Mental Ray getting slower with each release

#1

After 10 years of Maya development, we get this (see atttached image).

Lol. I don’t really know what to say about this test result… its a rendering thread but seems quite major so I put it in General.

Ok so the scene, light & materials are identical as are legacy global settings (a straight import). No radiosity eg. FG/GI.

So apparently for a scene this size the Mental Ray core engine is 40% slower than 10 years ago. That’s got to be pretty embarrassing if you’re a lead developer.

I could install as far back as 2013 I think on Subscription… bit of a hassle, if anyone already has a previous version installed pls msg me and I can send you the Legacy scene to test. It would be good to know where along the line Autodesk screwed up.

With a smaller scene the difference is more pronounced (up to a 50% performance hit, eg. 5s Vs 10s), on bigger scenes the difference is less but still around 30%, which is huge.

#2

if you like to render lagacy stuff use the old mental ray…

modern render engines work only fast with new materials, new gi solutions, new lights, new sampling methods… if you force them to work like in the past they wont perform at all…

and thats the same for other render engines…
compare vray1.2 to vray 3.3 and you will see the old one is faster if you work like in the past…

but you wont get the same quality of production images… you wont be able to render the same amount of data…you wont get clean motionblur… you wont be able to render a huge amount of lights…

but i agree is could be better with mental ray… but there is a reason for… they are in the middle of the defelopment of new mr core… and there is a lot changeing under the hood…

#3

No they only work fast with new materials if some of the previous libraries were re-written poorly for some reason. The renderer’s initial load time will likely be slower with a modern renderer, but process time certainly shouldn’t if there is no feature disparity and the renderer doesn’t need to make calls to unused libraries.

I’ve also had the chance to test it with a few versions later so both versions could use newer Mia shaders - same performance hit.

2016’s new Unified sampling is apparently more efficient - but quality wise its no better - getting it to the same quality as Legacy produces similar render times depending on the scene.

This is currently lodged as a support case with Autodesk - the thing is it might not be a legacy issue, and everything could be rendering slower. The new Mila shaders for example are certainly faster and better designed, but it would suck if users weren’t getting their optimal performance. So hopefully Autodesk will have something interesting to say.

#4

In my honest opinion, it doesn’t make any sense to compare scenes like the one above. Nobody renders scenes like that and you can’t measure real production performance with some specular spheres.

Of course rendering doesn’t perform faster with those sphere on a plane scenarios. Unified sampler was desinged for heavy scenes with motion blur, complex glossy and what not. And most importantly, mental ray indeed sacrifices performance for ease of use. This means you don’t have to control samples on every material anymore with mila.

#5

The discussion isn’t about why isn’t it faster… its why is it slower by such a large amount… it might be an issue which only an Autodesk software engineer can answer.

In many cases you definitely can extrapolate a simple test result like this to real-world scenes - especially in architectural work which is mainly just polycount, shaders & lighting. As long as the polycount of the test scene is high enough and there is some complexity by way of glossy reflections or whatever to get a sufficiently long render as a baseline.

That being said it would be worthwhile to test out Unified sampling with motion blur for when I do need animation, as I wasn’t aware it was faster with blur.

#6

our maya 2015 scenes using MILA shaders render about twice as slow in maya 2016. I went in and turned off all the importance sampling that we already had disabled in 2015, put the bsp settings back to bsp, and shadows back to simple, but the render times are still twice as long. They actually get significantly worse as render resolution increases - 2x slower at 4k while 15% slower at 540p

Slowly trying to get to the bottom of it. Thinking it might be one of the extra string options that were added to the 2015 scene. The renders are different between versions with the lighting. You can tell something about importance sampling is still happening even with it turned off.

I’ll experiment again with unified sampling and start taking shaders apart to try to isolate what the problem is.

I really don’t like how maya takes settings that were specifically saved with a scene in an older version and changes them when you open in a newer version - like the bsp settings etc.

#7

In my experience, the new default for shadow has a big impact on render times. Try switching from segment to simple.

#8

I have determined specifically where Mila is slowing down - I have not been able to confirm if its a 2016 specific issue.

I lowered the global reflection bounces but saw minimal improvement - even lowering to 1 bounce (making renders look worse), render times were still much slower than Mia with 5 bounces. However removing all bounces or eliminating roughness on the glossy reflection node saw a huge speed improvement.

So this issue is occurring with roughness on the glossy reflection node, i.e. perfect mirrored reflections render fast. Still… no modern renderer should be adding 2+ minutes to a low resolution render with only 1 reflective bounce… its not normal.

I have a 2016 test scene… in theory if someone could test that in 2015 and compare to my times it would determine if its a 2016 issue… problem is its too big to upload at 1.6MB :confused: