the 4090 (and exactly the same 5s estimate). Scrolling and zooming is not smooth without GPU but I can live with that. I export into PNG and it's very slow regardless of the GPU (apparently it's not optimized at all). To me the most annoying part is when those slowdowns start happening and I'm in the middle of intense editing. I can't restart LR instantly because it usually takes ~5+ seconds to shutdown the current instance. It won't allow to start a second instance while the first one is still running. It is pain, but I still like this machine because unlike my MBP it got some serious storage (HDDs) where I keep all my files. Recently I upgraded my RAID card that would allow to connect 24 SAS drives (the chassis can only fit 16, but now it has some room to grow ).
Long story short - nowadays I'm trying to avoid LR for anything but culling, denoising and exporting while other "intense" stuff is done in PS....Show more →
Good point, it did take forever to close. When the problem plagued me I would learn to shut down LR before it got too slow. I personally only use external storage so I can be free to edit wherever I please; with the compact size of NVMe storage I can fit 8TB in a pocket with room to spare. It's one reason I'm happy that TB is finally easily found on a PC now, as that makes it so much easier to connect very high speed storage, something we've had the good fortune to have on Macs for quite a few years now.
jhapeman wrote:
This is 100% where the problem lies--the local adjustment brushes. It has to do with how LR saves all of the edits so that you can do non-destructive editing. It keeps them loaded up in memory while you're working on any given image and also seems to have a memory leak so if you edit say 100 images and do a ton of the spot cleaning of dust (we do a ton of this with product photography) at some point it will just reach the point where the GPU runs out of memory and everything starts to choke. The more memory the GPU has the longer you can go before this happens. If you don't do this type of editing you may never notice the problem at all, as it won't affect you.
For my business this was a huge problem as we are taking macro images of products and no matter how careful you are about keeping dust off of the products you still end up having to edit out a lot of little dust spots. Before Apple Silicon came along I had a 2019 Mac Pro with an AMD Radeon Pro W6800X Duo GPU; while that was basically total overkill, the 64GB of GPU memory meant that we could edit almost all day without having LR hang due to the GPU memory leak.
I've never had the problem since with Apple Silicon, probably because I suspect the code is totally different, because monitoring usage of RAM over a day I don't see any of the creeping consumption of memory that one would associate with a memory leak. ...Show more →
the solution reported by @ohsnaphappy appeared to have prevented the gpu from running out of memory. so, i'm wondering if local brushes adjustment is the sole culprit?
"...Observations: with 5K monitor plugged into motherboard, the igpu inside of the 13900k operates the monitor. But LR brushes etc run on 4090. Confirmed. 4090 memory now utilizes around 16-22gb but never approaches 24gb."
tntcorp1 wrote:
the solution reported by @ohsnaphappy@ appeared to have prevented the gpu from running out of memory. so, i'm wondering if local brushes adjustment is the sole culprit?
"...Observations: with 5K monitor plugged into motherboard, the igpu inside of the 13900k operates the monitor. But LR brushes etc run on 4090. Confirmed. 4090 memory now utilizes around 16-22gb but never approaches 24gb."
Interesting option; I used to run two GPUs as well and I never thought about that being the solution. It would imply that there's a memory cap for the brushes, etc. but that cap doesn't leave enough GPU memory to run a higher-resolution display as well. That said it seems that fixing this is very low on Adobe's priority list as this has been around as a problem for years and years now.
tntcorp1 wrote:
the solution reported by @ohsnaphappy@ appeared to have prevented the gpu from running out of memory. so, i'm wondering if local brushes adjustment is the sole culprit?
"...Observations: with 5K monitor plugged into motherboard, the igpu inside of the 13900k operates the monitor. But LR brushes etc run on 4090. Confirmed. 4090 memory now utilizes around 16-22gb but never approaches 24gb."
It is interesting that the use of the GPU to drive the display seems to cause the GPU memory to fill and slow down the computer. I have tried everything to fill the GPU memory and cannot get it to fill and slow down the system, but the reason I don't see the slowdown may be that I am driving a 2K NEC. I can get the GPU memory up to ~20GB but not higher. Also periodically the GPU memory clears and drops down to somewhere in the range of 5-15GB. Sometimes it clears due to me changing to a new file, but usually it seems to occur for no apparent reason.
The observations in this thread are for GPU's with 24GB memory. What happens with a GPU with less memory? Is this a more severe problem?
dclark wrote:
It is interesting that the use of the GPU to drive the display seems to cause the GPU memory to fill and slow down the computer. I have tried everything to fill the GPU memory and cannot get it to fill and slow down the system, but the reason I don't see the slowdown may be that I am driving a 2K NEC. I can get the GPU memory up to ~20GB but not higher. Also periodically the GPU memory clears and drops down to somewhere in the range of 5-15GB. Sometimes it clears due to me changing to a new file, but usually it seems to occur for no apparent reason.
The observations in this thread are for GPU's with 24GB memory. What happens with a GPU with less memory? Is this a more severe problem? ...Show more →
It was a huge problem for me on my 2019 MacBook Pro which had an AMD discrete GPU with only 8GB of VRAM.
dclark wrote:
It is interesting that the use of the GPU to drive the display seems to cause the GPU memory to fill and slow down the computer. I have tried everything to fill the GPU memory and cannot get it to fill and slow down the system, but the reason I don't see the slowdown may be that I am driving a 2K NEC. I can get the GPU memory up to ~20GB but not higher. Also periodically the GPU memory clears and drops down to somewhere in the range of 5-15GB. Sometimes it clears due to me changing to a new file, but usually it seems to occur for no apparent reason.
The observations in this thread are for GPU's with 24GB memory. What happens with a GPU with less memory? Is this a more severe problem? ...Show more →
@dcclark - i'm curious which model of nvda gpu are you running?
).