VFX on the GPU

iray Visual Computing Appliance (VCA)

Is Visual Effects (VFX) ready for GPU acceleration and rendering? And where is this going for mental ray? NVIDIA ARC is already at work on porting features to mental ray including the preview feature of Global Illumination accelerated on the GPU.

With MDL coming to mental ray in 3.13, this opens the door to even more accelerated and simplified rendering.

Read a more about MPC and GPU rendering here: Studio Daily: The Future of GPU Appliances




About David

I am a VFX artist that specializes in Lighting and Rendering. I spend a fair amount of my time supplying clients with artistic solutions as well as technology solutions. With a background in fine art and technical animation training, I strive to bridge the divide between the artist and technologist.

Posted on April 6, 2014, in Uncategorized. Bookmark the permalink. 26 Comments.

  1. For years I’ve wondered why 3D/VFX took so long to render, presently as a freelancer waiting for renders can be a big time waster, I’m a person whom when knows the tools do things fast, literally.I can only hope that will make me very successful in the not to distant future, produce quality and fast, isn’t that the goal of all pipeline time frames ?

    Anyhow, with the rise of GPU rendering, I think the technology has caught up to the speed that I prefer to work, watching this industry grow from the sidelines, it has taken up to this point to reach this level only brings up the question of, why ?

    I recently watched a GPU video, show casing the new features in Maya as well as iray, when one of the presenters compared a real life photograph to the iray render, I was sold, instantly, especially on GPU rendering. The only catch is, it comes at a cost, recently I was looking at Octane Render and iray, both give superb results, except one is more costly then the next.

    CPU rendering won’t disappear overnight, it’s position is still strong. If you combine the two, I think for a small studio or a poor individual such as myself you have all, if not most of the power you need.

    That is my take on it, and I’m sticking to it 🙂

    • I would argue time per frame isn’t most important anymore (or else Arnold wouldn’t be a juggernaut)

      Simplicity is king. And GPU rendering allows for a simple (albeit restrictive) physically based rendering in an instant. Eventually I think GPU may take over the bulk of the work, but there are still things to solve for GPU rendering like good skin shading, hair, volumes, etc. Once these are resolved I think interactive to realtime rendering on GPUs will become a norm.

      Something to note on tech demos is time to converge. I see a lot of “fast” renders out there but few people actually let it converge to a noise-free solution. That’s the end goal: final frames.

  2. I’ve seen a few nice skin renders from GPU render engines.

  3. Some do a pretty close Job.

  4. I have been using Redshift a lot lately. Their philosophy is a biased render on the GPU for film production, so ALL features they implement straight on the GPU. To handle any shortfall in GFX memory, they have an out-of-core caching system (though I am yet to reach my 6gb limit, with clever instancing/proxy systems). It is brilliant. It certainly does IC, SSS, hair etc very well. So you have the speed of itteration from tyhe gpu, plus noise free 2K renders on a 780ti in 1-6 minutes. Best of both worlds. Given that some consumer Motherboards can handle 4 pci3 x16 cards, you can quarter that time. Makes large arrays of CPU boxes redundant, great for smaller opperators. Feels a lot like using Arnold or Vray in its workflows. AOV buffers, the lot.
    Anyways, to get back on topic, 50K! Ouch.
    2 machines, 4 780ti gpus a piece would cost a 5th of that. Admittedly the Tesla in that box have 12gb vram compared to the 3 (soon to be 6) in the Ti’s, but for redshift the out-of-core system and instances management nullifies that a bit.
    Still, cool article 🙂 Love your blog!

    • GTX cards, while cheaper, aren’t designed to render constantly at that output like a Quadro or Tesla (or the thermal envelope). The VCA also has a huge amount of RAM and infiniband. Not exactly a fair comparison for hardware. But it’s not designed for smaller installation where a few GTX being run every once in awhile or replaced often would be fine. Most larger studios sit on hardware as long as possible.

