Three years ago, Maxwell accessed the GPU rendering market with its release number 4. But too many materials support limitations created a lot of frustration and disappointment. Will release number 5 show the return of Maxwell in the GPU arena?
Major improvements but no equivalence yet between CPU and GPU rendering
One of the improvements is the support of materials in additive mode. Maxwell’s materials system works with layers. You pile features (brightness, scattering…) to create a material. They are frequently used in scenes and the lac of them in Maxwell 4 limited the interest of the GPU rendering. It is now available –at last- and works fine. Let’s make a few comparisons between materials calculated with the CPU or the GPU. GPU rendering is much faster but there are too many differences of configuration between my CPU (i7-990 / 6 cores) and my GPU (2x GTX 1080) so I can’t really draw any conclusions with figures.
CPU on the left image / GPU on the right image
CPU vs GPU render
Opaque CPU vs GPU render
A few materials are not supported, like SSS, coating, dispersion or procedural textures. Displacement is well supported though.
I noticed a few differences :
Two new procedurals : Random color and Random UV
Once again, I would tend to say « At last ! » It’s funny how things that seem to be obvious and simple for grap^hic designers and 3D artist can –litteraly- take years to be implemented in software. The interface is not really intuitive and having a bit more control would be nice. Why don’t they get inspiration from features that work well like in Fstorm or Corona? That’s a mystery. But it remains a must-have tool in a renderer. Bare in mind that, being procedurals, these features will only be rendered in CPU mode.
Native import of V-Ray scenes into Studio
There was a scene converter in the 3ds Max polugin of release number 4. In this one, you can now import V-Ray scenes (file: .vrscene). Once again: at last! Maxwell suffers from a lack of ready-to-use models. Having an easy conversion of V-Ray models and materials at the Maxwell format was necessary. Materials systems are quite different so you’ll need to add a finishing layer but it really helps anyway.
Rendering can be made on several GPUs. Nvidia
graphic cards supporting CUDA are mandatory. It is also recommended that the
memory available in each card is the same. You can activate and de-activate
cards that you want to use for a specific rendering but the Fire (interactive rendering) uses only one GPU.
A few limitations remain like the multilight handling, a few rendering channels like the shadow and the alpha customs, Maxwell Grass and volumetrics.
A few bonuses like:
A CPU cloud rendering service has been integrated in the interface of network rendering. A test image has been rendered on this service in 16 minutes for €1.06. Just to compare, the rendering took 2 minutes (8 times faster) with Ranch Computing in low priority for €2.38.
Maxwell 5 is the release we had been waiting for 3 years ago. The main new feature is a better handling of materials in GPU rendering, with the limitations I mentioned. But the GPU memory use seems slightly less good than with competitors. It means you have to keep an eye on the models’ number of polygones, the seize of textures and resolution of your renderings. A total switch to GPU does not really seem possible to me today but rendering duration can be really reduced for projects that are really under control (as far as materials and memory occupancy are concerned).