I do a lot of animations in Cycles and about to build a new workstation to render the animations faster, I want to get a 3 GPUs System(3 X GTX 1080 TI ) but I am afraid that the third PCI Slot of the motherboard will slow down the 3rd GPU and will not run as fast as the other 2, here are the specs, I posted in a lot of forums (MSI,AMD, NVIDIA) and I did not suitable answer , hope someone will help me with this build :
CPU : AMD Ryzen 7 1800x
MOTHERBOARD: MSIx370 xpower gaming titanium
GPU1 : GTX 1080 TI FE
GPU2 : GTX 1080 TI FE
GPU3 : GTX 1080 TI FE
RAM: Dominator Platinum Series 32GB (2 x 16GB) DDR4 DRAM
Since the 3rd slot is a PCIE x4, and run through the chipset, will it slow down compared to the other 2 GPUs?
Much appreciated, if someone has a better build or other advice please let me know, thank you so much.
Hi, don´t worry the mainboard switch to x8 x8 when both slots are used.
The lane speed does not affect render performance at all only loading the scene to VRam.
X4 is already very fast, several GB per second.
When loading the scene is finished all cards render with the same speed.
would it be better to get a 4 slot board so you can add another later?
intel out ranks amd for single thread performance, i7 7700K at ~2500 and the AMD at ~1900. unless you are cpu rendering, those extra cores arent going to help much. source: Passmark Benchmark Pages.
in the US, the i7 is closer 350 usd, vs the amd at around 450 usd according to passmark. a quick search agreed.
just make sure you get the baddest power supply on the shelf. gold rated and up, aiming for 150% or 200% of assumed wattage at full computer load. 50% power capacity is a sweet spot in efficiency. research OEMs, know who actually made the unit. Seasonic and Super Flower (yes, thats for real) are highly regarded.
I agree with Daedalus_MDW about the processor. i7 7700K is a very logical chioce. One more important thing - keep an eye on the physical space to fit the graphics cards.
The fact that the third PCIe slot is a different color implies it’s not the same spec as the other two.
I ran into this problem a few years ago when upgrading my video card. Cheaper motherboards, even though they may have more than one 16x slot, if you use two of them, they both slow to 8x, so you’ve got to read the specs very carefully. Also, read all the reviews you can find to make sure you’re getting what you think you’re getting.
A quick search and I came up with the ASRock Z270, a motherboard with four slots (there seems to be rare) and they all run at full x16 speed. You’ll want to dig deeper to make sure all slots run at full 16x speed when more than one is in use.
Since you’re planning to buy three GTX 1080x, I’m assuming an expensive motherboard isn’t going to break your budget. You may end up paying almost as much for the MB as you do for one of the video cards.
Right now, almost all the PCIe transfers that Cycles does happen at startup, so PCIe bandwidth does not really matter (4x should be plenty). If in the future Cycles starts streaming in data from main RAM during rendering, that could change.
The CPUs that run on this motherboard only have 16 lanes for the PCIe slots, so running more than one at full speed is not possible, even if the slots are all wired to 16x. Two will drop down to 8x, three or four will drop to 4x.
The Ryzen 1800x has 24 lanes, so in theory it should be possible to run at 8x/8x/8x with a suitable motherboard.
Either way, it usually doesn’t matter, because PCIe transfers are very slow (compared to GPU ram transfers) so programmers will avoid them if possible.
And if you wanna hear something really amusing, at the time I bought my dual-16x motherboard, I was using BI render. I had no idea ATT BI only used the CPU.
But thanks for the extra info here. I certainly don’t want to lead the OP astray.
How likely is that, do you think?
I ask because recent speed-ups in Cycles 2.80 are heading very quickly off the chart and I’m wondering if it’s already here… or perhaps the solution to adaptive displacement will be RAM > GPU transfers?
I don’t know how likely it is. There’s a patch on the tracker that lets Cycles spill out textures into main memory, that’s pretty easy to implement. It’s been a while though and it hasn’t been merged.
Pascal GPUs were announced with full virtual memory support, but I believe it turned out to be a TeslaP100-only feature. That would’ve made using system RAM simple, as well.
Then there’s the possibility of implementing a whole texture mipmap/cache system like in RedShift, which would be quite bit of work.
As for adaptive displacment or any other geometry: I believe it should stay in GPU memory at all times, for performance reasons.