    • redshift is great, but its important to be more realistic in terms of rendertimes. because what i read all the time is way to low for realistic comparision (its the same for octane, classroom noisefree render in 1080p took 8 hours for example). i used it for some projects with a gtx580 3gb with full brute force with noisefree rendering at the end. and my rendertimes were far above the 5 or 10 minutes thing i reading all the time. the more complex the scenes going, the higher the rendertimes, especially with brute force. rendertimes of 2 hours per 1080p are normal (not really noisefree, but good enough for most users), 4 to 5 hours for 4k. my heaviest scene had 24 billion instanced polys, took nearly 6 hours to render in 1080p and took 16gb of system ram. and yes, the scene would render alot longer in mr, same for arnold. and i could not render it in mr, rendertime would be around 20 hours i guess. thats the good thing of redshift, its gpu based and this makes things possible to render which you cant in mr actually when you dont have the cpu power. but this does not mean you can render something in under 10 minutes that would take 20 hours in mr or arnold. 😉

  5. Yeah, fair points :).
    Still, I cant help but feel if you made a 2 machine, 8 card ‘farm’, cooled the cards correctly (water even?), maintained default clocks or very slight underclocks and spent on modest infinband cards and router you could get close, for a LOT less. Doesn’t negate the Ram issue, so maybe limited when doing whole cities full of People/Cars/trees with 4k textures. Plus there is the time/money equation to consider when you build and maintain/support it, vs getting a ready-made turnkey system.
    Anyways, just hypothesis, I don’t have the spare cash to test the theory :D.

  6. I hear Red Shift is the in-expensive version of Arnold. I see to much ‘clean’ work coming out of Red Shift, compared with Octane for example, nothing wrong with ‘clean’ CG when required but when I want to make things dirty that’s where Octane seems to excel at, and it’s still a little faster then iRay from papers I’ve read.

    • If it’s biased rendering then yes, it’s probably faster than iray which is true brute force in photoreal mode. But then it’s not so much “push button rendering” which is what iray is targeted for.

  7. There is more control within Octane, which is unbiased. MR is biased, you knew that I just wanted to note it 🙂

    • Biased rendering will be faster, but mental ray already renders frames on a CPU faster than brute forcer renderers. It’s the process of tuning that users want to avoid. iray is truly brute force (single ray sample every time) that uses in core tech and materials to decide how to render the scene without added controls. Most everything in iray is automatic this way if you let it converge. The ultimate in simplicity is the goal. Biased renderers trade compexity for speed. But this is a vanishing model for rendering.

  8. What is vanishing, complexity or speed ? Most who read this blog know that Arnold stands as king of renders, either that or on par with Mental Ray. What would you consider a good render engine, excluding MR and Arnold ?

  9. I don’t see complexity benefited in the past, they were just extra knobs that didn’t add much support to the render, that is my opinion.

    As for my other question, you deliberately avoiding it ? 🙂

    • Complexity benefited a lot in the past. Have you tried rendering a frame using brute force on a Pentium circa 2003? These tools can still be used today for even FASTER renders. But overall simplicity is more desired. Something along the lines of Blinn’s law. Although we’re less maintaining render times and instead increasing them here for a bit.

      Other renderers?

      I like Renderman and Mantra. I think Renderman’s raytracer in the right pipeline would be a great tool for rendering. It’s just not always plug-and-play. Vray has good integration and lots of documentation support too. They all have something to like. And as we all remarked walking around SIGGRAPH, “There are too many renderers.” So it’s hard to choose just one to single out.

  10. I’ve read some nice things about vRay as well, now they have vRayRT. You’re right tough choices to make if your in the arena to buy a render engine.

  11. David do you personally think that AMD has made progress with the W9100 graphics card, compared to a Geforce Titan Black, which card would do you like between the two ?

  12. I was informed, ignore AMD cards, now your the third to sorta mention the same thing !

  13. Hey David , great posts lately and I’m hoping that Maya will be released soon.
    Two quick questions that I kinda know the answers I think.
    Mila Shaders will not be supported on the Viewport for 2015 just yet…right?
    And color swatches are still not gamma corrected…..still will have to be done by hand with gamma nodes.

    • Color managing improvement is ongoing. And viewport display may or may not. I haven’t actually seen the “gold” version. However, getting improvements will be much faster and consistent than before. There’s a good team for rendering on both the Autodesk and NVIDIA side for 2015.

  14. Hi David, Maya 2015 is out now and from the presentation I took that nVidia was going to make scripts/string options available to expose FG on the GPU and such. Any idea where I can get those? Thanks

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: