3D Rendering on Apple Silicon, CPU&GPU

[vel0city]

I prefer the look of Corona to Redshift already. The way it handles light and textures is far more convincing and aesthetic to my eye. It has a filmic, natural look out of the box and needs far less tweaking and pushing than Redshift. It just looks good by default. I'd say somewhere between Arnold and Octane for realism and aesthetic, which is pretty much ideal for my taste in looks.

Used the latest daily build for a few solid hours with no crashes or hangs in C4D R24. Never used Corona before but it's intuitive (if you can use renderer A you can use renderer B in my opinion, just get used to its workflow quirks) and I think I'll commit to a license after the very generous trial is up. It's not quite as responsive in the IPR as Redshift, but it looks great and is stable so far which is a fair enough compromise for me.

Redshift is really lagging behind with features. Still no Toon/NPR shaders (Maxon REALLY should have integrated Sketch and Toon into Redshift by now) no random walk SSS, still limited support for C4D gradients etc etc.

I'm really glad to have the option of another M1 native engine and with the current state of the GPU market, I'm all over Corona as a CPU option.

Did anyone have any luck with Octane X?

 

[hifimac]

I prefer the look of Corona to Redshift already. The way it handles light and textures is far more convincing and aesthetic to my eye. It has a filmic, natural look out of the box and needs far less tweaking and pushing than Redshift. It just looks good by default. I'd say somewhere between Arnold and Octane for realism and aesthetic, which is pretty much ideal for my taste in looks.

Used the latest daily build for a few solid hours with no crashes or hangs in C4D R24. Never used Corona before but it's intuitive (if you can use renderer A you can use renderer B in my opinion, just get used to its workflow quirks) and I think I'll commit to a license after the very generous trial is up. It's not quite as responsive in the IPR as Redshift, but it looks great and is stable so far which is a fair enough compromise for me.

Redshift is really lagging behind with features. Still no Toon/NPR shaders (Maxon REALLY should have integrated Sketch and Toon into Redshift by now) no random walk SSS, still limited support for C4D gradients etc etc.

I'm really glad to have the option of another M1 native engine and with the current state of the GPU market, I'm all over Corona as a CPU option.

Did anyone have any luck with Octane X?

Corona is my daily driver. I do a lot of Archviz and you can achieve great indoor looks easily without much fuss. I think it looks better than Redshift.

I love the way the materials are handled. They can be as simple as Standard/Physical for quick tweaks, but the node editor is there for more advanced stuff. Light Mix and the VFB are great for post processing.

The IPR is not as responsive as Redshift on a 3090, but it's not unusable. Rendering seems faster than Arnold. I wish they would pursue hybrid CPU/GPU rendering like VRay.

Aside from some beta bugs, it has always been super stable. Devs are responsive and always pushing things forward with new features.

I think they just changed their licensing model, but existing licenses get 3 render nodes, which is nice if you have some old machines laying around.

 

[LymeChips]

IMO Octane X won’t really shine until the next update where it’ll get the 2022 XB1 beta core. Then we’ll get the much improved ACES workflow, new color pipeline, native support and network rendering with PCs to name a few things.

 

[vel0city]

Corona is my daily driver. I do a lot of Archviz and you can achieve great indoor looks easily without much fuss. I think it looks better than Redshift.

I love the way the materials are handled. They can be as simple as Standard/Physical for quick tweaks, but the node editor is there for more advanced stuff. Light Mix and the VFB are great for post processing.

The IPR is not as responsive as Redshift on a 3090, but it's not unusable. Rendering seems faster than Arnold. I wish they would pursue hybrid CPU/GPU rendering like VRay.

Aside from some beta bugs, it has always been super stable. Devs are responsive and always pushing things forward with new features.

I think they just changed their licensing model, but existing licenses get 3 render nodes, which is nice if you have some old machines laying around.

All really good to hear. I had a poke around in the UI and found the Chaos asset browser and was impressed with how seamless the process of downloading an HDRI and adding it to your scene is. Loads of high quality assets available that are added to your scene as a proxy to keep viewport overheads down, and these are all included with your license? Impressive if so.

Tried converting some of my current projects to Corona from Redshift and they just look better, even working with the exact same PBR textures (Quixel) and HDRIs. Corona's lighting is instantly more pleasing and natural, but even for mograph/graphic scenes I find its lighting and texturing more pleasing. Redshift has that CGI/gamey look that even with an ACES workflow and post processing is hard to avoid and workaround. Biased vs Unbiased I suppose.

Redshift has M1 compatibility, speed and stability on its side but since the Maxon buyout new features have stalled and its integration with C4D hasn't improved. With the new subscription model I think I'd rather be giving my money to Corona - those additional render nodes will definitely be useful.

@LymeChips same old story with OTOY, then. Octane will be stable and fully featured, one day. We're still waiting. I love how their engine looks but just cannot work with such disruptively unstable software.

 

[LymeChips]

... same old story with OTOY, then. Octane will be stable and fully featured, one day. We're still waiting. I love how their engine looks but just cannot work with such disruptively unstable software.

To be fair it's been rock solid on PC the past two years I've been using it. I think they were thrown for a loop with Apple Silicone and having slowed updates until 2022 tells me they're finding footing in how they want to approach Octane X.

From this thread, hearing about Corona here I'm very interested in trying that out. But Otoy has their hooks in me ever since they released their RNDR platform. I've never used such a fast and well priced render "farm" - it's really changed the way I've been able to do commercial work.

 

[jujoje]

check out Shapr3d. It has a (heavy) mechanical / product design slant (and it’s missing quite a bit of complexity of a traditional 3d modeling desktop app), so it might not be what you need…

Shapr3d looks really cool; tried it a long time ago and remember thinking that it did a really innovative job of reinventing a 3d interface for the iPad and Apple Pencil (particularly after using NURBS in Studio Tools / Maya). Assumed that more apps would start to experiment with interaction models on the iPad for 3d software but so far not really. $240 a year is a bit much (although if you were doing arch vis / product design for monies probably not too bad).

@LymeChips same old story with OTOY, then. Octane will be stable and fully featured, one day. We're still waiting. I love how their engine looks but just cannot work with such disruptively unstable software.

Octane is an odd one; the renderer seems pretty great, but their inability to deliver timely updates and the afore mentioned instability means that it's hard to take seriously as a production renderer.

I prefer the look of Corona to Redshift already. The way it handles light and textures is far more convincing and aesthetic to my eye. It has a filmic, natural look out of the box and needs far less tweaking and pushing than Redshift.

Redshifts always had a look about it; the quality of the lighting always feels slightly off. A lot of the time it's fine (and sometimes preferable), but for more realistic things it tends to fall down a bit - I guess due to some of the trade-offs they make speed and being biased (one of the other reasons Octane should have been the popular gpu rendering to my mind).

 

[ed.]

$240 a year is a bit much (although if you were doing arch vis / product design for monies probably not too bad).

Agreed, it's too pricey to have it just sitting around, but the monthly subscription is ideal for when a job pops up, especially if you're already familiar with the interface and workflow.

 

[Xiao_Xi]

It seems that Apple has almost finished the Blender's Metal backend. Indeed, the new version of Blender will support AMD GPUs in macOS 12.3.

Cycles Apple Metal device feedback

The latest Blender 3.1 and 3.2 builds now have initial support for AMD graphics cards. However this requires macOS 12.3, which is still in beta. Obviously there is risk involved in installing a beta operating system, so I wouldn’t necessarily recommend everyone to go and do that. But if there’s...

devtalk.blender.org

However, it lacks the support of MetalRT and optimizations.

Cycles Metal device

**Tasks** For tracking Metal device development, targeting 3.1 release: - [x] Kernel changes required to compile for MSL (Metal Shading Language): - [x] Explicit address space qualifiers - [x] Adapt cross-platform layer (device/gpu) for MSL - [x] MetalRT kernel changes - [x] Host...

developer.blender.org

How fast could Blender get with the support of MetalRT and optimizations?

 

[Xiao_Xi]

Can someone explain why MetalRT uses less memory for curves?

Cycles Apple Metal device feedback

This scene is particularly heavy on curves. It is worth enabling the “MetalRT (Experimental)” checkbox if you haven’t already done so - MetalRT uses less memory for curves.

devtalk.blender.org

 

[jujoje]

Been a bit quiet, but looks like Houdini is finally heading towards Apple Silicon:

H19 Apple Silicon Tech Preview | SideFX

SideFX is pleased to announce the Houdini 19 Apple Silicon build which is a technical preview providing early access for macOS users on Apple M1 hardware. The current version of Houdini 19 works on M1 via Apple’s Rosetta translator. The Silicon build will provide a more optimized native build.

www.sidefx.com

And performance sounds good:

"So far we're finding improvements in several workflows (i.e. flip fluids, crowds, etc.) in the native macOS arm64 build compared to the main x86_64 build. The performance improvements have varied but we've seen up to 30-40% faster in some cases. And overall, the arm64 build just feels zippier."

And for good measure, gets a mention in the spec pages for the Mac Studio:

 

[treehuggerpro]

Anyone got a theory on why the M1 Ultra's relative performance isn't where you'd expect in Redshift? ~4x

 

[jujoje]

Anyone got a theory on why the M1 Ultra's relative performance isn't where you'd expect in Redshift? ~4x

That did strike me as a bit odd; gpu scaling should definitely be better than that.

Also I wonder what amount of memory is now available to the gpu. I think it was 1/4 ish being reserved for the system on the mbp.

