My Recent GPU Experience (Gigabyte Nvidia RTX 4070 Ti 12GB)

[Gnits]

Lightroom Version

Classic 12.2.1

Operating System

1. Windows 10

+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

"This is a relatively long discussion. It is worth reading the discussion in sequence, as posted."

To assist people returning to this discussion, the following are direct links to frequently used posts.

Post #20 contains a link to a raw file to use, if you wish to compare the performance of your rig to others.
https://www.lightroomqueen.com/comm...e-nvidia-rtx-4070-ti-12gb.47572/#post-1315509

Post #30 contains a summary of Denoise performance stats for a range of GPU’s and CPU’s from readers of this post.
https://www.lightroomqueen.com/comm...ia-rtx-4070-ti-12gb.47572/page-2#post-1315545

+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++

I am posting this on the Lightroom Classic forum, as I think it is most relevant to Classic users.

With GPU prices softening and the release of the 4xxx range I decided to purchase the Gigabyte Nvidia RTX 4070 Ti 12GB. I was aware, before purchase that I needed a min 750 Watt GPU. I already had a recently purchased Corsair RM850X model and plenty of spare Corsair supplied compatible internal power cables.

The GPU was tricky to install because of its size and the need to install an awkward bracket support system. The GPU was supplied with a Y cable, where the single end was plugged into the GPU and the two other ends plugged into separate power cables (not supplied by Gigabyte) which were then plugged into the PSU. I used Corsair cables, supplied with my PSU.

I triple checked that I had the polarity of all the cable connections correct, square pins fitted into square holes and rounded pins fitted into rounded holes etc. When I powered up my PC….. nothing happened. Nothing. No lights, no fans starting up, no status numbers on the internal motherboard display. Totally dead... and no clue as to what the problem might be. I feared the worst.

I completely disassembled my PC and removed the Corsair power supply. I then followed the Corsair instructions on how to test the PSU. All tests passed and all (28 I think) pin outs had exactly the correct voltage settings.

I rebuilt my rig, this time reinstalling my old GPU. With much relief, it booted perfectly. I was worried that many of my main components, such as motherboard, processor, PSU, etc had been toasted.

I sent a query to Gigabyte, requesting advice and trouble shooting steps, detailing the steps I had taken to test my PSU and my overall rig.

At the same time, I started researching any known issues with this range of GPU’s. I was horrified to discover many incidents where GPU’s, cables and other components had melted following installation of 4xxx GPU’s. My heart sunk. Many of the horror stories pointed to the supplied Y cable as the source. GPU makers pointed towards incorrectly seated motherboards, gpus and or cables. I can confirm that I had triple checked my connections and listened in each case for the ‘click’ as firm connections were made.

I ordered a Corsair 600W PCIe 5.0 12VHPWR Type-4 PSU Power cable , directly from Corsair. It was been shipped from France and would take a week.

6 days passed and two events occurred on the same day. 1). My Corsair 12VHPWR cable arrived and 2). I received a response from Gigabyte. Essentially, Gigabyte told me to check my GPU (which I had already told Gigabyte I had done) or contact the supplier of the GPU (Amazon)).

So, I installed the Corsair 12VHPWR cable I had just purchased, connecting the GPU to the power supply. There was no need for me to triple check all the connections as the connectors would only connect in the correct way. I gingerly turned on power to the PSU and pressed the start button on the PC….. … the fan on the PSU slowly started to turn …. lights appeared within the GPU, numbers started to appear on the motherboard led display and the boot sequence started.

My PC…came to life… with the sound of the 'Hallelujah Chorus' reverberating in the background.



My Summary.

1. The response from Gigabyte was totally unacceptable.
2. In my opinion, the Y cable supplied by Gigabyte was inappropriate for the intended purpose.
3. Gigabytes response was seriously underwhelming and totally ignored the elephant in the room… ie the Y power cable.

My advice to anyone installing a modern, fairly high powered GPU is to triple check the connections needed to support the GPU and procure a single, fit for purpose, cable assembly which connects the GPU directly to the PSU, without any 3rd party connectors. Make sure you have a modern PSU of sufficient wattage and make sure you have sufficient spare slots in your PSU to cater for twin connections if required to power your GPU.

Ps. I have noticed a very pleasing improvement in Lr responsiveness…... very timely… as the third event which occurred on the same day was the release of a new Lr version, with the new GPU focused noise reduction and mask features. Serendipity ??? ‘Sweet’.

I am posting this story in the hope that others will not have to experience the pain and anguish I have done and help avoid a potential catastrophic meltdown of lots of very expensive components. I know every config is different ... but the bottom line is that modern GPUs are power hungry and need cables fit for purpose.

Here is the cable I purchased directly from Corsair. In my opinion, this should be supplied with GPU's which need dual PSU power slots.

 

[Linwood Ferguson]

Keep them coming and I'll put together a spreadsheet. I'm hoping someone else has a card similar to mine, starting to doubt my effort, but cannot see anything I did incorrectly. I assume nothing is cached in a denoise run?

 

[James Waghorn]

This is a very useful exercise and it's quite surprising that the estimated time on my sysem is so inaccurate:
Estimated time: 16 minutes!
Actual time 5 mins 5 seconds (exactly the same as prbimages reported above with the almost identical spec pc)

CPU: Ryzen 5900x
RAM: 32 Gb DDR4 3600
GPU: Nvidia Geforce GTX 1050 Ti 4Gb
OS : Windows 11 Pro

I had Task Manager open whilst running and the CPU was hardly being used - maxing out at just 4.21 Ghz and 3% utilisation, whereas the GPU hit 99-100% utilisation and 3.9GB VRAM. System RAM was maxing out at 14.6 GB.

Time to order an RTX 4070 GPU I think.....

 

[Gnits]

Processor AMD Ryzen 9 3900X
Memory. 64GB
GPU:Gigabyte GeForce RTX 4070 Ti
O/S Win 10

Denoise : Est 15 secs Actual 18 secs

Very similar to post #20 ... which I was expecting as 4070 Ti is close to expected performance of 3080 Ti (but you have a higher spec cpu).

 

[Linwood Ferguson]

Very similar to post #20 ... which I was expecting as 4070 Ti is close to expected performance of 3080 Ti (but you have a higher spec cpu).

That's one of the things I am wondering. I have a 32 core Threadripper vs your 12. Short of finding someone with identical GPU not sure how we take the CPU out of the equation. Not just cores, it likely is more about paths to the GPU, how fast it can load/unload operations to it.

 

[Linwood Ferguson]

Data to date, updated 4/28/2023. I'll try to update this note as new info comes in. Corrections solicited, will leave in the order received for now so you can match up.

 

[Jim Wilde]

IIRC, the estimate for my MacMini test was 3 minutes.

 

[Gnits]

Very interesting to see these results. What jumps out to me so far is :
1. The spectacular difference between 3xxx/4xxx and earlier gpus.
2. I would deem the performance of the M1 Mac Mini to be somewhat disappointing.

For me… when I got my motherboard it was the latest and greatest, with specific focus on handling power to service cpus and gpus. Now the motherboard has AM4 cpu slot, while AM5 will be the standard for AMD cpus in the current release cycle.

I will have have a difficult decision to make sometime if it is worth my while to upgrade my CPU…

It looks to me that the emerging 4070 GPU models (not 4070Ti) may be the sweet spot going forward in terms of price and performance, with options for higher performance above that if desired.

It will be interesting to see the table above expand with a wider range of specs. For me, the trend is already very clear.

I am exceedingly happy with my 4070Ti… major improvement of editing in Develop mode and general use in Library mode. I have a very large panorama (80,000 pixels wide) and saw it pop into sharp display after only a second or so. Yesterday I started to work on my backlog of Sony A7rv images, which were so painful to edit with my old GTX1650 GPU.

I know we are never happy with hardware / software performance…. a form of perpetual motion.

 

[Jim Wilde]

Very interesting to see these results. What jumps out to me so far is :
1. The spectacular difference between 3xxx/4xxx and earlier gpus.
2. I would deem the performance of the M1 Mac Mini to be somewhat disappointing.

Another significant factor in the denoise speed is the amount of memory that the card has. There's a reason why Adobe recommend at least 8GB of GPU memory. A couple of examples:

On my creakingly-old Win10 system, with a GTX 680 card with 2GB or memory, the estimate for the image is 94 minutes! I didn't press the Enhance button, I fear the poor old system wouldn't survive!
On my similar age late-2013 MacBookPro (top of the line specs at that time) with an Nvidia GT 750M GPU with 2GB memory, the estimate was "only" 47 minutes. That one I did press the Enhance button and the system completely froze immediately, requiring a force power off and restart.

It might be a good idea if that data could also be captured here.

Regarding the M1 Mac Mini, "somewhat disappointing" was the same phrase that I used when providing feedback to Adobe. It should be better than that. However, I think I know the reason for the relatively poor performance, and hope to see that change before too long. If it does, I will provide an update here.

