Are there any performance figures for Apple M1 chip?
I would love to see some results from smartphone images like the iPhone 14 Pro Max and the Pixel 7 Pro. Their images tend to be noisier than big sensor cameras.
I tried the new denoise today as well as some other workflows on my new self-built PC. (I recently switched away from Mac to PC due to the cost of the full-spec Macs. I'm pretty glad I did to be honest.) My 16gb M1 MBP wasn't cutting it.
The bare-bones PC cost me around $1200 with a Radeon 6700 GPU (10gb vram). Denoise is very fast. It uses the dGPU 100% estimated 10 seconds but was clearly faster.
Could not be happier with LR Classic enhance performance except for the low-end CPU (1200f). Importing a few thousand old photos took quite a while. Going back to denoise old photos shot above 6400 ISO on old 12mp cameras will give them new life.
Based on this I decided to swap the Intel chip from 12100f to 13600k and upgrade from 32 to 64gb RAM.
To plug the state of PC today, all in all I miss the Mac a little but PC so much cheaper than a new M2 machine. I mean my AMD GPU has almost as much dedicated RAM as my whole M1 Macbook did. Adding 32gb of system memory cost me $68. Ridiculous. I love it.
InlawBiker wrote:
I tried the new denoise today as well as some other workflows on my new self-built PC. (I recently switched away from Mac to PC due to the cost of the full-spec Macs. I'm pretty glad I did to be honest.) My 16gb M1 MBP wasn't cutting it.
The bare-bones PC cost me around $1200 with a Radeon 6700 GPU (10gb vram). Denoise is very fast. It uses the dGPU 100% estimated 10 seconds but was clearly faster.
Could not be happier with LR Classic enhance performance except for the low-end CPU (1200f). Importing a few thousand old photos took quite a while. Going back to denoise old photos shot above 6400 ISO on old 12mp cameras will give them new life.
Based on this I decided to swap the Intel chip from 12100f to 13600k and upgrade from 32 to 64gb RAM.
To plug the state of PC today, all in all I miss the Mac a little but PC so much cheaper than a new M2 machine. I mean my AMD GPU has almost as much dedicated RAM as my whole M1 Macbook did. Adding 32gb of system memory cost me $68. Ridiculous. I love it....Show more →
I found a simple article that describes CUDA and Tensor core differences. Tensor is unique to NVIDIA so, apparently, Adobe has to do some custom programming for NVIDIA cards... I suspect the same for Apple. Tough on a developer to deviate from a common code base to run on different platforms.
For LR users, who do you think is the larger install base? Apple or Windows? I suspect the larger base will get the biggest gains sooner as the code gets refined.
BTW, I keep hovering over ditching Topaz and buying a new card (I have a GTX card now), but Topaz does such a great job for me... and it is really fast... just can't justify it yet.
I found a simple article that describes CUDA and Tensor core differences. Tensor is unique to NVIDIA so, apparently, Adobe has to do some custom programming for NVIDIA cards... I suspect the same for Apple. Tough on a developer to deviate from a common code base to run on different platforms.
For LR users, who do you think is the larger install base? Apple or Windows? I suspect the larger base will get the biggest gains sooner as the code gets refined.
BTW, I keep hovering over ditching Topaz and buying a new card (I have a GTX card now), but Topaz does such a great job for me... and it is really fast... just can't justify it yet.
This might surprise people but the larger install base for Adobe is Apple. Why? Because it's the exact same GPU and Neural Engine architecture on the iPhone, iPad and now all of the Apple Silicon Macs. That's part of the reason why Adobe spun so fast to Apple Silicon native versions of their apps when Apple made the switch--any app they have iOS/iPadOS versions of have a ton of native code that's been optimized over years.
Apparently it's not all that difficult to do the development anyway; Topaz is a great example of rapid implementation of Apple Silicon native versions of their apps that directly leverage the GPU and neural engines as well.
To your point on gains, right now the denoise function is optimized for the Tensor cores; Adobe explicitly states that and performance data backs it up--DeNoise is terribly slow if your GPU doesn't have them. Interestingly enough, on Apple Silicon it's currently NOT using the neural engines, as there's a bug in MacOS Ventura that needs to be fixed for denoise to work correctly. Adobe states that performance will significantly improve when that is addressed. In the meantime, my M2 Max MacBook Pro has performance similar to an Nvidia RTX 4070 GPU and my Ultra approaches the performance of an RTX 4090--all without the neural engines enabled. That's pretty impressive and I can't wait to see how good it gets when the bug is fixed.
