hey man.
scene setup is only one side of the coin, and certainly for working on the scene it's best to have a powerhorse. but @ rendertime, pure raw cpu power counts. and maxwell needs A LOT (maybe you also want to work on your computer while rendering...). check out the specifications of the renderer. it basically simulates a full camera setup (every shot has a shutter time set, film material, f-stop, ... all as a real camera), all based on physical laws and optics. and it calculates all in 32bit (.hdr) range. the other important point is that it renders the full frame without any buckets. so the quality of the rendering simply increases over time. the quality itself is measured in SLs (sampling levels) with exponential timesteps from SL to SL. so basically you can render forever on an image.
now. a standard production image in architectural visualization as i've done a lot takes from 4 to 15 hours to render on an overclocked quad today, but i've seen images rendered on multiple nodes over a period of 100 hours to clear up noise. for example images with chromatic aberration, caustics, or subtle diffuse lighting need a long time to render. and as the field of architecture is... a lot of different versions are needed sometimes in a very short time (six images in an afternoon... i've had that.)
certainly one could just pick a different renderer with a biased approach (photon mapping like mental ray, vray, brazil, ...) which render the image itself a lot quicker by sort of faking the lighting. but the actual lighting setup takes MUCH longer (i worked on the lighting setup for an office interior for a full day. just tweaking parameters in mental ray !!). maxwell ? just place light emitters or the sun and you're done. there are no settings for nothing except for the camera options. this is why i love that renderer. same quality every time. no tweaking !
also, the problem with renderfarms is that most of the time they're running at low usage and sometimes, the peaks seem to explode where one wishes to have access to blueGene. (
http://www.top500.org/ )
so the bang-for-buck must be worth more than in a gangbang. money, space and technical knowledge also are limiting factors in the broad field of architecture.
my dad also says that he'd rather pay a lot more for one huge computer (let's say 4xquad xeon under one hood), but this is a waste of money based on raw cpu power that is accessible with lo-price slaves (a rule of thumb: 20% more power costs the double amount of money). of course, there are a lot of nice labs out there that sell very sweet rigs, but at what prices ?? see for example
http://www.appro.com/
this is also why i started this thread here, to find out what can be done againts horrific prices:
http://forums.anandtech.com/me...id=27&threadid=2187473
the helmer computer mentioned at this thread there has six nodes and costs around 3500$ ... a great price for that amount of power !
so the main question exactly was... how should those machines be built ? what are the rules behind looking out for possible hardware ?
blades ? off-shelf computers ? self-made lego computers ? single components just mounted to a wooden frame with some metal around ?
difficult... i know !
Facebook
Twitter