 

[James Waghorn]

Up until now Lightroom has run acceptably with a GTX 1050 Ti 4GB, but with all the recent masking features it has been noticeable the GPU is very much the weak link in my system when multiple masks and the healing brush are used. This has been further confirmed with the slow performance with Denoise AI, where CPU usage is virtually nil and the GPU is maxed out at 100%.

Puget Systems still recommend the RTX 3060 Ti 8GB on their website, but as far as I can see, they haven't re-run any of their GPU tests since Adobe dramatically increased the use of the GPU in masking and in particular Denoise AI. I agree with Gnits, that for now, the RTX 4070 12 GB (non Ti) appears to be the sweet-spot for Lightroom/PS, providing some extra power and VRAM compared to the 3060 Ti to future-proof a little longer.

 

[Linwood Ferguson]

I added the M1 estimate though frankly I don't find the estimates particularly interesting -- I mean, are you NOT going to apply because of it.

I haven't been adding GPU memory because I assume each GPU has the memory each GPU has, i.e. the model implied it. Is that not true?

I probably should have asked about @prbimages 1050 TI OC if it was overclocked or just the model, and the reverse if @James Waghorn's was an "OC" model, since I think "OC" is a model variant.

But yes... would be nice to flesh this out with more examples. The RTX 4070 Ti matching my RTX 3080 Ti is reassuring I did not imagine my 18 seconds, but yes, that's a huge gape to the nearest other.

 

[johnrellis]

Another anecdotal data point: My shiny new Macbook Pro (2023, Apple M2 Max, 64 GB, 32 graphics cores sharing system memory) runs Denoise on 6000x4000 Canon R6 II raw in 10 seconds.

 

[johnrellis]

Adobe employee Rikk Flohr said in the Adobe forum that the Denoise time estimates should get better after you've run it a few times.

 

[Linwood Ferguson]

Another anecdotal data point: My shiny new Macbook Pro (2023, Apple M2 Max, 64 GB, 32 graphics cores sharing system memory) runs Denoise on 6000x4000 Canon R6 II raw in 10 seconds.

Any chance you could take the image in reply #20 that everyone else is using and try it, so we are working with the same data. And note comments in there how to time.

 

[johnrellis]

One of the differences between newer and older graphics cards is that the newer ones are starting to include hardware optimizations for running AI models, not just graphics computations. For example, newer Nvidia cards include "tensor cores":

https://www.nvidia.com/en-us/data-center/tensor-cores/#:~:text=Tensor Cores enable mixed-precision,performance computing (HPC) tasks.

Tensor Cores enable mixed-precision computing, dynamically adapting calculations to accelerate throughput while preserving accuracy. The latest generation of Tensor Cores are faster than ever on a broader array of AI and high-performance computing (HPC) tasks. From 6X speedups in transformer network training to 3X boosts in performance across all applications, NVIDIA Tensor Cores deliver new capabilities to all workloads.

"Mixed-precision computing" means computing with 8, 16, and 32 bit integers as well as 32 and 64 bit floating point numbers. This can provide large speedups to workloads that are optimized for it.

I'm wondering if Adobe has optimized the Denoise models for using such newer capabilities, and the Apple and Windows open-source libraries LR uses to execute the models are doing a bad job of running such models on older hardware.

It would be interesting to time the execution of Denoise when Windows and Mac OS (on Intel Macs) are running in safe mode. Safe mode disables the GPU at the OS level, and the AI libraries used by LR automatically fall back to using the CPU. (LR's AI commands ignore Preferences > Performance > Use Graphics Processor.)

With AI masking, using the CPU is consistently about 5x slower than a good GPU, still quite usable for occasional use. If that 5x holds true for Denoise, then we might see CPU execution take on the order of 60 seconds (5 x 12 seconds) rather than 20 - 30 minutes.

I'd normally do this experiment myself, but I can't update my Windows ARM test machine to LR 12.3 due to a Creative Cloud bug.

 

[johnrellis]

Any chance you could take the image in reply #20 that everyone else is using and try it, so we are working with the same data. And note comments in there how to time.

17 seconds (average of three runs). The dialog estimates 20 seconds.

 

[Linwood Ferguson]

It would be interesting to time the execution of Denoise when Windows and Mac OS (on Intel Macs) are running in safe mode. Safe mode disables the GPU at the OS level, and the AI libraries used by LR automatically fall back to using the CPU. (LR's AI commands ignore Preferences > Performance > Use Graphics Processor.)