 

[ader42]

Yes, that is odd, going from 2.2x to 2.6x is a smallish step (18%), not a leap.

Those graphs seem to show that the software itself is a major limiting factor - e.g. how many threads it can make use of for example? One would expect more and smaller buckets to give more improvement than that when going from 32 cores to 64 cores.

As someone who has an M1 Max on order and is considering cancelling it for an M1 Ultra it makes me wonder if I should just stick with M1 Max.

 

[Boil]

That did strike me as a bit odd; gpu scaling should definitely be better than that. I wonder what amount of memory is now available to the gpu. I think it was 1/4 ish being reserved for the system on the mbp.

Redshift folks themselves reported that the 64GB M1 Max could use slightly more than 40GB RAM for the GPU...?

 

[anticipate]

That did strike me as a bit odd; gpu scaling should definitely be better than that.

Also I wonder what amount of memory is now available to the gpu. I think it was 1/4 ish being reserved for the system on the mbp.

I suspect they are running into limitations with the interposer between the two chips. It’s not SLI or anything but that discrepancy is huge.

 

[jujoje]

Those graphs seem to show that the software itself is a major limiting factor - e.g. how many threads it can make use of for example? One would expect more and smaller buckets to give more improvement than that when going from 32 cores to 64 cores.

More and smaller buckets is not always better; it's very dependant on what you're rendering, although one would assume that the scene Apple and Redshift were using would be tweaked towards the strengths of the M1 GPUs.

Redshift folks themselves reported that the 64GB M1 Max could use slightly more than 40GB RAM for the GPU...?

I wonder whether its a proportion of memory or a reserved 25GB chunk; 100GB of gpu memory would be pretty ridiculous.

I suspect they are running into limitations with the interposer between the two chips. It’s not SLI or anything but that discrepancy is huge.

This does sound likely, although given that it apparently has a 'a massive 2.5TB/s of low latency, inter-processor bandwidth', to quote the PR, I wouldn't have thought the hit would be that high.

 

[UBS28]

Redshift folks themselves reported that the 64GB M1 Max could use slightly more than 40GB RAM for the GPU...?

Seems in line what I see with the 32 GB M1 Max, which seems to pull 20 GB of RAM for the GPU.

 

[singhs.apps]

100GB of gpu memory would be pretty ridiculous.

Not.

 

[Boil]

Wait until Apple gets LPDDR5X RAM into the ASi Mac Pro...

M1 Ultra could support 512GB of LPDDR5X RAM...!

But at what cost...?

 

[jujoje]

Not

Don't get me wrong, I'm all for it; you could load the entire Moana data set onto the GPU with that amount of memory. Other than batching things, I'm struggling to find a task where that amount of GPU memory would be necessary (not that it never would be).

But at what cost...?

Too much

I was thinking of waiting for the Mac Pro, but the Studio is kinda in the sweet spot for pricing for a workstation; the Mac Pro feels like it's going to be too expensive (but curious as to what crazy hardware and design Apple will come up with for it).

 

[jujoje]

My main misgiving on the Mac Studio is over the gpu speed; it looks to land, at best between a 3060 and 3090, depending on optimisation, leaning more towards the former. Which seems a bit weak for a pro workstation designed to last several years.

 

[sirio76]

Anyone got a theory on why the M1 Ultra's relative performance isn't where you'd expect in Redshift? ~4x


View attachment 1969972

The answer is quite simple, Redshift do not scale linearly using more GPU, it’s a known limitation no matter if you use AS or Nvidia GPUs.

 

[treehuggerpro]

The answer is quite simple, Redshift do not scale linearly using more GPU, it’s a known limitation no matter if you use AS or Nvidia GPUs.

Redshift doesn't scale 100% perfectly in all scenarios, but for the most part it does, or comes very close. Maxon collates user results in the Redshift forum for their Vultures Benchmark tool. Relative examples as per the colours:

 

[sirio76]

I’m talking about real scene not simple benchmark, Apple tests are made on a large scene, over 1GB.
With only two GPU you won’t notice that much, the more cards the worst it will be it was explained by a RS developer, I’ll see if I find the topic.
Also you should make the comparison on the same machine using the same GPUs, those benchmarks come from different systems, different CPU, different OS, different drivers, maybe OC cards etc.

 

[diamond.g]

More and smaller buckets is not always better; it's very dependant on what you're rendering, although one would assume that the scene Apple and Redshift were using would be tweaked towards the strengths of the M1 GPUs.



I wonder whether its a proportion of memory or a reserved 25GB chunk; 100GB of gpu memory would be pretty ridiculous.



This does sound likely, although given that it apparently has a 'a massive 2.5TB/s of low latency, inter-processor bandwidth', to quote the PR, I wouldn't have thought the hit would be that high.

Did they mention what kind of latency hit the interposer adds?