[choo-chee]

well I'm still on version 4. I'm thinking about version 5 but can't find any difference that matters to me (arch-viz). I'm not a substance user, I don't do product shots... and most of the scenes I work consume 10-30 Gb of RAM. So actually I have no use for GPU rendering as long as it can't take some of my RAM ? because maxwell's GPU engine can't handle those scenes (too much memory needed) ? is there a plan to extend the memory usage of the GPU engine to handle physical RAM or do I stick to version 4 ?

[luis.hijarrubia]

Well, using CPU ram on GPU is not an option, that will be so slow that probably is better to use CPU engine.

We are working on 5.2 needing less memory for voxelization. We can try to compress textures. And with time 16 GB cards will be the standar.

Next step above that can be out of core technology. But that's like making a new engine. Big film engines like disney's or pixar's have that.

[choo-chee]

so just a dumb question than, as 16 Gb cards are still expensive and not enough for a typical heavy arch-viz scene... is it possible maybe to have the GPU engine use the CPU for voxelization and use the CPU RAM to store part of the information of the scene that (I really don't know how things work ...) are not getting the speed down ...? I know GPU can be fast but it's like saying the engine is good only for small scenes

[luis.hijarrubia]

Well, actually Voxelization is done on CPU and result is transferred to GPU, we still have to make the process work on GPU (and that GPU voxelization should be used for CPU production engine too, speeding up CPU preprocess). But you cannot work on GPU, having data like the Voxelization Structures on CPU RAM. Every tome you cast a ray (and you are casting some billions rays here) you should wait for RAM to check if the ray hits a certain geometry. For having a quick paralelism, GPU memory is like going to another building and take the data. CPU memory is like going to the moon for the data.

All of this, we are talking about short term. At long term we can make things. One would be Out-of-core. That should make virtually scenes of any size work. Basically you do things on pieces, and order your rays/samples/textures whatever. So when you go to the moon and bring something back, everybody that needs that uses it and then go to the moon again for the next thing. It works better than you may think. But you need to rebuild all the engine and sort things a lot.

Another way to go can be use parts of GPU engine on CPU one. Something like do trace of rays on GPU, but shading still on CPU. But we would addapt engine for that, as you would need to go for thousands of samples at the same time on CPU instead of #number of CPU threads.

[choo-chee]

Thanks for the explanation. I understand things much better now. So eventually we'll use GPU in a few years time but as for now the only way to make things faster is still optimize the scene as much as possible (also get a faster CPU... ). See, since most of my renders demand today about 20-30 hours for a good SL without denoiser and 60-80 hours for interiors that need denoising anyway... making CPU 10-15% faster doesn't really matter (I'll still leave renders for the nights and weekends) - unless there's a way to make some combo so make calculation 30% faster for minimum

[Alexander Kuhnert]

Hello everyone,

i felt the urge to respond to this recent discussion since we are still hesitating to use MW 5 for our "real" production workflow, although we already bought licenses in an early excitement about the new version months ago.

Besides some obvious flaws in the current MW5 version (no SSS in GPU, still fireflies in GPU, overall memory consumption and the incompatibility to our normal project / file size, too many crashes) which makes the everyday usage of the engine unfortunately still questionable , i realized that this specific discussion feels rather preposterous.

I mean it is 2020 and maxwell render exsists alongside engines like unreal engine where you get raytraced images within some seconds to minutes (if you use it as an archviz renderer for highres stills) and not in 60-80 hours....
I know tha the overall technical philosophy and raytraced precision in these two engines is not fully compareble. But still - just compare those numbers. In an odd way that fits in the comparison with going to the moon or to the next building (regarding the memory speed).

No offense but i hope you guys at maxwell find a way to cope with these problems anytime soon since otherwise there wont be many reasons left to use maxwell.

[choo-chee]

I have to add, 15 years ago, working with mental ray or turtle render, for arch-viz, and then one day testing early maxwell was also like going to the next building for getting a decent result or going to the moon for getting something good enough with mental ray or turtle... but today it's changed, render engines are fast. maxwell still has great quality but it's gonna be obsolete without getting significantly faster. My clients say, we like your quality and service but almost any of the other guys give us final images faster than you....

[luis.hijarrubia]

I think 5.2 will be a major update. And GPU memory usage will be one of the improvements. Multilight for GPU (wich I teased some weeks ago) too. And many other things that we will be announcing when it's possible. For me the most important it's that after some years in wich Maxwell looked like it was stopped, now we are developing and realeasing new versions frequently, and the product gets better day by day.

I know that you compare products, but maybe comparing with Unreal it's too much. I mean Epic is an international 17 thousand million company, we are a local small company.

But i feel we are worlking better than ever, re-building many things, so that in the future we can have some very nice features. And speeding up some parts of CPU engine with GPU, making kind of an hybrid, can be an option.

