All highend 3d software supports mental ray. Most comes bundeled with it.
Lightwave doesn't. And for the software that does, some people still prefer renderers other than mental ray- Vray, Brazil, etc., and lest we forget, Renderman. Studios may and do use these renderers interchangeably. Many of the 'high end' studios have their
own renderers (ILM, R&H, et al, and Pixar is using the 'next version' of Renderman)
100x faster is in doubt. Look at CPU vs. GPU-based encoding for video- you don't see a 100x improvment. GPU-based rendering does not render using the graphics card's engine, it renders using its own engine and uses the GPU to process instructions it is particularly suited for. Within a few years, rendering for today's
television visual effects may be done on the GPU, but film will likely remain GPU-
assisted for a bit longer. Remember that the lovely shots you see rendered real time in games are highly optimized- years worth of man-hours spent getting it to work within the engine.
A 'good'
television visual effects shot these days will use area lights, global illumination w/ ambient occlusion, traced shadows, and anti-aliasing set much higher than 16x AA yields, at very high polygon counts. (Usually millions for 'hero' objects, the
Galactica was around two million polygons.) Objects also tend to have shaders and textures in the color, diffuse, luminosity, specular, glossiness, reflection (if reflective), and bump channels. The texture maps themselves can be large and multilayered, 512 MB in and of themselves for a single model. And this is television, not feature films, and we're essentially discussing a hard surface model because I didn't even mention having to do things like sub-surface scattering to simulate skin, or the amount of polygons a good hair/fur simulation adds to the scene. Add on top of that the possibility of volumetrics and/or volumetric lighting in the scene, and the necessity of the objects to reflect that.
As for the amount of memory, yes, more is better. But if you're designing scenes on a machine with only 4GB, and have done full-up test renders on that machine, then you're likely to only need 4 GB on your nodes as well. But the idea was to present a budgetary and system baseline, and the article states as such under the assumption that you might know how much memory your scenes use while rendering.
But it boils down to making your own decisions.
Do you want a few render nodes to increase your productivity? (i.e. ability to get work done...)
Will it be worth the $140 per node to use the operating system you already know and use on your workstation, or is it cost-effective for you to learn how to administer a set of Linux nodes, including the a render controller that runs under Linux? (Note, at the low end of animator pay, a day and a half spent learning Linux is the cost of putting Windows on two nodes...)
Do
all the renderers
and plugins to that renderer that you use run under Linux? Will it cost you extra to get the version that runs under Linux? Will it cost you more per system than it would have to just put windows on the machine?
I could very well have spent an extra page discussing the pros and cons of putting Linux on your render nodes. At this point, i basically have.