I currently have a GTX 1080 that has been fine for most things but I have seen some issues with larger files in other 3d software and on really large RC constructions the speed will slow down a lot. I have read on here a bit and have seen a bit of varying information on the best set up when it comes to GPU.
What is best?
Keep GTX1080 - Add second GTX1080
Keep GTX1080 - Add GTX1080 Ti
Keep GTX1080 - Add Quadro P5000
Run just a P5000
Remove GTX1080 and replace with 2x GTX 1080 Ti
Buy a P6000 (i will… but i hope this is not the best option)
I have seen mixed things about SLI in RC. I know in a lot of computing aspects and in gaming a single GPU that is more powerful will be beneficial than SLI, but this is not true in most computing aspects. However does having matching cards matter? The biggest thing to look for here is the largest amount of cores and onboard RAM? Very excited to get something more powerful, thanks in advance for any input / guidance.
Looking forward to answers here. Only thing I can contribute is I think I saw Wishgranter say is that RC ignores SLI but that adding ‘any’ card(s) is good.
Hi
SLI is ONLY for games, nothing to do with COMPUTE and so is advised to DISABLE the SLI ( as it can cause issues )
RC can work with 1 GB VRAM so there is no need to have card with highest GPU VRAM, only if want inspect-see them in 3D view. just large VRAM will NOT speedup the calculations
Its better to have 2x 1080(Ti) as single P6000
Just add more gtx1080 or gtx1080ti is your best gpu/$ IMHO. But really I wouldn’t invest heavily in GPU - the CUDA compute options are changing too rapidly!
Our pascal systems (gtx10, titanXp, P100) are about to be replaced with another generation (Volta - V100s are becoming available but the machine learning crowds are pre-ordering. DGX Station… zoom zoom!)
Maybe first check your various processing times…
All my (admittedly small) projects finish their GPU processing component very quickly and then spend quite a lot more time soaking up the CPU time (2xgtx1080 and 2xE5-2623… I find the CPU portions of a straight through reconstruction pass take about twice as long as the GPU compute phases. It does use all 16 threads really well but for balance I’d say pay for faster cpu for the RC once you have 2-3 gtx1080s)
Don’t skimp on the cache disk(s) either - quite a lot of traffic to the cache files even when there is still lots of free ram.
All about balance but there will always be a bottleneck somewhere…
Happy shopping
Jen
Thank you all for the responses! I had not seen my cpu hitting any max numbers during any renders (40 million polygon+) but I will look more in to that as it sounds like I may be seeing some things slowing there. I had not realized there were aspects that were so CPU intensive. Without diverting this thread too much does RC make more use of a higher clock speed or is it better to have more cores/threads? (I am assuming more cores)
From more reading it seems that the P5000 is essentially a 1080 (Ti) with a good bit more RAM which along with price makes me lean more toward the 1080 TI. I am all the way down to wait for Volta but to a point, depends on how long.
My main question at this point I guess is can anyone explain to my why having a second GPU is more beneficial in 3D work than a single GPU of “equivalent” power? I have always understood/been under the impression that a single stronger card will outweigh the performance of multiple. Im running in to issues with not having enough VRAM, would adding a 1080 TI effectively put me at a total of 19GB of memory? I thought that only was the case in an SLI format? I appreciate all of the help on these somewhat basic questions, just really wanting to keep expanding here (new lidar scanner arriving tuesday to incoporate in to workflow) and I am wanting to make sure my hardware is not too much of a bottleneck for what I am trying to accomplish (which is all relatively new to me still)!
The P5000 is a bit different than a gtx1080, mostly about the way the internal software splits up the hardware.
I think part of your confusion is thinking of the GPU as something you connect to a monitor and draw pictures with.
You can do that, but what RC and other programs are doing is actually using it as a little super fast compute appliance.
Dump in Data and programs, hit go, and it dumps out results! This uses the CUDA cores/programs to run a program.
Generally a CUDA program running in one GPU can’t access the data/memory of any other GPUs in the system. Generally.
SLI (Scan Line Interleaving) was about loading the scene into several GPU and splitting up the rendering work. Still no data sharing, just both doing pretty much the same thing with the same data (scene).
For CUDA processing, you just have to think about the CORES (how many, what speed) and how much space they have for the data and programs to feed the cores. If you have a fast and a slow card in the system, they will just do parts of the job at whatever rate they can handle. If you have a new and and old card, it may be harder since each might support different CUDA instructions
(Nvidia calls this “compute capability” and the pascal chips are CUDA 6.1 - see https://developer.nvidia.com/cuda-gpus for comparisons)
So no, having a second GPU doesn’t act like a single larger pool of memory, but you can run multiple instances of the accelerated code, each having more space to play in)
As for CPU cores - that comes down to compromises. More cores at a given clock rate mean more heat and more expensive chips. So you can drop cores and go fast, or add cores and go slow, or add more fast cores and go hot or add a second CPU and then go back and think about how fast, how hot, how expensive you want to go. There is always someone willing to sell you faster 
Same goes for things like the disks - more disks, faster disks, more faster disks, trade speed,size, connections, reliability, cost.
As I mentioned before - best to check all the steps in the process and see where you could get the largest benefits for the available investment. GPU got a huge boost driven by the gaming community and now again for the Machine learning crowd so you are able to get a LOT of compute/$, Disks got a great boost with the NVMe drives… but a lot of systems can’t realize the full potential of those drives yet (bottleneck in the PCH bandwidth). x86 CPUS unfortunately haven’t changed much which is why the serious processing stuff is being done on the GPUs now.
Hope that helps…
Jen
Speed queen…
Thanks Jen! That makes sense about the software and the drivers running differently (price tags are different enough) and I believe I saw that they were using a higher grade silicon in the Quadros. I thought that the 1080 and the P5000 were both on the same GP104 with 2560 cores (what I meant by saying they were essentially the same since the cores are what im learning matters most).
Speed is an issue here (somewhat) however my main thing to overcome here is an amount of vram, my inability to render full scenes as a mesh due to their size is less of an annoyance like the speed aspect and more of a necessity. However after all of these answers it sounds like I should either throw another 1080 in here for now to solve the problem… or take the 1080 out, throw it in the unraid box for a vm and get a TI or two :?
CPU wise I am currently on a Ryzen 1800x and am planning to hold off for the Threadripper when they get all, or at least most of its growing pains sorted. They seemed like a good balance of cores/threads and clock speed (plus quad channel memory and more pci-e, etc.) without my bank account hating me too much. I was just unaware on how the optimization of RC worked with lets say a choice between a 8/16 @ 3.8 GHZ and a 16/32 @ 3.4 GHZ would there be any benefit to the larger chip or would there be more benefit with the faster chip? I guess the main question there is does RC take advantage of any and all available cores or is it optimized to stop at a certain number? thanks again to everyone for all the answers and guidance!
Gavin Calaway from wrote:
… lets say a choice between a 8/16 @ 3.8 GHZ and a 16/32 @ 3.4 GHZ would there be any benefit to the larger chip or would there be more benefit with the faster chip? I guess the main question there is does RC take advantage of any and all available cores or is it optimized to stop at a certain number? thanks again to everyone for all the answers and guidance!
I can’t say over all - but I know that in the main cpu bound phases, RC keeps all my cores (2 x 4/8) busy. Given your choice I’d suggest the 16/32@3.4ghz will give you much better throughput. (Of course it is possible RC tops out at 16 threads in which case only your real cores would be busy but that’s not all bad either.)
One big advantage in the new chipsets is access to more options for cpu attached PCIe lanes. This means you can make up a 4 drive NVMe raid0 and have it not bottlenecking at the PCH. I wouldn’t want to trust booting off it for a while - but great for a scratch disk!
Jen
Hello guys!
i have a suggestion that may help evryone ^^ : If we publish a set of images and everyone run it on their configuration, we could have after a time quite some answers on those hardware aspects 
my current hardware: i7 6800k oc at 3.8ghz - ssd 960 evo 500go - 2 nvidia 1070 - 32gb ram running at 3ghz
Count me in - I can cover the lower end…
i5 4460 @ 3.2 GHz - 1 GTX 750 @ 1 GHz 4GB - 850 EVO - 32 GB Ram @ 3.2 GHz