[Matthew Hermans]

I can add a little bit more to this.

I've used Maxwell since the Alpha days 15 years ago. I still hold onto it for total ground-truth rendering although it's a process to get data into Studio as i work in Houdini and Unreal Engine mostly.

A few years ago we rebooted as a realtime VFX studio and have been pushing realtime workflows since - and it's really amazing what you can do with instant feedback, motion capture, virtual cameras and realtime compositing...there's never been anything like it.

But the truth is is that realtime costs quality - we've produced a few projects that have rendered out of Unreal at 1fps or slower and still had a lot of quality issues. So i started to recognise that quality was dropping so I started looking at ways to get the benefits of realtime but still have a process to render higher quality final results. At the moment that's happening through Octane which has a plugin for Unreal that makes it quite easy - although no less easy than my Houdini -> Maxwell Studio process - and this lets me get all the expected quality out.

So based on my experience it's still wise to consider having a hybrid pipeline and keeping as much asset/lookdev outside of the render packages. We use Substance to help with this and this was one of the reasons I started putting attention back into Maxwell Studio (because of the import wizard)

I hope that adds something : )

[Brian Buxton]

"Next step above that can be out of core technology. But that's like making a new engine. Big film engines like disney's or pixar's have that."
Not true. Every other established GPU renderer has out of core tech. It is not something exclusive to Pixar or big film Studios.
Octane has it, Redshift has it, VRay has it, Arnold has it, Keyshot has it, Mavericks has it, Cycles has it. Maxwell does not.

Additionally, given that most or all of the host plugins are feature incomplete or problematic due to other issues, perhaps offering a +Studio upgrade at a vastly reduced cost might help you hang on to those on the verge of walking away from Maxwell.

[luis.hijarrubia]

My bad. I ment full out-of-core. Most of those examples have out-of-core only for textures. That one is easier, and we can do that without remaking the engine.

[Nasok]

As much as I love the ability to render in Maxwell straight form Maya - I believe that the best way is to reduce the complexity of the plugins and rather make them functional exporters for Studio.

There was a time when I thought completely the opposite - now I understand better how most of us work. And I think the best way is to work exactly like Matthew - personally I started doing the same thing - it is a bit a curve to learn but it is definitely taking off pressure from having a fully functional plugin for my host app.

Consider this - majority of CAD apps are not really suited for a convenient rendering workflow anyway. Some apps don't even have the essential UV mapping tools, or IBL workflow, or light targeting, or light filters (like barn doors), emitter blockers, and etc. Why would you want to start rendering workflow inside such app ?

So - instead of having plugins for every app that could benefit from having Maxwell connection - why don't we just have an exporter to the Studio (say, a free plugin) that does one thing - packs your scene and exports it to MXS (with all the goodies) - that's it - no FIRE, no Maxwell Sea, no Grass, scatter - no realflow import - none of those things. Imagine how lightweight this plugin will be.
Development - waaay easier.

In the end - NextLimit will be able to afford to have such "connections" for pretty much any DCC app.
And MAxwel lStore is super simple now - instead of dozens of versions - you now have Studio - that you buy - and free exporter for your host app. That's that.

... and plus render nodes - for the good measure.

Easy

[Brian Buxton]

My bad. I ment full out-of-core. Most of those examples have out-of-core only for textures. That one is easier, and we can do that without remaking the engine.

I see. That is a useful distinction and makes things a little easier to understand.

My frustration being stuck with a limited and problematic plugin for FormZ without Studio to fall back on still stands.
It is too late to switch it to Studio but that is not desirable either.
I think most if not all of the plugins fall short of expectations next to Studio and Studio lacks the convenience of having Maxwell Fire in a CAD or Modelling environment.
I would be interested in a fair price to have Studio as part of the same license.
Without this, I don't think it is viable to continue using Maxwell (there is no option to recalculate normals in the plugin, making it currently unusable when rendering nurbs and other smooth surfaces)

[choo-chee]

first thing I didn't know that most other engines "out-of-core" means textures only. so those are useless too if you have a typical arch-viz scene. second I must say that to me - at least Maya's plugin - is very good, works almost flawless 99% of the time and I see constant improvement. As it appears with GPU (to my eye at least) it will remain quite useless unless display cards with tons of RAM will be very cheap, sometime in 2040

[Brian Buxton]

I'm not sure about the textures only thing. I have an old 12 gb GTX Titan but my system has quite a lot of Ram. With it, Redshift can render huge scenes as long as I remember to turn off progressive rendering and use buckets instead. I just tested the beach scene from Disney Moana Island and Redshift was able to render. I could not do much more than that because the scene is so big it is not practical to do much more than just load the files and render in Cinema 4D.