Maxwell Render | Official public forum

Yesterday, I was very pleased when I saw Improved performance in scenes with a lot of grass and/or hair in release notes for new beta 3.0.99.6

I did a quick test with one of my complex scenes (1.6 milion hair blades, multiple lights sources, displacements, SSS,...) and I was very surprised about result.

But today I decided to make in-depth test, focusing on how dramatically it was improved and I must say it is freakin awesome improvement !

It was very simple scene:
- one plane scaled to 2x2
- added Grass modifier
- all three types were tested Flat, Curve, Cylinder
- Points per Blade was changed to 15
- Density to 100 000 blades/m2 [total count was 400 000]
- length to 18 cm
- rest of settings weren't changed
- V1 is lit by default Sky + Sun setup
- V2 is lit by single emitter from left side
- every scene was rendered at 1200x900px SL12

I used Pin Hole Camera and simple material (only one bsdf with roughness 97) to be sure that depth of field or material have minimal effect in final result.

Both scenes were rendered on:
- 2x Xeon E5 2680v2 (40 cores total)
- Asus Z9PE-D8 WS
- 32GB RAM


TEST SCENE V1


TEST SCENE V2


RESULTS


NOTE:
only "downside" of this improvement is memory consumption which is increased now about 50%. In 3.0.99.5 scene V1 took about 2,2GB, in 3.0.99.6 took about 3.3GB during the rendering process.

Anyway as you can see, or try it on your own, it is really nice improvement

Thank you NextLimit and MaxwellRender team for providing another great early access build We are looking forward on 3.1 release ! Woah!

very nice comparaison
i installed this version this morning ;after few hours windows stop working than i can't use it !

Very promising indeed!

That looks incredibly promising !

Thanks for the detailed tests I wonder also if you set less points, perhaps the Curve type would be even faster compared to before. 15 points per blade is quite a lot.

Awesome comparison!

Thanks for this very interesting comparison!
Almost 4 times faster looks great, but what bothers me is the 50% more RAM that is needed now...
I think a more balanced solution would have been better - maybe 25% more RAM and 2-3 times faster (if it is possible to control it this way)
I was glad to see an improved RAM usage with grass (i'm not sure if it changed in 3.0 or 2.7 or even earlier). I still have 3 machines with only 16GB RAM
and with 50% more RAM usage for the grass many projects would have run out of memory on these nodes...

I was hoping to see a less memory intensive solution for the grass with maybe a limited number of different grass objects - instanced. Similar to a scatter solution like ForestPack, but with parametrically generated objects. Should be possible in combination with maxwell scatter.

Keep in mind in his test the RAM usage is exaggerated because he used 15 points per blade. Especially for the curve type, this is not needed. For the curve type since it's an analytical perfect curve, only 3 points are really needed.

I like it!
Thanks for the test.

Mihai wrote:Keep in mind in his test the RAM usage is exaggerated because he used 15 points per blade. Especially for the curve type, this is not needed. For the curve type since it's an analytical perfect curve, only 3 points are really needed.

Sure, but there is no differentiation between the different setups, so i would read this as percentage for all situations.