geo_n wrote:With cpu renderer, a ram module adds to your total memory and if you have 6 ram module of 4gb each then that totals to 24gb of memory which you can use for a scene.
Look to GPU
(Your graphic card) as a PC
(that has "CPU" - chip, "RAM" - vRAM, comunicating interface inside..),
in theory You're multi GPU rig is somethink like a small render farm
(as each GPU is working individually).
(for the record, new interface is coming that will enable GPUs to share memory, but for that You should wait few Years, few generations at least. if You're interested about this topic, look the GTC 2014 event where they present & explain the technology =)
Now answer a quaestion:
which CPU engine let's You to add RAM from several machines? In my knowledge: each scene should fit to every PC/node (that means into every RAM) & then CPU on that machine comunicates with RAM (on that machine), taking info from it to solve the puzzle. That's how it is..
(The same works Intel's Xeon Phi cards for HPC. They are basically interpreted as 'computers having 60+cors' inside a host computer.)
Now about scaling and all the licence issues, consider this comparison:
* if You want to get the best enthusiast level CPU it would cost You 1000$ (six core),
* if You want double the power there are two routes - get 2 systems (waste money, for two sets of RAM, two mobos, two PSUs,..) or buy dual Xeons system
* if You want even more power You can buy four builds or 4 socket system..
- the problem here is that extreme enthusias CPU will cost You 1000$, while top of the line Xeon (for dual socket) 2500+$, for quad socket even more..somewhere 3500$..so why am I writing this? Just to prove the fact, that scaling GPU power is way more easy (& GPU vendors are not charging You like Intel does). You can get way more value compared to CPU scaling - that's considerably more restrictive & surprise surprise =) noONE talks about that,
'cos that's the way it is..& You have to get used to it!Before You start talking about free licences better build a system with at least 4GPU
then start thinking even about bigger, server based system..
You can easily get 7-8PCIe card, lower-end xeon some risers & have a 7-8 GPUmonster..
(without any bottlenecks (1x), server boards have 40lanes from cpu
..pluss, PLX chips that will be able to connect every card to at least 8x =)
I do understand some arguments..but we already have a very good deal scaling GPUs.