First off, a question: does GPU acceleration merely speed up the rendering process, or does it actually increase Lux's ability to handle "heavy" scenes with lots of figures?  Because my bottleneck with rendering in Lux has always been "it will choke on too much geometry and you have to guess how much is too much."  The memory conservative option helps a bit, but ultimately it comes down to me running this stuff on a 3 and a half year old midrange computer with 32-bit OS and 3.5G RAM, and having a thing for Optitex's high-poly dynamic clothing.

Second off: I remember looking into this matter about a year, year and a half ago and getting the impression that my ATI Radeon 3870 was not one of the supported GPUs (or one of the GPUs that had not received driver updates to support this feature).  Does anyone know if that is still the case?

Your biggest issue is going to be your 32 bit system more than which mode luxrender is running in as it severly limits how much memory you have in use.

from my understanding the GPU acceleration is still in beta form and not a reliable option for luxrender yet. I run a high end 64 bit system with 12 gig DDR3 and this feature slows down loading and render times for me.

Yeah, I suspected that was the case, but it's good to have it confirmed.  Thanks!

Gpu acceleration is exactly that, it accelerates the render. It actually requires more memory than CPU-only rendering so you are better of, with only 2GB per process, by avoiding GPU acceleration. To save memory you can use the texture collection option, that can dramatically help.

Cheers 

Thank you, Paolo!  I'd played with the texture collection feature once or twice without being quite sure what good it would do, and it did seem to help.

I'm an enthusiastic user of Reality/Luxrender—in fact I use nothing else. I have a powerful graphics workstation so graphics capabilities and memory are not an issue. Nevertheless I will not consider using GPU acceleration because of its limitations. In fact it can create more noise than is compensated by its efficiency, and there are a number of other issues, some obvious (e.g. limitation of GPU memory), some less so. The best way to increase efficiency is careful use of lighting. For example, use only mesh lights (and I rarely need more than two of these) and enclose your scene in a cube primitive to reduce wasted sampling. If you observe such techniques renders can be remarkably efficient without GPU contributions.

Having a 32 bit system certainly limits your capabilities. I had such a system until quite recently. I had to become resigned to being able to render limited scenes. I didn't find GPU rendering (although I had a decent card) much help with this.

Oh, is that what the stick your scene inside a cube thing is about?  Cool!  I will have to try that.

Thanks for sharing your experiences with GPU rendering on a 32-bit system.

It helps to hide anything that's not visible. A leg in a pair of pants, a shoulder unnder a shirt, feet in shoes, if the mouth is closed hide the teeth and gums, etc. The less geometry exported the smaller the file, the less RAM needed to complete the render.

and if you use windows 32 bit with more than 3GB RAM you should turn on the 3GB swtich to allow a single app to use 3GB RAM instead of 2GB, but you should use LuxRender without Studio running if you use this option.

Yes, indeed. Hide everything in your scene tab that is not in the camera frame, unless it is required to create shadows. Anything hidden in this way will not be sent to Luxrender and will reduce the memory requirement. Note, though, that although making surfaces null materials in Reality will save wasted sampling, the mesh is still exported, so it does not reduce memory requirements.

Although Reality does a good job in reducing such things when it converts materials, keep an eye on diffuse colour to make sure it is not pure white (255, 255, 255) as this is very wasteful, and unnecessary, and keep specularity very low, even on a fairly glossy surface. If you are using metals (which are fantastic) you might think about sacrificing some polish to avoid reflections on adjacent surfaces (say if it is jewelry next to skin). Typically with gold they look like red or orange speckles and never seem to clear up. You can also have this problem with non-metal texures with high specularity.

The cube trick really works. I use it by default, and although it cannot be used in wide open scenes (without postwork), I prefer tightly framed scenes in any case and use it for both indoor and outdoor scenes.

Yeah, I try to be careful about hiding stuff for images.  The diffuse color thing and metallic issues are new to me, so I will keep them in mind.

Quote - The cube trick really works. I use it by default, and although it cannot be used in wide open scenes (without postwork), I prefer tightly framed scenes in any case and use it for both indoor and outdoor scenes.

Wouldn't this be the same as using a light sphere (or whatever they're called)? Like with UberEnvironment, where it encases everything in a giant globe.  Doesn't that work the same way?  Or is it too big?

That is Image Based Lighting (IBL) where the surrounding mesh is used to calculate the quality and intensity of light based on the image attached to it. Using a cube to enclose the scene is almost the opposite. It does not, of course, a source of light; it limits the calculations required in tracing the light from the mesh(es) that are the source of light inside it. You may also have parts of your scene model outside the cube, those that will not be seen, and calculations will not be wasted on those parts. Yes, the IBL sphere itself would be too large for that purpose.

Quote - That is Image Based Lighting (IBL) where the surrounding mesh is used to calculate the quality and intensity of light based on the image attached to it. Using a cube to enclose the scene is almost the opposite. It does not, of course, a source of light; it limits the calculations required in tracing the light from the mesh(es) that are the source of light inside it. You may also have parts of your scene model outside the cube, those that will not be seen, and calculations will not be wasted on those parts. Yes, the IBL sphere itself would be too large for that purpose.

I could be wrong, but I believe Paolo said at some point that the IBL sphere isn't actually exported to Lux. It's just kinda there to help line up your IBL image. It'd be easy to try and see, as if it does, it'll be in the material section. :-)

Just for reference, I use IBL's almost exclusively for all my pictures, as it does add one big, colored light if you use the image with it. It can also be run without an image in the slot. Doing it this way--for me--adds a light that doesn't bog down big render scenes nearly as much as a slew of mesh lights. Granted, you don't get all the pretty shadows like you do with a mesh light, but its fast! I pair this with setting the render to 'Exterior'.

One other trick I forgot to mention for speeding things up (a little, as in a kilosample or more, scene dependant) is nulling skin shaders that are also hidden by costumes, like the arms, legs and torso if inside a business suit, for example. If it doesn't have to calculate textures, or a lot of them, you'll definitely see a speed increase.

The cube enclosure does work, but it's best used for scenes with props that are out in the open (the theory of scene enclosure is that Lux uses 'photons' of light to bounce around your scene, then traces their intersection with obstacles, ala characters and props. When these are out in the open, the photons bounce out into space, rather than being enclosed, which lengthens render times.) If your charcter is already in a room, a cube isn't really needed. Or if your scene has an incomplete set of walls, like a corner piece, just add scaled flat plane primitives and set them to matte in Reality.

It's little tricks that add up in the end in the name of efficiency!

Yes, of course, you are right. The sphere isn't imported, unless you rename it, I presume, although the image itself can be rendered.

I use the cube mainly on open sets such as you describe. I have one made to a useful shape, size, material and location and saved as a prop. I can easily re-size its various dimensions to suit. I much prefer these semi-open to rooms as you have so little elbow room when setting up—in some cases there's hardly room to swing a mesh light. Yes, you can turn things on and off or make them transparent, but it ends up being very fiddly. It drives me nuts. On the other hand, I rarely do long or wide angle shots so partial sets suit me very well.