Bad GPU scaling in 3ds Max 2016 Iray.

[Out-Of-Light]

Hi,

I was having a discussion about the new Titan X for iray rendering on another post.

There were a few misleading benchmarks and a few misleading posts from users that didn’t actually own the card yet, but eventually it was sorted out, that the Titan X is not crippled in iray, and that it works fine in Max 2015 SP2 with the Maxwell patch applied, and in 2016.

For anyone interested there are some reliable benchmarks on the Migenius website under products, Nvidia Iray, Benchmarks Max 2015.

I am holding off on getting the new Titan X, A: because stock is limited so prices are still high, and B: AMD will launch their new 380X 390X cards soon and they are supposed to be on par or slightly faster than the Titan X so this should force prices down.

I found one person using iray and owning a Titan X and asked if they could perform a speed test for me, and it was proven that the card is around 45% faster than the older Titan 2688 core. Meaning say a 1280x720p render at 11,000 iterations of an architectural interior taking 25 minutes using 4x Titans could be reduced to around 13 minutes using 4x Titan X, so if you are needing multiple HD noise free images in a day like me for client projects, the cards could be worth the investment.

But,

During the speed test comparison, I noticed something on my own system that is a little worrying about Max 2016 and Iray, in that although rendering speed seems to have improved using a single card, multiple cards do not scale as well as previous versions.

I used Nick Clarks 500 iteration iray test on maxforums which is no longer available but I still had the Zip file to hand.

I ran the tests several times not letting the CPU’s/GPU’s clock down.

My system uses 4x GTX Titan (2688 Cuda cores each) and 2x Intel Xeon E5-2687W with 40 pci-e lanes per CPU, which reduces bottlenecking.

I carried out the 500 iteration test in Max 2014 about a year ago, but the results have not changed using 2015 SP3.

Max 2014/2015

1x E5-2687w 46 secs
2x E5-2687w 29 secs

1x Asus GTX Titan 25 secs
2x Asus GTX Titan 13 secs
3x Asus GTX Titan 09 secs

4x Asus GTX Titan 07 secs

4x Asus GTX Titan + 2x E5-2687w 06 secs

Max 2016

1x e5-2687W 54 secs

2x E5-2687W 32 secs

1x GTX Titan 20 secs

2z GTX Titan 14 secs

3x GTX Titan 11 secs

4x GTX Titan 09 secs

4x Gtx Titan + 2x E5-2687w (HT) 07 secs.

The Titans scale almost perfectly in Max 2014/2015.

HD 11,000 iteration render taking 25 minute to complete in max 2015.

1x GTX Titan = 25 minutes

2x GTX Titan = 13 minutes | 25 / 2 = 12.5

3x GTX Titan = 09 minutes | 25 / 3 = 8.33

4x GTX Titan = 07 minutes | 25 / 4 = 6.25

So the scaling is very close.

In 2016, it starts off looking promising using one card with a reduction of 5 minutes, but then it all goes wrong.

This is exactly the same scene used in 2014/2015, no changes needed, at the same iterations and resolution.

The same HD 11,000 iteration render now taking 20 minutes to complete in max 2016.

1x GTX Titan = 20 minutes

2x GTX Titan = 14 minutes | 20 / 2 = 10

3x GTX Titan = 11 minutes | 20 / 3 = 6.66

4x GTX Titan = 09 minutes | 20 / 4 = 5

The scaling is very bad compared to 2014/2015

So using 3x Titans in Max 2015 is actually as fast as using 4x Titans in 2016, and using 4x Titans in 2015 is fasters than using 2016, but using one card shows an impressive improvement. There also seems to be a slight performance hit using CPU's.

Why is the GPU scaling so bad in 2016?

I hope Nvidia/AutoDesk are aware of this and a fix will come soon, or I will end up having to do all my production rendering in Max 2015.

I do not own 4x Titan X so I cannot say whether or not the scaling works better on Maxwell cards, but I will probably upgrade in a few months when the prices come down or wait for the Titan X2.

Can Anyone else using multiple GPU's test and confirm this?

Are AutoDesk/Nvidia aware of this and working on a fix?

Many Thanks　

[BenBisares]

Hmmmm...I'm not aware of any scaling performance issues with 2016. Took a quick look at our database and there doesn't seem to be anything related.

Still you did a lot of leg work and your results don't lie. Let me write this up and submit this to the devs.

[Phillilp_Miller]

Thanks for the detailed report - we will look into it.

While we support Titans (and GTX in general) we do all of our testing on Quadros, so it's possible there's someting slipped through. The renderer underwent a massive change between Max 2015 and Max 2016, so it's pssoble that there's's tuning still to be improved.

Now, I am surporised (and bothered) you didn't see a better speedup between 2015 and 2016 though, as most scenes will go 1.5X to 2X faster now. Could you describe the lighting charactersistics of your benchmark scene and what Max materials/shaders you are using in it?

