The fxguide tech lab was lucky enough to have the new M2 Ultra Studio in use for a few weeks. As a VFX professional, here’s why you care about it.
First, the Studio M2 Ultra is fast, really fast.
Secondly, it’s very good for VFX machine learning…and to prove it, Foundry’s NUKE (machine learning/CopyCat powered) now runs natively on M2 silicon and Apple M2 Studio works with
Let’s start with the Studio M2 Ultra…
Announced at WWDC, fxguide will be testing the new Mac Studio (almost exclusively) in our lab. As most readers know, if there’s one topic fxguide prioritizes first, it’s machine learning (ML). So we’re really excited to see Foundry fully Apple Silicon in Nuke’s next release. Native including PyTorch support for very nice CopyCat ML functionality.
With so much happening in the cloud today, it’s easy to forget that there is a set of major VFX applications that truly benefit from the power of the desktop. Remote cloud rendering and other applications work well, but for a range of applications the overhead of uploading materials for use in the cloud is unwise. For example, his VFX editing for a 4K UHD feature film requires gigabyte versions of shots that can be played back and reviewed without the delays and compression that typically occur with remote viewing.
The new Apple Studio is virtually identical to the new Apple M2 Mac Pro. If you’re transcoding or ingesting huge amounts of material, you’ll benefit from the slots offered by the much larger Mac Pro, but when it comes to applications like VFX editing and ML, Mac Studio is just like the Mac Pro. Is the same. We tested Mac Studio with an M2 Ultra with a 24-core CPU, 76-core GPU, 32-core Neural Engine, and 192GB of unified memory. This is the same configuration. So that it can be installed on Mac Pro. There used to be an argument that adding a graphics card required a tower configuration, but Studio does not. The integrated “graphics card” is roughly equivalent to a powerful high-end NVIDIA card. It comes in the same form factor as the original tiny Mac Studio.
The Ultra M2 is quiet and small. This may seem trivial unless you’ve actually worked on a real project with a bunch of other VFX artists on a typical open floor plan. Fan noise and desk space are at a premium, and the Mac Studio is nearly invisible when compared to larger high-end gaming PCs.
Back to machine learning…
Autodesk Flame and Maya both run native Apple Silicon, while Nuke previously only ran via the CPU on Apple Silicon computers. Flame has strong machine learning, but Foundry’s ML CopyCat has changed the game. Released in Nuke 13 and enhanced in 13.2, Copycat was great in putting the power of machine learning in the hands of artists. Using the principle of providing source frames that match ground truth sample frames, CopyCat learns from a small number of frames how to transform an image in the way the artist intended it to understand or “learn” the entire clip. can be inferred with the Inference node. Nuke 14 expands on this with the addition of Cattery, further expanding the set of applications artists can deploy. CopyCat is used to save time on lotto improvements, comp improvements, rigs and cleanup, beauty work and more. Deblurring, smart vector generation, segmentation, real-time test comp – the list grows almost daily. An experienced artist can use his CopyCat to greatly improve the workflow of the VFX pipeline. Unlike the stable diffuse AI examples, CopyCat is an implementation of controllable, reproducible, artist-driven, comprehensible, explainable AI, offering a whole world of new VFX approaches.
Conclusion: For example, take the Ballerina example ‘style transfer’ demo that Foundry uses to show off Nuke, just run the new Nuke in Mac Studio, load the file and it will work. nothing else is needed. CopyCat runs as promised on Apple Silicon, performing both training and inference. And it’s fast.
It’s unfair and premature to quote stats as this new version of Nuke isn’t even in beta yet, but it can run CopyCat more than twice as fast as high-end PCs with today’s high-end graphics cards. It seems that. It’s important to note that this isn’t just caused by newer Apple hardware. The Foundry has improved his CopyCat internally to make it faster. This new version of Nuke will soon enter private beta. Foundry adds even more features in that release. My educated guess is that SIGGRAPH in LA will be a great time to learn more about all the other features of his NUKE’s next release.
Why Do Foundries Care About Apple Silicon? (Hint: VFX Supervisor + ML)
Some avid PC users may be wondering why the foundries put so much effort into implementing such a powerful Nuke on Apple Silicon. The answer actually has to do in part with how much ML is changing the way we work. The foundry has been receiving numerous requests for a while from VFX supervisors and field staff to port Nuke to Apple Silicon. Here the Apple MacBook Pro seems to be the laptop of choice for most of his VFX supervisors, so I prefer to run the script on the MacBook Pro (even before the latest M2 upgrade) . But it’s more than that.
CopyCat inference makes supervisors think about training inference nodes for sequences instead of shots. For example, he trains on a set of 25 images, and it makes sense to use an inference node not just for that one shot of her, but for all similar shots in the sequence. As with all ML, choosing the right training data requires real skill. Applying a single training node to multiple shots requires careful selection of sample frames that define a training data set wide enough for all relevant shots. In other words, when using CopyCat to get the most out of ML, think in terms of multiple shots. I haven’t tested this yet, but the Foundry dev team tested it with his full CopyCat node running on an old M1 MacBook Pro, and by all accounts it works perfectly. Not surprisingly, the Max M1 doesn’t have the graphics RAM to compete with the M2 Ultra. But Nuke works just fine with his base M1 Pro Apple hardware. This is exactly what managers use when thinking about assigning shots to teams. The M2 Ultra Studio may be the best option for your office, but it’s not the only option. This is important as teams move to new, faster Apple Silicon.
Why is Apple obsessed with VFX?
VFX may seem like a small market for Apple, but ML is not. Apple is moving towards providing solid and robust tools for ML, and Nuke is just one example of a powerful Apple Silicon PyTorch implementation that opens up the use of Apple’s hardware to a wider audience.
Until recently, many ML applications had to run on NVIDIA thanks to NVIDIA’s CUDA. NVIDIA has made outstanding contributions in enabling ML applications. However, there were times when the CUDA GPU application programming interface provided the only option for certain types of GPU ML processing. When the Mac Pro was first released, we were told that the main problem was that the Mac Pro would lock users out of his NVIDIA CUDA solution, thereby excluding Apple from a range of ML applications. reported that there is
The new Apple Silicon solves the CUDA problem. that’s all Solution to Existence It is possible Solved. In this respect, the competition should be good for both companies. Apple’s silicon implementation is power efficient, so it generates less heat and delivers incredible GPU power in a small form factor. Small form factor comes at the cost of not being as upgradeable as plug-in cards, but ask yourself how many times you’ll just upgrade your GPU card instead of updating both your PC and your graphics card. Plug-in cards offer flexibility, but the elegant, fast, and quiet form factor works right out of the box with no fuss, and has a very strong position in VFX as well.
Machine learning is a great application that you can control directly on your desk without using cloud services. Apple provided a superior solution, evidenced by the effectiveness of a fully functional Nuke CopyCat solution. Feel free to say that this only appeals to Apple fanboys. But the Apple Studio M2 Ultra is what makes fanboys. This is a nice implementation.
(You can read the original Mac Studio article about color grading here, and the original CopyCat article here.)
In case you were wondering, this is not a sponsored story, we have already purchased our own new M2 Mac Studio through testing.
