3D Rendering on Apple Silicon, CPU&GPU

[jmho]

I have the creative cloud version. I'm not sure what you mean by using the GPU? Do you mean for rendering, because for IRAY rendering, no, it's still using CPU, but the main app itself is definitely using the GPU and is nice and responsive.

Baking still seems to be using CPU though.

 

[BootLoxes]

I have the creative cloud version. I'm not sure what you mean by using the GPU? Do you mean for rendering, because for IRAY rendering, no, it's still using CPU, but the main app itself is definitely using the GPU and is nice and responsive.

Baking still seems to be using CPU though.

The main app part is whats most important to me. The 2023 edition should release this month on Steam so I will jump on it.

 

[vel0city]

Octane Render C4D plugin now running natively on AS, I'm running it in C4D R23 right now and finding it remarkably stable and responsive on a 16" Macbook Pro Max. Instantly prefer the look of it to Redshift which has been my go-to renderer for years. The way Octane handles lighting and textures is so much more aesthetically beautiful and accurate to me.

 

[LymeChips]

Has anyone done the math and have a guess as to what the Octane Bench score of the unreleased Mac Pro is? Otoy CEO earlier this year said their M1 Max laptop was about 220 OB so maybe we’ll be looking at ~1k OB?

220 M1 Max
+25% M1>M2
x4 1 SoC > 4 SoC
- assuming the 220OB score is correct do my other estimates add up?

 

[innerproduct]

I guess we were a few that expected that kind of scaling but the m1 ultra showed extremly bad scaling for renderers. About 50% instead of 90-100% which is the scaling we see even over pcie from other manufacturers. And apple bragged so much about their ultra fusion one might have expected more than 100% scaling due to benefits of increased bandwidth. I really hope you are right and Apple have solves these issues for the next generation. Well see. The issue is that a basic PC could easily pack two 4090 and reach about 3000 OB for about the same price as a mac studio ultra. I actually almost expect something unexpected for the macpro that makes it stand out. Price will not be reasonable in any way though as they sadly showed all to well with the mac studio.

 

[Boil]

I guess we were a few that expected that kind of scaling but the m1 ultra showed extremly bad scaling for renderers. About 50% instead of 90-100% which is the scaling we see even over pcie from other manufacturers. And apple bragged so much about their ultra fusion one might have expected more than 100% scaling due to benefits of increased bandwidth.

So 2+2=5...?

I really hope you are right and Apple have solves these issues for the next generation. Well see. The issue is that a basic PC could easily pack two 4090 and reach about 3000 OB for about the same price as a mac studio ultra.

Not really a "basic PC" if it has dual 4090 GPUs...?

Those 4090s are $1600 each, so the rest of your "basic PC" (chassis, mobo, CPU, PSU, RAM, SSD, cooling) has a $400 budget; the CPU alone will eat that up...

I actually almost expect something unexpected for the macpro that makes it stand out. Price will not be reasonable in any way though as they sadly showed all to well with the mac studio.

Pricing for the base M2 Ultra Mac Pro should stay at the current $6K that the base 2019 Intel Mac Pro is at, base M2 Ultra Mac Studio is $4K; upgrading the PSU, mobo (with PCIe slots), and chassis from Mac Studio to Mac Pro should not cost more than the current $2K pricing differential...?

 

[innerproduct]

Maybe I choose bad wording. a great but kind of basic pc with a 300$ cpu on a 300$ mb in a 100$ chassis can always have two GPUs. Oc dual 4090 doesn’t make it basic anymore. But still at a very reasoble price. A macpro will be priced and specced like a high end workstation like a threadripper pro using rtx A series card most likely. But a real pro machine have massive ram capacity, ecc ram, actual pro features in graphics cards etc while Apple sells glorified mobile parts across the line. I do hope I am wrong though and that we all will be blow away. (Oh and regarding the 2+2=5 comment, actually caches and mem bandwith can make immense differences for some workloads and with apple marketing it sure sounded like the ultra was more that just glued together)

 

[vinegarshots]

Maybe I choose bad wording. a great but kind of basic pc with a 300$ cpu on a 300$ mb in a 100$ chassis can always have two GPUs. Oc dual 4090 doesn’t make it basic anymore. But still at a very reasoble price. A macpro will be priced and specced like a high end workstation like a threadripper pro using rtx A series card most likely. But a real pro machine have massive ram capacity, ecc ram, actual pro features in graphics cards etc while Apple sells glorified mobile parts across the line. I do hope I am wrong though and that we all will be blow away. (Oh and regarding the 2+2=5 comment, actually caches and mem bandwith can make immense differences for some workloads and with apple marketing it sure sounded like the ultra was more that just glued together)

To be fair, it is extremely difficult to fit 2x4090s into one PC. You need an 8-slot motherboard, which is going to cost you $800-$1200 all by itself. Plus, one 4090 running at full 600W requires a 1200w PSU, so you're not going to be able to power two of them on a standard PC PSU unless you run it out of spec...