- Phil

[Out-Of-Light]

Hi Phillip,

AutoDesk Support are currently investigating the issue.

As scaling works very well in 2015, and not in Max 2016, I don’t think its something to do with my hardware set-up, or something I am doing wrong with materials or lighting, and more likely as you say some more tuning needed with the software or drivers.

Titans in past versions of max, including 2015 have been certified by AutoDesk for 3ds Max, and passed all the tests, but not as yet for 2016.

I tried two different tests, one using low poly models with different textured materials, and another concentrating more on high poly count , no textures, using Mental Ray Arch & Design materials, and Photometric Lights set to raytraced, spherical. I have been using Iray as my main renderer for over three years now, and the light and material set-ups I use have never caused scaling problems. I calculate how long renders will take every day, usually between using 3x GPU’s if I want to continue working while rendering, or 4x GPU’s to speed up render times, so I am quite used to making these calculations, that’s why I noticed the scaling is off in 2016.

Although I rely on the Iray material plugin for most of my scenes, there is not currently an official 2016 plugin release, so was not used in the comparison tests.

The test I ran using a high poly count and standard Arch & Design material is below, but the results are the same, with one GPU showing a 39% performance increase compared to Max 2015, but the more GPU’s added in Max 2016, the worse the scaling gets.

I used a 16,000 ploy model, a component of a detailed sci-fi model I created, with subdivision applied, and set the render iterations to 200. This is a 1280x720p image.

　

Max 2015 SP3

GPU’s Only (GTX TITAN (2688 core))

1) 49 secs

2) 25 secs 49/2 = 24.5

3) 18 secs 49/3 = 16.33

4) 14 secs 49/4 = 12.25

Max scaling difference = +/- 2 secs

Max 2016

GPU’s Only (GTX TITAN (2688 core))

1) 30 secs

2) 20 secs 30/2 = 15

3) 16 secs 30/3 = 10

4) 15 secs 30/4 = 7.5

Max scaling difference = +/- 8 secs (This would be minutes on a HD noise free 11,000 iteration image so quite a difference)

Although the tests with low ploy+ Materials, compared to High poly varies a little, both tests conclude that GPU scaling in Max 2016 is way off compared to previous versions. One GPU performs better than previous versions, but the more GPU added, the worse the scaling gets.

Its important for me to be able to accurately calculate render times, so really hoping the scaling problem can be fixed in Max 2016.

I am also pretty sure that the scaling problem will be evident using multiple Quadro cards, as there is a big difference between optimising viewport performance by drivers, and the way Cuda cores are handled for rendering, so the results using say a Titan X and an M6000 should be about the same.

Definitely need to be fixed.

The image attached is the 200 iteration 1280x720p image I created, using a component from a much larger sci-fi model.

Many Thanks.

Simon Pestridge.

[Out-Of-Light]

Attached is the basic interior scene i used for the 500 and 11,000 iteration tests.

I also used Nick Clarks 500 iteration test from Max forums that showed the same poor scaing in Max 2016.

The 500 and 11,000 iteration scene is an old client project, and the 200 iteration test is a component from one of my own concepts, so prefer not to share the files publicly, but I am sure a simple iray scene set up will show similar scaling problems.

[Out-Of-Light]

Autodesk support came back to me and have confirmed that there is a GPU scaling issue in 3ds max 2016, and are handing the information over to the development team. Hopefully for a fix sooner than later 🙂 Autodesk had to wait for the necessary hardware to arrive to replicate the problem themselves.

Anyway, glad its not just my system.

I will have to continue with Max 2015 SP3 for the time being for Iray rendering. With multiple GPU's its about the same speed if not faster in some scenes than Max 2016, but at least I will be able to accurately calculate render times.

If Autodesk can get the same 39% speed increase using one GPU to work on a multiple GPU set up in Max 2016 , this would be awesome 🙂

Thanks to Ben Bisares for his help on this issue.

[junz]

3ds max 2016 sp1 on windows 10 Iray rendering on maxwell card even slower than windows 8.1 pro

Autodesk plz fix this so annyoing

[waracana]

I have 2 Tesla K8 coprocessors and 2 quadro 5000 in a HP Z800 workstation at work. I was rendering a scene with 3dsmax 2015 using iray.

All 4 devices contributed to the rendering process. I actually tested 3dsmax 2017 only because I wanted to use the iray section helper to broaden the focal length of the camera, which took the camera out of the model. Iray section plane helper corrupted the file, but that´s a whole different story.

Back to iray gpu rendering, I tested the scene, this, again, was an exact copy of the one that was working with 3dsmax 2015. It never got to use any of the 4 cuda devices, it could only do cpu rendering. Last, just to diagnose if it was related to that file in particular, I decided to test the studio sample file available when you first launch 3dsmax 2017. Nothing, this is a very simple file with only a rough concrete floor texture, everything else is plain white..and some photometric web lights, that´s all. Any of the 4 devices contributed to iray rendering, only the 2 cpu were working..

??