• Welcome to the Lightroom Queen Forums! We're a friendly bunch, so please feel free to register and join in the conversation. If you're not familiar with forums, you'll find step by step instructions on how to post your first thread under Help at the bottom of the page. You're also welcome to download our free Lightroom Quick Start eBooks and explore our other FAQ resources.
  • Stop struggling with Lightroom! There's no need to spend hours hunting for the answers to your Lightroom Classic questions. All the information you need is in Adobe Lightroom Classic - The Missing FAQ!

    To help you get started, there's a series of easy tutorials to guide you through a simple workflow. As you grow in confidence, the book switches to a conversational FAQ format, so you can quickly find answers to advanced questions. And better still, the eBooks are updated for every release, so it's always up to date.
  • 14 August 2024 It's Lightroom update time again! See What’s New in Lightroom Classic 13.5, Mobile & Desktop (August 2024)? for the bug fixes. Hopefully this will fix many of the sync issues reporting in Classic 13.3 and 13.4.

Sonoma fix? Adobe/DxO found bugs in Apple Mac Neural Engine on Ventura

Status
Not open for further replies.
B

bakubo

Member
Joined
May 16, 2023
Messages
68
Lightroom Version Number
12.5
Operating System
  1. macOS 13 Ventura
  2. macOS 14 Sonoma
We had a thread here months ago about Adobe and DxO publicly saying that Ventura introduced Neural Engine bugs. Adobe said that they could not use the NE for Denoise AI until Apple fixed the bugs. DxO said the bugs are not in Monterey. Here is the thread that I cannot update:

Adobe has found bugs in Apple Mac Neural Engine​

https://www.lightroomqueen.com/comm...-found-bugs-in-apple-mac-neural-engine.47745/

It seems that despite there being many Ventura updates since Adobe and DxO reported the problems to Apple last year the NE bugs never got fixed. Are they fixed in Sonoma?
 
Sort by date Sort by votes
Denoise on My 2020 MacBook Air M1 is not any faster with Sonoma.
 
Upvote 0 Downvote
Thank you for that update. Not what I hoped hear though. Now I am hoping that Apple has fixed the broken Neural Engine and it is just a matter of LrC 12.6 being released to finally make use of it. When/if Adobe can use it then it is likely that on Apple Silicon Denoise AI will speed up. See the link I provided above for more complete info.
 
Upvote 0 Downvote
There will likely be no releases until version 13 which should be out within the next few weeks.
 
Upvote 0 Downvote
Okay. The release number does not matter to me. I just hope the next release is finally able to use the Neural Engine like Adobe had planned to from the beginning. Assuming that the Ventura Neural Engine debacle has been fixed in Sonoma. And it is not safe to assume with Apple. :)
 
Upvote 0 Downvote
Me neither. I just didn't there there would a 12.6. I hope it is fixed too. I'll see when the next update/version comes out.
 
Upvote 0 Downvote
Unfortunately, it seems likely the news is bad. I just saw a report from someone who tried LrC 13.0 Denoise AI on an Apple Silicon Mac with Sonoma and the times were the same as they have been since the beginning of this Apple/Adobe/DxO debacle so it seems likely that the Neural Engine is still not being used.

Apple and Adobe have been very quiet about this for months so I think the message that is being sent is that Apple has no plans to ever fix these bugs. New emojis for MacOS are much higher priority, I suppose.

It seems to me that Eric Chan really should update his Denoise AI paper that he published on 2023/4/18:

https://blog.adobe.com/en/publish/2023/04/18/denoise-demystified

He wrote this, but soon afterwards Adobe said they could not use the Neural Engine until Apple fixed new bugs in Ventura (DxO also wrote about their troubles with it in Ventura, but no problem in Monterey):

Finally, we built our machine learning models to take full advantage of the latest platform technologies, including NVIDIA’s TensorCores and the Apple Neural Engine. Using these technologies enables our models to run faster on modern hardware.

The paper really should be updated to remove mention of the Apple Neural Engine or add a note that until Apple fixes the Neural Engine bugs that Adobe cannot use it.
 
Upvote 0 Downvote
14.1 is out today but I have not installed it yet.
 
Upvote 0 Downvote
Yes. Adobe Denoise wasn't any faster on the M1.
 
Upvote 0 Downvote
Zenon said:
Yes. Adobe Denoise wasn't any faster on the M1.
Click to expand...
Thank you for the info about 14.1. It is clear to me that Apple never intends to fix the Neural Engine bugs introduced in Ventura over a year ago. There have been about 16 Ventura releases and 2 Sonoma releases. The message is that a core Apple Silicon component is broken for Adobe, DxO, Topaz, and probably others and it will not be fixed. New emojis, desktop photos, etc. are a much higher priority. Randomly ejecting external drives for no reason which is a bug introduced in Big Sur has also still not been fixed. Ongoing Bluetooth problems. And there are more longstanding, basic bugs that Apple will likely never fix. Oh well.
 
Upvote 0 Downvote
Topaz Photo AI has a problem for a long time where it sometimes produces results with a checkerboard/tiling problem. I asked on the Topaz forum if this is the same problem DxO reports and one of their engineers replied and said they use the Apple Neural Engine and that is why Photo AI sometimes creates checkerboard/tiles:

"We do use the Neural Engine for Apple Silicon devices which is why this issue happens."
 