Well, it's really good of course. Per my post it's giving new life to my old photos. It seems surprising that Adobe didn't update this sooner since the competition has been doing AI noise removal for a long time now. Maybe more surprising I never took the step to purchase a 3rd party plugin. But anyway I'm glad it's baked into LR now.
I sorted by Tensor cores and came up with an efficiency number on the RH column. The value does not mean anything, just compare GPU to GPU if interested. What we care about is the Actual column. Might be interested to add a column for $$s to seconds but costs vary too much. No data for the Mac's here. If anyone wants to download the same test file (link below) and run it on their Mac or other RTX cards I'll add the numbers. My M1 Max MBP with the low end GPU core count took 38 seconds.
Ondrej wrote:
Short test with A7RIII 42Mpx RAW files - it takes about 11-12 minutes on my computer (default 50%) - i7 9700K, 32GB RAW, SSD drives, integrated GPU.
Thanks for the chart, that is pretty helpful info. Makes my GTX1050TI look pretty sad, heh. But it also keeps the 3060TI looking like the best bang for the buck still.
What I find interesting is the difference between TI and non TI cards, because in many cases, depending on the series, you can only get the NON TI cards with more RAM, yet, the TI cards are still faster even with 8gb vs 12GB, I guess from what you have shown is because they usually have more tensor cores or whatever... so, there is that. Plus, I have also noticed depending on what brand/flavor of card you buy, some have DDR6 and others have DDR6X memory, and different stated memory speeds and bus widths which I am not sure your chart takes into consideration or if it even matters? Just throwing it out there. I'm guessing that all comes into play at some point, maybe with batch processing more so than the single test image.
I keep putting off upgrading my CC 2015, because I own it, and am using the Topaz suite and new Photo AI... but at some point when things are more certain, it looks like I'm going to grab a new video card and want to try the new Photoshop since mine is severely outdated now, and I would really like to use the new masks and ACR w/AI. Fortunately, working speeds are tolerable, though still not exactly zippy, with my current software / hardware combination.
I originally started up a chart with clock speeds and GPU ram but it was too much work and there seemed to be different flavours for the same card (maybe from different manufacturers?) as well as memory bandwidth and speed, base and boost clock speeds. In any case I dropped it for this simpler approach. Even though there's not a lot of data you can see that the 20 series don't do as well even with their high core count. Nvidia call them Gen 2/3/4 for the 20/30/40 series so it's possible they have different designs to improve efficiency.
I'll post it later but I started a separate table for M series Macs that show the speed to correlate quite well to GPU cores.
Also, Passmark appears to have up-to-date pricing for cards... my spot checks show they denote the cheapest price available for each card. BTW, their performance benchmarks are pretty good for real world but given Adobe is writing custom code for these cards, I wonder how their rankings match when running Light Room.
mcbroomf wrote:
I originally started up a chart with clock speeds and GPU ram but it was too much work and there seemed to be different flavours for the same card (maybe from different manufacturers?) as well as memory bandwidth and speed, base and boost clock speeds. In any case I dropped it for this simpler approach. Even though there's not a lot of data you can see that the 20 series don't do as well even with their high core count. Nvidia call them Gen 2/3/4 for the 20/30/40 series so it's possible they have different designs to improve efficiency.
I'll post it later but I started a separate table for M series Macs that show the speed to correlate quite well to GPU cores.
Yes, I've been eyeballing the 4060 Ti as well, but man, the gaming card reviewers really trashed that card since I guess from their standpoint it should have had more ram or they might as well have stayed with the 3060 Ti for the price and very similar performance.
For our use, the 4060 Ti might make more sense than the for the gamers and find a home in photo editing? I dunno, but one thing I DO LIKE about the 4060 Ti spec is the lower power consumption, putting less stress on the PSU. I hope someone with one provides a result to update your chart soon, heh.
The performance based on your chart it makes me think that since the speed differential going from the 3060 Ti all the way up to the 4070 Ti is only 3 seconds, that the 4060 Ti will only be a very incremental improvement, like in the 21-22 second range for a WAG, but it still might be worth it for the right price to have the more modern card with less power consumption.
Power consumption was a consideration for me as my desktop only has a 650W PSU and I didn't want to add a upgrade for that in addition to the GPU.
Note that the difference between the 3060 Ti and 4070 Ti is 5 seconds not 3 seconds (the 4070 is 3 secs better). Hell of a cost difference too.
Bruce, thanks for posting the links to cost and benchmarks. I'm not going to add them to the table as the cost will change too quickly, but very useful for parsing a final choice after setting an upper threshold on processing time.
I've ordered the list by the number of GPU cores and it seems to hold well. Interesting that the efficiency is different between Max and Ultra
As a reminder if you have a Mac M1 or M2, any configuration that I can add to this table (or PC/Nvidia for the table above) and you feel like testing it then download the file below and run the Denoise test. Start your timer after the Preview has finished, set to 50, and stop as the DNG appears in the filmstrip. I set the Detail sliders to zero but I don't think it affects the time, just the final quality. If you want to run it a few times purge the LR cache (in Preferences) between runs https://drive.google.com/file/d/15UAKmx9IA1k2Zt877ShHeUewCcHquyhu/view
EDIT 6/15 : MBP 16" with M1 Pro and 16 GPU, M2 Ultra with 76 GPU cores and M1 Max with 32 GPU cores added
The thing about Topaz is that most of their apps don’t stick around. Perhaps they have limited employees to support them. Do you remember Simplify, or Glow. Looks like Topaz Studio is gone too. My favorite app was/is Topaz Impressions, the full version, not the Impressions filter in Studio 2. The thing is that I know folks who built businesses around Impressions, left hanging. Fortunately, all of these apps still work and I still use Impressions with presets I built and modified for myself. So, Denoise, Sharpening AI will soon go away, as will Photo AI too at some point. But hey, hopefully, maybe replaced by the new improved Impressions AI.
I've ordered the list by the number of GPU cores and it seems to hold well. Interesting that the efficiency is different between Max and Ultra
I have a friend with an M1 Pro MBP that I hope will do the test for me, and Jeff will fill in some other blanks soon I think., but as a reminder if you have a Mac M1 or M2, any configuration that I can add to this table (or PC/Nvidia for the table above) and you feel like testing it then download the file below and run the Denoise test. Start your timer after the Preview has finished, set to 50, and stop as the DNG appears in the filmstrip. I set the Detail sliders to zero but I don't think it affects the time, just the final quality. If you want to run it a few times purge the LR cache (in Preferences) between runs https://drive.google.com/file/d/15UAKmx9IA1k2Zt877ShHeUewCcHquyhu/view ...Show more →
I'll read back a bit, but I'd be interested to know how these newer M-chip MacBook Pros compare to my somewhat older 2.9 ghz i9 performs. It does switch to an internal Radeon 560X 4 gig graphics card when I connect it to the larger 4K display. But it's a 2018 model, so it's getting a little bit old.
Charlie San wrote:
The thing about Topaz is that most of their apps don’t stick around. Perhaps they have limited employees to support them. Do you remember Simplify, or Glow. Looks like Topaz Studio is gone too. My favorite app was/is Topaz Impressions, the full version, not the Impressions filter in Studio 2. The thing is that I know folks who built businesses around Impressions, left hanging. Fortunately, all of these apps still work and I still use Impressions with presets I built and modified for myself. So, Denoise, Sharpening AI will soon go away, as will Photo AI too at some point. But hey, hopefully, maybe replaced by the new improved Impressions AI....Show more →
It all depends. As long as the company exists we will still be able to download purchased apps. I still have access to AI Clear and B&W effects. I got B&W effects maybe 10 years and they still show my license key. Who knows if it will run the latest OS I have.
OK so the M2 Ultra Studio just arrived and quick testing shows that a number of tasks are faster now, improvements range from 7% in 1:1 preview rending, to 17% in jpeg export speeds, to a much more noticeable 28% improvement in LR AI-based denoise. If you've bought an M1 version, rest easy. If you haven't bought yet, you will be better served by an M2. More details to come.....