But the i7 7700k have only 16 PCI-E lanes(Ryzen 24 lanes), do you think it can run 3 GPUs or even 4 GUPs? theoretically, it can x8/x4/x4 or even for 4 GPUs x4/x4/x4/x4 since PCI-E speed does not matter for cycles x4 is plenty enough? and what about SSD ? does it also take a chunk? I am afraid that the 7700k will bottleneck, what do you think it’s better, keeping the Ryzen 7 1800x(3.7GHZ,8cores,24 lanes) with 3 GPUs setup , or upgrading to 4gpu mobo + I7-5930K (3.5 GHZ,6 cores,40 lanes) both CPUs have the same price in my country , it is worth it to sacrifice little amount of core speed and count to add one more gtx 1080TI ?
as far as I know, the gains you will get from 3 gpus vs dual begin to diminish substantially… basically dual gpu’s are worth it… 3 or 4 are not as far as blender is concerned… Supposedly, each gpu will only offer half the performance increase of the previous number of gpus… so, if you have one and add another, it will cut your times approximately in half… if you add another, it will increase performance only 25% so the gains become less and less… Additionally, Blender will only utilize the memory of one card as I understand it… so, based on all that, I would probably recommend a dual card setup with two cards that have a vast amount of memory per card… and any system, AMD or intel can run a dual card setup at x8 for each card which is far more than enough at this point.
also, I have a 5930k setup with a quadro m4000 and it runs strong in blender… I also have an 1800x with a single gtx 1080, which does in fact outperform my intel setup substantially… but they are both great systems…
I got this weird idea about splitting the build into 2 micro atxs , one with the Ryzen+2 GPUs and other with i7-7700k+2gpus LOL ,idk, its the same price if not lower.
That’s how rendering with Cycles works - if you render image for n minutes and then for 2 x n minutes you get 50% less noise, if you add another n minutes (3 x n now), you lower the noise by one third and so on so you keep adding less with the same amount of time(or one more GPU). It doesn’t change much. 4 cards will render the same amount of samples twice as fast as two. It’s not like SLI for gaming, they should not be connected to each other and work independently. If you have 4 cards, PCIe connection works slower, so copying the scene to the memory of the cards might get slower, but that’s not a big issue.
Haha!!.. I mean, I guess you could… but honestly, the best thing to do is to make a plan… know your softwares and build accordingly… I think the truth of the matter is, that today, the biggest performance increase you are ever going to see is going to be on the software side of development… so basically, build a system that should technically tear through whatever program you throw at it and hope for the best, or take a closer look at what your softwares utilize and in which way and build a system optimized for that… then of course I guess you could also build a system based on a workflow you would like to practice,… based on how your hardware, software, coffee breaks and lunch dates are all aligned throughout the day… I don’t like waiting much for anything, but there isn’t any hardware out there yet that substantially takes the agony out of the Blender waiting game.
well… the 5930k is overclocked… and the 1800x at stock clocks… for whatever reason, the 1800x easily beat out the 5930k… both in rendering and in smoke simulation processing which seems to be a single threaded operation… and on the same blend, it was a substantial difference… so… I gotta give the points to the 1800x on that one… but, I would think that that would be expected anyway…
Okay I think am going to take Daedalus_MDW advice and go with the i7-7700k and 4 GTX 1080 TI with a Z270 4 GPU capable ATX, since the i7-7700k have 16 lanes it will be x4/x4/x4/x4 I think 4gbs is more than enough to load the scene to the VRAM so 1 or 2 sec delay for bigger scenes is not too much , I am only concerned if all lanes are used what about USBs and SSDs and M.2 and other devices do they also use those lanes ? I am spending an entire year savings hope I don’t regret this lol.this PC will be only usd to render cylces animations and editing those animation with adobe premiere , nothing else , I realy think I will need those higher core speed of the i7-7700k for faster physics and fluid simulations and also for buillding the scene to send it to the GPUs .please if someone thinks I am doing something not right it will be great to corerct me , hope I am on the right path.
thanks again
I think that system should serve you quite well… the only area i might worry about is core count… especially for work with other software outside of blender that you might be using for compositing sequences, rendering to any number of intermediate formats, along with any adjustments you might make in those programs… They all have thier own dependencies, some of them heavily CPU based… Otherwise, that might be the absolute sweetest PC you can put together!!!