Upvote 0 Downvote
For those who do not know about the Apple Silicon GPU and the similar Nvidia Cuda cores and Tensor cores which are similar to the Neural Engine then see this:

Video: What are Apple's GPU cores?

https://www.youtube.com/watch?v=-TOdEjcFldI

Ever wonder why Apple lists their GPU Cores like 64 and 76, but an Nvidia RTX 4080 has 8,704 CUDA cores?

At the highest level is the Graphics Processing Unit, aka the GPU which is a parallel processor, which is optimized for multiple instructions to be executed in parallel, opposed to traditional CPUs being optimized for sequential processing. The data is funneled to streaming multiprocessors, which in Apple's vernacular are referred to as cores.

In Apple Silicon, at least for the M1 series, each core is split into 16 Execution Units, each with 8 Arithmetic Logic Units (ALUs). For example, the top end M1 Ultra has 64 Cores with 1024 Execution Units or 8192 ALUs.

Minor Correction: In the video, I mentioned Intel Macs had a shared memory design which is true, which I really should have clarified as the GPU has a reserved pool of RAM and not the same as Apple Silicon which is a much superior design as the processing (vertex shaders, pixel shaders, texture units) are part of the same pipeline.

Neural Engine - Nvidia CUDA Cores vs. Tensor Cores: What's the Difference?
You'll find both on your Nvidia GPU, but what do they do?

https://www.makeuseof.com/nvidia-cuda-cores-vs-tensor-cores-whats-the-difference/
 
Upvote 0 Downvote
Today Topaz updated the issue with the Neural Engine:

https://community.topazlabs.com/t/topaz-photo-ai-v2-1-1/55916

There is also an update for fixing tile artifacts in some upscaling models (Standard / High Fidelity). This should be automatic for Windows, but on Apple Silicon Mac machines you will need to use the new “Use Neural Engine” toggle that shows underneath the upscale model selection to fix it. Since the issue is caused by Neural Engine, disabling it should fix the tile issues. However, doing so will make processing slower.

There is also the NE problem of color changes that DxO has mentioned, but Topaz does not talk about that.
 
Upvote 0 Downvote
On another forum there is an interesting update concerning LrC and the Apple Neural Engine debacle.

He reports that LrC 12.5 Denoise AI using Monterey (Neural Engine still worked in Monterey) took 10-11 seconds on his 45.7mp Nikon Z9 photos. Now with LrC 13.0.1 using Sonoma it takes 25-30 seconds. I suspect Adobe detects that Monterey is running which has the working Neural Engine so uses it, but with Sonoma and Ventura it uses the much slower GPU. In this case it shows the GPU takes 3 times longer.

I tried LrC 12.3 using Ventura and 13.0.1 using Ventura on the same 45mp Canon R5 photo. With 12.3 it took 34 seconds and with 13.0.1 it took 38 seconds so 12% longer. Topaz Denoise AI takes 9 seconds. Topaz has said in the past they use the Neural Engine and they also seem to care less about problems. They seem to continue to use the Neural Engine for the speed and accept the Neural Engine problems that sometime show up in images. DxO says they implemented some "temporary" kludge that allows them to use the Neural Engine and then fix the problems it produces (which probably makes it a bit slower). Since it appears the Neural Engine bugs are now permanent and will never be fixed then maybe Adobe should do something similar.
 
Upvote 0 Downvote
I forgot to mention in the post above that he is using an M1 Max 16" Macbook Pro 64gb. I am using an M2 Pro 16" Macbook Pro 32gb.

Here is one more bit of interesting information. He reported also that on his Windows 11 PC (i7-8900K, 64GB, Nvidia RTX 3070ti) that it takes 11 seconds, so the same time as on his M1 Max when on Monterey. As Adobe's Eric Chan wrote months ago they use the Nvidia Tensor cores (similar to Apple Neural Engine -- except the Tensor cores work).

The M1 Max, M1 Pro, and M1 all have the same 16-core Neural Engine so we can expect that all of them on Monterey would be about 11 seconds, I think. The M2 Max, M2 Pro, and M2 all have the same 16-core Neural Engine and Apple says it is 40% faster than the one in the M1 series. So, we can make a reasonable guess that on Monterey the time would be about 6.6 seconds. The M3 Max, M3 Pro, and M3 all have the same 16-core Neural Engine and Apple says it is 15% faster than the one in the M2 series. So, we can make a reasonable guess that on Monterey the time would be about 5.6 seconds.
 
Upvote 0 Downvote
Hmm, maybe M3 on Monterey (or a fixed Ventura/Sonoma) would be more like 4.4 seconds. Apple says that the M3 Neural Engine is 60% faster than the M1 Neural Engine. Does this sound right?

Of course, all of these are just guesstimates. It is possible that the Neural Engine is used in conjunction with the GPU, not totally replacing it.

Screenshot 2023-11-01 at 6.42.50.jpg
 
Upvote 0 Downvote
With today's release, the Apple Neural Engine is active in the Adobe desktop applications. (Camera Raw, Lightroom Classic, Lightroom Desktop)
 
Upvote 0 Downvote
Status
Not open for further replies.
Back
Top