But im wondering the following is does scene building profit more from multicore or clockspeed. and how do lanes join in on this story
PROCESS
as far as i understand what happens at Scene building is the following.
HDD + RAM (POLYS + TEX + OBJECT TAGS (motion blur, sub-d, hairs))
-->1. CPU CALCULATES -->
-->2. SENDS TO GPU's over PCI-E
-->3. GPU RECEIVES
-->4. GPU CALCULATES
-->5. BACK IN RAM (or hdd??)
What im trying to figure out is, step 1 and 2. What would be the best setup to run 4 cards. A 6850k Hexa (@4Ghz) with 40 lanes? Or a higher clocked 6700k quad 28lanes (@5.2Ghz e.g.) ?
Because in regards to calculation, it should not matter. As the 40 lanes are not used to their full bandwith.
However with the loading scene, you would say the bottleneck is in ACTUALLY calculating.
TRANSFER SPEED
CPU 40 lanes 4x GPU all GPU's will operate @ 8x lanes - or possibly 1*16x + 3*8x
CPU 28 lanes 4x GPU all GPU's will operate @ 4x lanes - or possibly 2*8x + 2*4x (OR can a 28 lane CPU not do 4 GPU's?)
Bandwith of PCI-E 3.0 == 15.75 GB/s (×16) == 0.98GB/s per lane
Taking GTX 1070 or 1080's with 8GB Vram (if thats used entirely)
So with a 40 lane cpu where all cards work in 8x the transfer should be 1sec.
While with a 28 lane cpu where the slowest cards work in 4x the transfer should be 2sec.
Which is nothing compared to the building times i guess. Because that sometimes takes up to minutes with very big scenes. (currently on 5820k)
So what i wonder about, does scene building profit from multicore in the same way scaling in multithreaded rendering does (e.g. cinebench) or is it more like programs like AE where clock speed counts more.
If you have a theory, or tested this, or are an octane developer who just knows
i would love to hear more about this! Also curious how people with the 7x gpu setups experience the loading of a scene. Thanks for listening
