3D Rendering on Apple Silicon, CPU&GPU

[singhs.apps]

Help

help.autodesk.com

Apple native CPU support: Arnold natively supports running under Apple M series CPUs, such as the M1 CPUs. This allows Arnold to natively run under DCCs that support Apple native CPUs and in our tests gives speedups of up to 1.45x compared to running with x86_64 Rosetta 2 translation. (ARNOLD-10481)

 

[deconstruct60]

Hmm... Brad Peebler, former CEO of Luxology, is now with Apple; maybe Apple buys Modo...?

Apple is on a wide spread hiring freeze. If they don't have enough money to pay salaries for new people how do they have money to buy Modo (and more people to pay)?

Apple doesn't sell sky high price software packages either. So if they chopped Modo from $1800 to $500 it isn't like revenues are going to jump sky high to pay for extra programmer seats either. (they might get some volume bump with a lower cost but more than 3x increase? maybe , maybe not. ) Highly likely part of Modo's limited developer resources allocation problem now is that paying off previous acquisition costs overhead.

Apple probably won't go to zero acquisitions in 2023 , but probably going to be much smaller and far more focused on new growth (greenfield relative to Apple) gaps in new R&Dm, than in Sherlocking the existing Mac application space. Kicking your software partners when they are down (overall market has problems ) doesn't buy them more partners in the long term.

 

[sirio76]

Apple is on a wide spread hiring freeze. If they don't have enough money to pay salaries for new people how do they have money to buy Modo (and more people to pay)?

You can’t be serious.. So suddenly the largest/richest company in the world became poor?

-Apple is valued about 2.5 Trillion $ and have a cash reserve of 50 Billions $
-it will spend little less than 200 millions $ in acquisition this year
-on average they earn about one Billion $ a day, every day
-Cook clearly said that they are still hiring, just with more caution

I‘ve no data about Modo value but my guess is that is pretty cheap and that Apple wouldn’t have any financial issues to buy it if they really want to.
Of course I’m sure this is not going to happen, but not because of lack of financial resources.

 

[jujoje]

Hmm... Brad Peebler, former CEO of Luxology, is now with Apple; maybe Apple buys Modo...?

There's little point in buying Apple Modo, but could totally see him working on an app for AR/VR creation (if they ever releases their headset), that being an area where they need solid authoring tools and Brads experience would be put to good use. Reality Composer is far to limited and high level. They also need to improve their usd tool and docs, but that's another story...

Help

help.autodesk.com

Apple native CPU support: Arnold natively supports running under Apple M series CPUs, such as the M1 CPUs. This allows Arnold to natively run under DCCs that support Apple native CPUs and in our tests gives speedups of up to 1.45x compared to running with x86_64 Rosetta 2 translation. (ARNOLD-10481)

Nice catch! Surprised to see anything Autodesk getting AS support at this stage; hopefully they'll surprise us with Maya in the not too distant! That's also a pretty significant uplift.

 

[vel0city]

Help

help.autodesk.com

Apple native CPU support: Arnold natively supports running under Apple M series CPUs, such as the M1 CPUs. This allows Arnold to natively run under DCCs that support Apple native CPUs and in our tests gives speedups of up to 1.45x compared to running with x86_64 Rosetta 2 translation. (ARNOLD-10481)

I nearly missed this important post thanks to all the off-topic PC posts. Can a mod clear up the last few pages please? Anything to do with Windows or Nvidia cards is completely off-topic in this thread. You might as well be posting about lawnmowers in a thread about toasters. Zero relevance.

 

[innerproduct]

Help

help.autodesk.com

Apple native CPU support: Arnold natively supports running under Apple M series CPUs, such as the M1 CPUs. This allows Arnold to natively run under DCCs that support Apple native CPUs and in our tests gives speedups of up to 1.45x compared to running with x86_64 Rosetta 2 translation. (ARNOLD-10481)

This is very good news. Auto desk have been silent for a long time so many have been worried. But how do one use this in Maya? Doesn’t that have to be native as well? Anyway. 40-60 % improved when going native seems to be vey common. Same goes for other cpu renders and for Houdini Sims. (And the ultra is very fast for sims in Houdini! a good bit better than a 5950x, almost 7950x level)

 

[singhs.apps]

This is very good news. Auto desk have been silent for a long time so many have been worried. But how do one use this in Maya? Doesn’t that have to be native as well? Anyway. 40-60 % improved when going native seems to be vey common. Same goes for other cpu renders and for Houdini Sims. (And the ultra is very fast for sims in Houdini! a good bit better than a 5950x, almost 7950x level)

“This allows Arnold to natively run under DCCs that support Apple native CPUs” , suggests Arnold will run natively on AS native Apps.

It may work in Houdini AS Beta, most likely right away in C4D.