I may be able to do that later, though safe mode changes a lot of things, and not sure if I can get the same resolution (which is relevant to the test since it includes building the new preview and displaying it).

 

[johnrellis]

To get the same resolution of the previews, set the option Catalog Settings > File Handling > Standard Preview Size to a specific value rather than Auto.

 

[Linwood Ferguson]

To get the same resolution of the previews, set the option Catalog Settings > File Handling > Standard Preview Size to a specific value rather than Auto.

That's actually something I did not consider in these tests either, some people might have old, small monitors. I have two large ones (largest is 3840x1600)

 

[BobT]

Hal,

On further reflection, applying the name DNG to a species of TIFF seems to muddy understanging, not clarify it. Of course Adobe could refine the meaning of DNG.

Yes, I must admit I was rather disappointed to find that DNG is a container rather than a standardised raw format. I accidentally converted some jpegs and TIFFs to DNG. It probably doesn't matter but yes, it does muddy things. I guess I could dig them out, but it would be nice if they where clearly flagged somehow for us dummies.

 

[Linwood Ferguson]

It would be interesting to time the execution of Denoise when Windows and Mac OS (on Intel Macs) are running in safe mode. Safe mode disables the GPU at the OS level, and the AI libraries used by LR automatically fall back to using the CPU. (LR's AI commands ignore Preferences > Performance > Use Graphics Processor.)

Well, this was a fiasco. Adobe's subscription service kept me from doing anything useful. I thought I got 90 days, but it would simply not
let me get into develop. Heck, I couldn't even get the image I was using to display on the preview screen (though interestingly a few others did).

I'll post elsewhere, but so much for Adobe allowing 30 days.

Maybe someone else can figure out how to make that work, I had no luck at all, I tried numerous times, got the warnings, then LR would run a bit, and after 5 minutes or so just completely crash and disappear.

 

[Linwood Ferguson]

OK, safe mode with networking worked, and it took.... ready for this.... 16 and a half minutes as opposed to 18 seconds. And that's with a tiny preview. Oh what a difference a GPU makes. And that's on a 32 core (64 threads), 128g, very fast CPU.

 

[johnrellis]

Very interesting. That's a slowdown of 55x. And 16 mins is comparable to what people on older cards are observing. So there's no advantage in using some of these older cards over just using the CPU.

The AI masking commands only slow down about 4 - 5x when executed on the CPU. This suggests there's something about the Denoise models that's been optimized for the newer GPUs and which executes proportionately much slower on older GPUs and on CPUs (e.g. the mixed-precision arithmetic supported by Nvidia's tensor cores).

 

[prbimages]

I probably should have asked about @prbimages 1050 TI OC if it was overclocked or just the model, and the reverse if @James Waghorn's was an "OC" model, since I think "OC" is a model variant.

Yes, "OC" is part of the model name. I have NOT overclocked it.

I agree with Gnits, that for now, the RTX 4070 12 GB (non Ti) appears to be the sweet-spot

Interesting, James, that we have such similar setups. I put my machine together mid-2021, when graphics cards were horrendously expensive, so went with a lowly-specced card with the idea of upgrading it later. I think the time has come. I'm not sure about the definition of "sweet spot" though - an RTX 4070 is double the price of the 3060 that I am thinking of. I see in the table above that the RTX 3080 Ti and the RTX 4070 Ti both took 18 seconds. It would be good to see more comparison results between these two generations of cards.

 

[James Waghorn]

I'm not sure about the definition of "sweet spot" though - an RTX 4070 is double the price of the 3060 that I am thinking of. I see in the table above that the RTX 3080 Ti and the RTX 4070 Ti both took 18 seconds. It would be good to see more comparison results between these two generations of cards.

Yes, perhaps "sweet spot " is too subjective. In the UK the price is basically mid-point between a 3060 and 4070 Ti. I was also thinking of the 3060 12GB until the 4070 came out, but the former is getting for 3 years old now, so I would prefer to upgrade to a newer model. Having said that, it would be very useful to see some timings for the 3060 card with this particular denoise AI test before splashing the cash!

 

[James Waghorn]

Further to my earlier post, I did some Googling and found several people recommending an RTX 3060 12GB GPU and for example on another forum someone running an Intel i5 CPU upgraded from a GT1030 to RTX3060 12GB and claimed Denoise AI processing time was cut down from 15 minutes to 15 seconds!