there and back again, a Render Journey

[schwungsau]

I've start test Houdini 18 and Arnold 6. the first test was simple splines rendering, 250.000 splines instanced 25 times. it loads a 140MB alembic file.

rendered in 6 core Xeon CPU and Nvidia Quadro RTX 5000. (windows 10 pro)

the startup for Arnold GPU is slow, it renders faster, so it seems but for clear up the final image it takes for forever or just dropped /crashed, hard to tell on the GPU. the CPU is quite fast but much slower then GPU if it ever would finish. (adaptive sampling was on)

As soon as Arnold finishes rendering the scene, it stops and do not refresh any more on parameter changes. so far i am not impressed with the Arnold GPU rendering.

 

here is the same scene Arnold CPU with only direct Lighting. (on my MacBook)

some test with Arnold GPU. it performed much better with just direct lighting.

 

Edited January 31, 2021 by schwungsau

[schwungsau]

2 minutes render time, left CPU and right GPU with high poly mesh 1.9GB data. (6 core Xeon and Quadro RTX5000)

CPU

 

GPU

 

in this cache, the GPU was faster and more stable. its response time on shaders changes not too bad. only sampling changes took 3-4 seconds to respond.

Octane was crashing on this scene while loading the geometry. Blender Cycles needed 4 minutes to load the alembics.

 

 

 

the CPU is much more efficient in cleaning up the final noise than the GPU. the render result is line up on both, but sampling is quite different. it would be impossible to mix CPU and GPU renderings.

In the end, the GPU rendering is a little faster the CPU rendering (once you optimized the render setting.) Optix GPU rendering does not support GPU BVH branching only pure AA samples, that's properly why I am not getting the excepted massive render speedups on the GPU.

Tests in Blender showed CUDA rendering are faster then Optix rendering (with RTX).

 

 

 

[schwungsau]

Renderman 23 sampling. i am testing the different integrators of Renderman 23 in Houdini. the limit was 2 minutes render time on a 6 core Xeon, one HDRI dome light. Alembic geo 1.7 GB file.

direct lighting

Path tracer

 

pxr unified

pxr unified bidirectional

 

pxr bi-directional path guiding

pxr vcm

 

the shader settings for pxr unified integrator need to be adjusted to match the look. it took 30 seconds for the first pixel. the interactive responds time is excellent. the clamped highlights come the buggy Houdini 18.0.287, below some scene with Houdini 18.0.237

the dull renderings / images comes from houdini production bug, newer houdini build fix it. 

Edited March 9, 2020 by schwungsau

[schwungsau]

Renderman 23 CPU vs Arnold 6 GPU with 2 minutes time limit.

 

Renderman 23 CPU

Arnold 6 GPU

 

the reflection differences come from normals, Arnold calculates it's on normals, Renderman took the normals from the Geometry. i've tried with Karma but i had issues and crashes with it.

Edited January 7, 2020 by schwungsau

[schwungsau]

rendered with Arnold in Houdini. I've tried the atmosphere volumes the first time. easy to setup. The render time was quite slow on the volumes, typically for volumes, but much faster then render the screen envr with VDB cloud. I've used a mesh light for inner character illumination.

 

it's a shame I could not use Arnold 6 GPU, because of its missing atmosphere and meshlight features. Volumes would boost get a huge speed boost with volume raymarching on a GPU. I had to use a denoiser from Affinity Photo in dark areas. with Arnold 6 GPU it does not drop any errors, you have to wait 1-2 minutes if data is loaded to GPU and starts the rendering or not. even a wrong file pass to an Hdri image makes Arnold GPU drop out, you sitting and wait in front of a black screen and don't know if it will render or not.

 

 

 

 

Edited January 10, 2020 by schwungsau

[schwungsau]

I've tested RenderMan with spline rendering. it's much faster compared with baked Geometry. the spline primitives are very good in Renderman.

I've stopped the test with mantra after 10 minutes, it is super slow even with optimized render settings.

the Renderman images rendered for 3:20 on a 6 core Xeon
2.7ghz with pxr pathtracer. 4k resolution

geometry:

splines primitives:

rendering splines directly saves a lot disk space and meshing time. interactive rendering was super responsive.