Arnold ships as default with all Maya licenses and one would imagine it has more Maya based seats on macOS than any other DCC suggesting we will see Maya native on AS at some point.
I don’t imagine Maya sales are so poor in macOS that they wouldn’t bother.

Porting over to AS may also help windows on ARM ports if Microsoft ever decides to offer it as alternative to x86 version.

 

[Xiao_Xi]

Apple is investigating how to achieve better XPU performance.

Blender Archive - developer.blender.org

developer.blender.org

Would other renderers have a similar performance increase?

 

[sirio76]

Even if they do, you will never notice that small speedup in real life

 

[Pressure]

Even if they do, you will never notice that small speedup in real life

Up to 20% give or take seems great on the M2.

 

[Xiao_Xi]

Is the investigation on improving XPU performance a step on the road to enabling multi-GPU rendering?

This patch stems from investigations into multi-device performance on Metal.

Blender Archive - developer.blender.org

developer.blender.org

 

[deconstruct60]

Apple is investigating how to achieve better XPU performance.

View attachment 2118608

Blender Archive - developer.blender.org

developer.blender.org

Would other renderers have a similar performance increase?

Is there really much of real work increase on something other than a M2/M1 ? The root cause motivation here is that the workload job scheduler is deficient. That problem is only going to get substantially worse as both scale up the GPU core number that you are ineffectively scheduling and also skew the GPU core count to be 2x - 4x larger than the CPU core count. The bulk of the 'solution' here is to use CPU cores to plug the gaps in workload allocation. When there are more GPU cores there will be more gaps. When the CPU:GPU ratio then pragmatically will have less "gap fillers" .

Part of what is helping here is that the CPU cores are about the same number as the GPU cores ( while not matcing in math units in each type of core , but a measure that there are not 'too many' GPU mis-scheduled gaps to fill. ).

The problem with this "work harder , not smarter" is that brute force only scales so far. If wasting compute resources that will only grow as scale up.

If had a more accurate scheduler then the CPU cores contributions would pragmatically be more additive. Here really just filling in workload pragmatically lost by wasting GPU resources.


P.S. This 'apply brute force' approach is generally generally not what XPUs are about. The bigger synergies are when trying to offload and/or segreate certain work onto which processoing unit is better tuned for that. For example, moving TCP/IP network station overhead off the CPU onto some specific accelearator. Of extremely embarrasing parallel compute off the CPU and on GPU cores. It is generally not about shifting work because leaving 'core type x' cores idle too often and/or starving them of data.

 

[deconstruct60]

Is the investigation on improving XPU performance a step on the road to enabling multi-GPU rendering?

Blender Archive - developer.blender.org

developer.blender.org

Going to an "cheaper and even dumber" scheduler isn't particularly likely to deal with NUMA effects very well since not trying to localize or mitigate workload placement at all.

An upside tradeoff though would be able to reallocate scheduler 'time' to bucket assigning time. Would have to save the dumber scheduler from itself by having different 'buckets' assigned to the more distance CPU/GPU/etc core groupings.

 

[Xiao_Xi]

It looks like NanoVDB will soon work with Metal.

NanoVDB support for Metal is also being worked on.

2022-11-30 Render & Cycles Meeting

Attendees Brecht Van Lommel (Blender) Thomas Dinges (Blender) Weizhen Huang (Blender) Christophe Hery (Meta) Sebastian Herholz (Intel) Nikita Sirgienko (Intel) Michael Jones (Apple) Brian Savery (AMD) Notes Many Lights: getting close to ready to merge. Weizhen is still cleaning up the...

devtalk.blender.org

What is the use case for NanoVDB? What is the difference between OpenVDB and NanoVDB?

Could other 3D software benefit from this?

 

[singhs.apps]

It looks like NanoVDB will soon work with Metal.

2022-11-30 Render & Cycles Meeting

Attendees Brecht Van Lommel (Blender) Thomas Dinges (Blender) Weizhen Huang (Blender) Christophe Hery (Meta) Sebastian Herholz (Intel) Nikita Sirgienko (Intel) Michael Jones (Apple) Brian Savery (AMD) Notes Many Lights: getting close to ready to merge. Weizhen is still cleaning up the...

devtalk.blender.org

What is the use case for NanoVDB? What is the difference between OpenVDB and NanoVDB?

Could other 3D software benefit from this?

It’s an Nvidia initiative to accelerate openvdb (an open source volumetric data exchange format - think smoke, clouds, etc…) on the GPU.

NanoVDB

NanoVDB adds real-time rendering GPU support for OpenVDB

developer.nvidia.com

 

[hifimac]

https://twitter.com/i/web/status/1598420299154096128

It will be really interesting to see the Apple Neural Engine get expanded into areas like this.

 

[jmho]

Very interesting that the base M2 Air is significantly faster than the M1 Pro. I just looked and Apple do claim that the M2's Neural Engine is 40% faster, so it makes sense.

The Neural Engine has been incredibly under-utilised so it's nice to see that Apple themselves are working on libraries that can be used by 3rd parties.

 

[bcortens]

https://twitter.com/i/web/status/1598420299154096128

It will be really interesting to see the Apple Neural Engine get expanded into areas like this.

It would be nice if the Neural engine gained a bit more granular API access at some point to make it easier to use.

 

[sunny5]

Apple Silicon still sucks for 3D and game. Blender is the only hope so far for AS Mac but for gaming, I highly doubt it.

 

[jujoje]

Apple Silicon still sucks for 3D and game. Blender is the only hope so far for AS Mac but for gaming, I highly doubt it.

Houdini's running pretty great on Apple Silicon, as is RE Village, Hades and Disco Elysium, so you're pretty much wrong on both counts.

 

[sirio76]

Not only Houdini, C4D works great too on AS, surprisingly (at least for my workflow) is the most stable platform I’ve ever worked on (comparing to Intel Mac or Windows Intel/AMD machines).
Anybody reading this topic should be noticed that most users here are gamer/3Dhobbyists who pretend to be professional that are actually using 3D software to generate income, so anyone reading should take everything with a large grain of salt.

 

[ader42]

Apple Silicon still sucks for 3D and game. Blender is the only hope so far for AS Mac but for gaming, I highly doubt it.

Apple Silicon is brilliant for ZBrush

 

[Xiao_Xi]

Apple Silicon still sucks for 3D and game. Blender is the only hope so far for AS Mac

Unlike FOSS projects, commercial software benefits from behind-the-scenes development to surprise users. Therefore, some of them might even be more suited to Apple Silicon than Blender. As Apple developers who help Blender have shown, adapting software to Apple Silicon takes a long time.

 

[sunny5]

Houdini's running pretty great on Apple Silicon, as is RE Village, Hades and Disco Elysium, so you're pretty much wrong on both counts.

Haha, you are the one being false and wrong. The truth is, Apple Silicon Mac is meant for 2D, not 3D. Pretty great does NOT mean it's comparable to PC. Take a look at RE Village. M1 Max is TWICE slower than mobile RTX 3070TI. TWICE. This is more like equal to mobile RTX 3060. RE Village is a NATIVE game and yet it's not even close to mobile RTX 3070TI. If the power consumption is your concern, Apple Silicon Mac is limited by itself while PC can control the limit. And yet Apple advertised that M1 Max is close to mobile RTX 3070TI or 3080. Truth hurts, huh?

Also, almost all 3D related software are CUDA based and therefore, Mac is not even close. If you gonna argue with benchmarks, they dont really represent actual and real life performance.

Apple Silicon is brilliant for ZBrush

Brilliant does NOT mean it's comparable. So far, AS Mac is good only in limited uses. Adapting takes a lot of time and yet Blender is the only 3D software willingly supporting Apple Silicon. Mac is never known for 3D works for a long time especially after they ditch Nvidia GPU which is dominating 3D software. Even now, 3D is essential to use 3D related software.

It seems many people are confusing between comparable and works great. For example, Blender works great with M1 Max but the performance is comparable to GTX 1080 based on my testing. If you are working in 3D, you really need high performance to do rendering as they take a lot of time to finish the task. Even with RTX 4090, it takes a few hours to render. Yes, Mac still can render but the rendering time will be... devastating. Because most of 3D software are optimized for Nvidia, especially CUDA.

If you still wanna argue, bring me some actual test results, not benchmarks. Every benchmarks work differently and they dont really represent the actual performance while you are working. I've seen many 3D artist and they always hate Mac for 3D works as they are TOO slow to work with.

No idea what this is even supposed to mean.

Video, Music, Photo, Design are 2D based.

Well, duh. RTX 3070TI is a much faster GPU. It also uses 3-4 times more power depending on the model. Why is it surprising that a faster GPU is faster?

I said MOBILE RTX 3070TI. And it used twice power power, not 3~4 times duh. Also, M1 Max is limited up to 60W. So who blame who? Ironic.

So what? Native does not mean perfectly optimised. It wasn't developed for Macs from scratch and I doubt that they use every trick in the book to make it as good as it could on Apple hardware.

Now you are complaining that it's not OPTIMIZED yet lol.

 

[leman]

The truth is, Apple Silicon Mac is meant for 2D, not 3D.

No idea what this is even supposed to mean.

Pretty great does NOT mean it's comparable to PC. Take a look at RE Village. M1 Max is TWICE slower than mobile RTX 3070TI. TWICE.

Well, duh. RTX 3070TI is a much faster GPU. It also uses 3-4 times more power depending on the model. Why is it surprising that a faster GPU is faster?

RE Village is a NATIVE game

So what? Native does not mean perfectly optimised. It wasn't developed for Macs from scratch and I doubt that they use every trick in the book to make it as good as it could on Apple hardware.