GPU Rendering. Again.

The Month of Not-the-GPU

If GPU rendering has you confused, don't worry. You're not alone, and anyway, it's just not important yet. This is all getting talked a bit to death and I've probably written more about it on CGArchitect than about any other subject.

Here's how it stands. There are two categories of GPU rendering implementations right now, both somewhat useful but neither a complete solution to most architectural renderers' needs. These are brute force raytracing and game engine type interactive output.

Executive Summary

Don't make plans to give up on traditional rendering. As it stands, the usefulness of GPU solutions is quite limited.

Brute Force

First, brute force. This includes Vray RT-GPU, iray, Octane, Arion and probably others. Here's what I wrote in a recent CGArchitect thread on the subject:

Here's the problem with relying on the GPU to do "unbiased" (as if a little bias is such a bad thing, whatever those [PR people] were saying on that hype video about [that new CUDA render project] that we all saw, and no complaining, you know it's true) rendering. You compare a CPU and a GPU, architecturally, the CPU handles fewer simultaneous operations which can be quite complex, while the GPU handles many simultaneous operations that are not at all complex. Factors working against the GPU are lack of complexity, and the limitations of the concept of many in computation. GPUs have more transistors than CPUs, but only by a factor of two or three. A lot of those GPU transistors are going toward managing all the protocol for that absurd number of simultaneous operations. So the number of transistors that can be actively used for doing tasks isn't very different.

What this all adds up to is that the assertion that a Geforce card has far more computing power than an i7 is absurd on its face. What it has is much more very specialized capacity for doing a pretty small number of things.

Which is why it can, by hook (just monte carlo brute force it, you've got freakin' 800 threads) or by crook (just put the damnable model through DirectX already like we should have been doing years ago) either very quickly produce an image that you can't really use in a high end presentation, or very slowly produce one that you can. Which you've been able to do for years.

So don't put all your eggs in the GPU basket. Somebody's going to have to make some kind of breakthrough before they'll be useful for the same tasks as CPUs, and there is no way to tell when or if that will happen.

What I mean by all of this is that brute force GPU rendering has limitations that keep it from being as good as we'd hope for production rendering work. In practice, a skilled user of Vray, mental, Final Render, Cinema4d, etc., will make a better result that renders faster than a user of a brute force GPU renderer. If you've been hoping for a killer app, this is not it.

Brad Peebler from Luxology made a good video on this a few months back.

Game Engines

Direct3D, part of Microsoft's DirectX software, is a powerful 3D output engine built into Windows and modern video cards. There is a new generation of software emerging that brings in your models and puts them into a system based on the type of interactive 3D environment, using Direct3D, that most popular Windows video games use. This has the advantage that it uses the GPU for exactly the purpose for which it was designed, and actually requires no rendering at all. All 3D output is generated on the fly, allowing walkthroughs to be fully interactive.

This type of software includes Twinmotion 2 and Lumion. It runs on any good video card supporting DirectX. Current implementations will run on DirectX 10.1 hardware, but DirectX 11 hardware is better - a good new series Quadro, Geforce 4xx series, Radeon 5000 or 6000 series or current generation FirePro will be equally useful.

Unfortunately, there are a few downsides here as well. First, your clients will want to run the walkthrough on their own computers, which might not be sufficient, and you'll be forced to explain technical issues. Second, you don't get the full GI lighting and advanced rendering effects you're used to using. Third, interactive walkthroughs aren't run through Photoshop or video editing/compositing packages. No post - you need direct output that is presentation worthy. And fourth, the software is not well matured - Lumion is not in release as of this writing, and Twinmotion users are complaining of bugs, though this should all clear up in the near future.