Octane 2019.2 renderings is very fast, I've used the geo to render it. 2 minutes, as except its lot faster the CPU renderings. the spline rendering in octane does features custom width yet.

Octane gives physical correct shading out of the box. Renderman the shading is extreme flexible with NPR renderings, which is hard to get in Octane. the IPR responsibility was a little slow with shader tweaks.

[schwungsau]

test with previs Scene but with Blender realtime Eevee render. 2 million polygons. 10-20 fps on RTX 5000. (live viewport below)

the offline rendering took 3 seconds not sure what is causes, probably loading time

houdini openGL rendering:

the prman 2-minute rendering for comparison :

 

renderman direct lighting 

 

 

Realtime engines are on the rise! I will keep an eye on this!

 

 

 

 

Edited January 31, 2021 by schwungsau

[Iskander]

Did you test 3Delight?I found it pretty much faster than cpu Arnold/Mantra/Renderman,x10+ on some tasks and it really wonder me..

https://www.3delight.com/
https://discord.gg/MGtJx4q
https://gitlab.com/3Delight/3delight-for-houdini

Edited January 14, 2020 by Iskander

[schwungsau]

On 2020-01-13 at 7:16 PM, Iskander said:

Did you test 3Delight?I found it pretty much faster than cpu Arnold/Mantra/Renderman,x10+ on some tasks and it really wonder me..

https://www.3delight.com/
https://discord.gg/MGtJx4q
https://gitlab.com/3Delight/3delight-for-houdini

yeah, I've heard they back from the dead. its looks great so far, currently i can't beat Renderman. plus its beta, I had issues with portrait camera format.

image below is with 3delight, clock in 4:20 min

 

Edited October 3, 2020 by schwungsau

[schwungsau]

older scene.
Houdini openGL rop  4 sec a frame in 4k and DOF. Arnold and Mantra took both 2-3 minutes to render it in 4k.

 

Edited January 15, 2020 by schwungsau

[schwungsau]

testing new Arnold 6.0.3 GPU renderer update on particles

 

Recent HtoA 5.2.1 made it possible to actually use Arnold6 GPU in Houdini, i tool it on a test ride with particles. (5 million particles). 

the setup is quite simple, pure diffuse shading only AA samples. I've chosen a darker frame to test the sample quality. 

CPU 2 minutes on 6 core Xeon :

 

 

 

GPU 1 minute (Nvidia Quadro RTX5000),  same sample count the CPU render above but more little more noise.

 

 

GPU 1:30 minute, increased the sample to get a noise free render.

 

the Arnold GPU render it getting slowly faster the Arnold CPU in some cases, there is still long road ahead to for the GPU renderer speed to catch up with competition.

Edited October 3, 2020 by schwungsau

[schwungsau]

testing Arnold 6 in Solaris for lookdev in Solaris. its has really potential and its fun to work with, only a Arnold render HDA is missing. Pixar provide one, but its easy to build a renderROP HDA. 

the scene was running on my Thinkpad

 

 

GPU renderspeed comparsion with indigorender

 

close up with indigorenderer

Edited January 31, 2021 by schwungsau

[schwungsau]

Arnold with spline rendering from a USD file. first time pixel on CPU and GPU is about 3 seconds.

[schwungsau]

the testing Lee Griggs tricks with a textured volume inside a class object

Arnold rendering

 

 

 

this is rendered with Cycles in Gaffer, next I will try Hydra version from Cycles inside Houdini 

 

[schwungsau]

Indigo renderer with a simple glass medium 

[schwungsau]

its older but thats a hypercube with glass shader in arnold with fake caustics, its quite fast. (6 minutes)

 

same cube with physical correct light simulation in Indigorender, true caustics. it renders in 2 hours, but at least 10 faster than any path tracer (CPU or GPU) without fake.

 

[schwungsau]

I am starting to test Arnold Renderer in Houdini Solaris. it took me a while to figure how to makes it work in Solaris context, especially with NPR shaders. I’ve need to create beauty render pass and render product with contour sampling.

 

[schwungsau]

playing with idea to use the dot product of incoming ray and surface normal to generate superfast fake caustic’s. its renders fast ! (1minute) and its full "art-directable"

here are some tests:

 

[schwungsau]

light cone in Arnold with athmos volumetric :

[schwungsau]

water rendering with arnoldGPU