 

[mi7chy]

Those 4090s are $1600 each, so the rest of your "basic PC" (chassis, mobo, CPU, PSU, RAM, SSD, cooling) has a $400 budget; the CPU alone will eat that up...

Complete system with two 4090 FEs is well under budget of $5K Mac Studio M1 Ultra while being about 17x faster on Blender.

To be fair, it is extremely difficult to fit 2x4090s into one PC. You need an 8-slot motherboard, which is going to cost you $800-$1200 all by itself. Plus, one 4090 running at full 600W requires a 1200w PSU, so you're not going to be able to power two of them on a standard PC PSU unless you run it out of spec...

No it isn't. 4090 FE is a 3-slot card while the motherboard mentioned above has 5-slot spacing until next PCIe slot at 6th slot. You just need to make sure your case has space for two additional slots below last PCIe slot which usually isn't an issue. Puget Systems measured total system power draw with two 4090 FEs under load at 890W from wall.

https://www.pugetsystems.com/labs/articles/1-7x-nvidia-geforce-rtx-4090-gpu-scaling/

 

[iPadified]

Complete system with two 4090 FEs is well under budget of $5K Mac Studio M1 Ultra while being about 17x faster on Blender.

No it isn't. 4090 FE is a 3-slot card while the motherboard mentioned above has 5-slot spacing until next PCIe slot at 6th slot. You just need to make sure your case has space for two additional slots below last PCIe slot which usually isn't an issue. Puget Systems measured total system power draw with two 4090 FEs under load at 890W from wall.

https://www.pugetsystems.com/labs/articles/1-7x-nvidia-geforce-rtx-4090-gpu-scaling/

What has this to do with the title of the thread? Please stick to the topic.

 

[jujoje]

No it isn't. 4090 FE is a 3-slot card while the motherboard mentioned above has 5-slot spacing until next PCIe slot at 6th slot. You just need to make sure your case has space for two additional slots below last PCIe slot which usually isn't an issue. Puget Systems measured total system power draw with two 4090 FEs under load at 890W from wall.

890W; lol. And let's hope the cables don't melt.

Pricing for the base M2 Ultra Mac Pro should stay at the current $6K that the base 2019 Intel Mac Pro is at, base M2 Ultra Mac Studio is $4K; upgrading the PSU, mobo (with PCIe slots), and chassis from Mac Studio to Mac Pro should not cost more than the current $2K pricing differential...?

Wouldn't that be a bit underwhelming? I'm sure the case would be very nice and all...

Feels like the market segmentation is a bit off between the high end studio and Mac Pro, which makes me thing the Mac Pro will be very expensive (well even more so than I was anticipating).

There's also the question of what you can actually do with those PCIe Slots...

 

[tomO2013]

Raw speed in blender is one thing, no doubt it’s a good reason to buy a 4090 PC if that’s your primary and sole driving force.

Alternatively, if you are working with massive textures, large detailed scenes that require more than 24gb memory, then you’re out of look with the 4090.

You’d be much better off working with a rtx a6000 quadro and hoping that your scene fits in its 48GB memory footprint. That card alone retails for $4650 USD.

Alternatively if your scene can’t fit in the 48GB memory of the a6000…. Apple will happily sell you a cheaper m1 ultra with 128GB of unified memory that can service your needs.

Personally if blender performance was a big driver and I was serious about 3D rendering work, driver support for my software, working with Adobe substance painter 3D with huge memory footprint requirements etc… in effect a production workflow that required more than the 24gb consumer limit… I’d be looking at a professional graphics card or m1 ultra.

I’d probably be looking first at comparing the a6000 nvidia, not the 4090 (even if it’s faster in raw blender).

 

[Pressure]

Maybe I choose bad wording. a great but kind of basic pc with a 300$ cpu on a 300$ mb in a 100$ chassis can always have two GPUs. Oc dual 4090 doesn’t make it basic anymore. But still at a very reasoble price. A macpro will be priced and specced like a high end workstation like a threadripper pro using rtx A series card most likely. But a real pro machine have massive ram capacity, ecc ram, actual pro features in graphics cards etc while Apple sells glorified mobile parts across the line. I do hope I am wrong though and that we all will be blow away. (Oh and regarding the 2+2=5 comment, actually caches and mem bandwith can make immense differences for some workloads and with apple marketing it sure sounded like the ultra was more that just glued together)

The problem will probably be finding an affordable motherboard that has two PCIe Gen 4 x16 slots that both are wired to x16 and not one at 16x/8x and another at 4x. Which again leads to the problem of having a fast enough processor to feed the GPUs data with enough PCIe lanes available.

In the end you will need either a Xeon or Threadripper to accomplish both.

 

[Xiao_Xi]

This presentation from the latest Blender conference explains how Blender Cycles can render massive scenes using multiple Nvidia GPUs.

Could Apple distribute the rendering across multiple Mac Studio using Thunderbolt as it does with Tensorflow distributed training?

 

[ader42]

This presentation from the latest Blender conference explains how Blender Cycles can render massive scenes using multiple Nvidia GPUs.

Could Apple distribute the rendering across multiple Mac Studio using Thunderbolt as it does with Tensorflow distributed training?

If Mac is your only option (I would rather retire than use Windows) and are on a budget then CrowdRender for Blender works on Mac… so if you need to save cost and can live with Blender then why not buy the most cost-effective Macs (Mac Minis?) as a local render farm?

For Cinema4D I believe Arnold, Octane and Redshift all have the capabilities for distibuted rendering of a single image.

 

[LymeChips]

Yeah I wasn’t trying to start a Mac/PC performance per dollar comparison. I think everyone knows how that goes….


But thanks for all the info. I wasn’t aware scaling was that poor, I wonder if Octane has a better scaling model vs RS?

I’m an Octane user so in the end most of the rendering will be handled by a PC either via their net render feature for IPR or Deadline. I just hope the AsMP can deliver ~800 OB so it can stand alone without the PC net rendering.

 

[vinegarshots]

No it isn't. 4090 FE is a 3-slot card while the motherboard mentioned above has 5-slot spacing until next PCIe slot at 6th slot. You just need to make sure your case has space for two additional slots below last PCIe slot which usually isn't an issue. Puget Systems measured total system power draw with two 4090 FEs under load at 890W from wall.

Good luck getting a 4090 FE, though. I have a PNY 4090, and it's a 4-slot card. In many cases (such as mine), the PSU is directly under the bottom of the mobo, which means the bottom slots are not usable for a multislot GPU.

4090s aside from the FE can draw up to 600W each, but the big issue is that the ATX 3.0 spec requires the PSU to handle double the wattage for transient power spikes.

 

[Boil]

Wouldn't that be a bit underwhelming? I'm sure the case would be very nice and all...

Feels like the market segmentation is a bit off between the high end studio and Mac Pro, which makes me thing the Mac Pro will be very expensive (well even more so than I was anticipating).

There's also the question of what you can actually do with those PCIe Slots...

Four standard uses off the top of my head...

• 4K/8K video I/O cards
• Audio DSP I/O cards
• High-speed networking cards
• M.2 NVMe SSD RAID cards

Why do some folks insist on thinking that PCIe is good for nothing but discrete GPUs...?!?

 

[mi7chy]

In many cases (such as mine), the PSU is directly under the bottom of the mobo, which means the bottom slots are not usable for a multislot GPU.

That's a user problem with case selection. Not all cases have bottom PSU. Two GPU configuration is easy with >2 common in the mining world which can be adapted to mine 3D renders.

4090 build is actually easier since you can half the number of GPUs to get performance of build with twice the number of A5000 or 3090 GPUs.

https://render.otoy.com/octanebench/results.php

 

[sirio76]

Yeah I wasn’t trying to start a Mac/PC performance per dollar comparison. I think everyone knows how that goes….


But thanks for all the info. I wasn’t aware scaling was that poor, I wonder if Octane has a better scaling model vs RS?

I’m an Octane user so in the end most of the rendering will be handled by a PC either via their net render feature for IPR or Deadline. I just hope the AsMP can deliver ~800 OB so it can stand alone without the PC net rendering.

Scaling is not poor. Misinformation is so diffuse here that is indistinguishable from reality. Some software do not scale that well, other will scale perfectly. Of course troll will just report of the software that not scale well ignoring anything that works perfectly.
Also, the majority of the people writing here in this topic are hobbiest, as said before nothing wrong with that but you may find that professional users have very different opinion.

 

[vinegarshots]

That's a user problem with case selection. Not all cases have bottom PSU. Two GPU configuration is easy with >2 common in the mining world which can be adapted to mine 3D renders.

Yeah, you're still not showing how a 2X 4090 build is "easy," because it's not. And you still have that pesky PSU max wattage requirement to solve, too. You do realize that Puget mining rig has multiple power supplies attached to it, right?

 

[innerproduct]

the big problem here is that 3d rendering on the mac currently we only have 3 gpu renderers (octane, redshift and blender) All of these show worse than expected performance scaling for the ultra. This has severely dampend enthusiasm. The ultra does however scale very nicely in some sim software like axiom for houdini where it touches 3090 perf. But as we say here, for rendering there is no good solution and a nvidia pc even in the lowest tier performs better.

 

[Xiao_Xi]

3 gpu renderers (octane, redshift and blender) [...] show worse than expected performance scaling for the ultra.

The ultra does however scale very nicely in some sim software like axiom for houdini.

Do you have a link that shows this?

 

[tomO2013]

the big problem here is that 3d rendering on the mac currently we only have 3 gpu renderers (octane, redshift and blender) All of these show worse than expected performance scaling for the ultra. This has severely dampend enthusiasm. The ultra does however scale very nicely in some sim software like axiom for houdini where it touches 3090 perf. But as we say here, for rendering there is no good solution and a nvidia pc even in the lowest tier performs better.

Would you be comfortable sharing samples of your own custom 3d modelling work (not benchmark source). I’m asking because benchmarks can test extreme situations that may not actually reflect how one would use the software.

Given the discussion course on 3090, blender etc… versus advocating for Maya or other vendor backed software, I’m assuming (possibly incorrectly) that you work for a small independent studio or are an individual hobbyist.
Giving us an indication of your ‘As Is’ workflow with actual real world outputs that you personally have created will possibly allow somebody here to guide you on whether an alternative workflow might fit your need, given your interest in the M1 ultra and price/performance for your professional workflow?

Would you be comfortable sharing some of your personally created 3d modelling work - I can appreciate that if you work for a small shop that you are not allowed to share your work externally for copyright / client NDA reasons. However most folks working for such shops have a personal portfolio - it’s quite common in the industry ?

Obviously if you are hobbyist there should be nothing really holding you back from sharing your own creations.

Same ask goes to @mi7chy - I’ve asked you personally before to share your 3d modelling work samples on this and other threads - but I guess you are probably very busy with work so may not have had the chance to do so.
If you could squeeze in a minute or two to share some sample output, it would be really beneficial to helping you once and for all determine if the Mac is the right tool for you from a 3d rendering perspective and see if we can find you an even better Mac optimized workflow.


Thanks

 

[jujoje]

Four standard uses off the top of my head...

• 4K/8K video I/O cards
• Audio DSP I/O cards
• High-speed networking cards
• M.2 NVMe SSD RAID cards

Why do some folks insist on thinking that PCIe is good for nothing but discrete GPUs...?!?

Fair point; wasn't thinking about Audio or storage (this being a thread on rendering and all). Curious as to whether we'll get a discrete Apple raytracing / compute card than anything, but given the elliptical thread on this on the Mac Pro forums the topic has been kinda done to death. Guess we'll have to wait and see.

The ultra does however scale very nicely in some sim software like axiom for houdini where it touches 3090 perf. But as we say here, for rendering there is no good solution and a nvidia pc even in the lowest tier performs better.

Do you have a link that shows this?

There were benchmarks on the Axiom sites showing it beating a a 3090 (I think I posted it somewhere in this thread), however the site has been redesigned and can't find them. It's also pretty much the only solver I can think of that uses Metal rather than OpenCL / Cuda.

Curious as to Houdini sim speed though; from what I recall SideFX haven't optimised the pyro solver on AS yet and it slower than it should be (although having 96Gb memory means that it theoretically should be able to produce nice high res sims, unlike Nvidia cards). Not sure if any of the other GPU accelerated solvers would benefit as much, but curious to hear how it performs, even anecdotally.