Gadgetory

while some people use all of their GPU stockpiles for mining a few dollars a day we have real work we need to do with them and that is rendering our new intro animation so this animation was redesigned by Andrew for the site it's currently rendering on this machine this is our main camere and blender machine for production workloads you often hear us talking about this is a production card or a production CPU well this is a use case where that's actually at hello so this is where the real use case is for those production products and for today we're focusing on GPUs or using a lot of CUDA power to do the rendering it's taking about 50 minutes per frame to render you know 120 frames get through that's 120 hours of rendering and that's for a 2 second animation 60fps 2 seconds so pretty long time this is on a Titan X I'm a Maxwell GPU it's not good enough to right now so what we're going to do is build a system that's probably going to use three to five GPUs around what fits for PCIe lanes and what fits in the motherboard and all that and once we have that system built we can run the renders again on CUDA it does not need sli bridges it just needs the video cards and then see what our pickup is for render time with a real production system so this serves two tasks one it gives you insight as to what everyone means when they say this is a production GPU or this is a production CPU and it also lets you see the behind the scenes for how long something like this actually takes to render and so the reason for updating our animation blender actually updated its shaders so we're using new shaders with the renders and we also have some texturing applied so this is a this is a non final version of the animation we have a couple of tweaks to it but you can see some of the texturing and new shaders in play and also it'll be in 4k so that will match our new camera everything will be in 4k 60fps and that's the point of doing this before getting to those that this coverage is brought to you by EVGA and their 1080p is c2 which we've recommended fairly highly for its build quality and the icx which are kind of fun to play with you can check our full fc2 review for the 1080i if you're curious to learn more or you can click the link in the description below to find the product page for the 1080i sc2 so I'm gonna grab this Quadro this just came in and in fact I've study so this just came in this is I think a P 4000 we're using it for some benchmarking on basically what is a non gaming card but we're gonna use it first on rendering stuff the reason I'm grabbing this is not because the most powerful thing we have for rendering because the Quadro actually this one is not our 1080i is in our Titan XP will probably render faster they have more cores higher frequency but this is one slot so we'll be able to fit an extra card in by using a single slot GPU that's still relatively high power should be about 10 70 class for rendering the reason it's expensive is because of the drivers certification but that's our first one and then we're going to go to our 1080p I shelf to the rest so all these cards are GTX 1080p eyes that we've reviewed plus there's one more over here that's in pieces type are the Lightning card but we're going to grab I have a hybrid in the other room already for rendering we're gonna grab an FCW three because two slots and it's going to be hopefully cool enough so we have what we need let's go build that system okay so this is our X 299 production bench one of the ones we've been using for Vega production tests and I've got it all hooked up right now to our current clamp for testing that I'm doing for build Zoid separately but we'll talk about that later and we also have a plug load meter down here which will allow us to see what sort of load this thing generates with all these GPUs to make sure it's within the specs of the power supply so what I'm going to do is mount these into an i-9 systems that has the lanes I need this is why people complain about Lane count by the way and why thread Ripper will be interesting for this application what I'm thinking it is Quadro here I'm probably I'll hook up these bands so that the Titan axe can breathe Quadro here and then I'm thinking we might be able to squeeze in a hybrid there that'll take those two slots and then maybe an FTW three there so let's try that first oh yeah right there okay six pin we got that will be six pin power for this one okay that Titans going to stop for a little bit but I'll work it out okay hybrid that was hiding in the corner that's going to be a six and an eight so we're still good for that please fit wonder if it's better to put thermals are the main concern here so I'm trying to think through it so this is a blower fan of this on the side it's got a big back plate on it could pull a minor and run like a riser cable out of there if I'm really worried about thermals but this thing should be pretty low temperature anyway you know six pin means it's going to be low power draw to 4,000 is not the highest power dropping on video makes so it's going to generate lower temperatures so you know we're just we're going to run with this we'll check all the temperatures once I get a light load on there and decide if I need to make any modifications anything or hybrid cool anything and I'm going to actually space this a little bit somehow let's maybe we will end up running a riser let's let's get this in here to let socketed yes okay one two three four that's the value of a single slide GPU even though it's lower power could run a riser but we're going to try that a lot of power cables now we'll check thermals and power drawn next before going too crazy with the render this is going to be a little problematic I think so take the shortest one okay those will reach the head so those are candy and IMAX really needs to make the cables longer this is always going to problem I've had with enter max otherwise it's a 1350 watt 80 plus Platinum power supply so it's great except the cables are too short okay so that's good two more seriously this would not be a problem if the cables were standard length all right there we go okay okay so then this can go over here that should be will it turn on oh okay all right we're on so next thing to do is plug in all of these fans I didn't think we would need them but that clearance is just so tight that let's go ahead and do it for safety there's one okay should have good airflow okay all right so all the fans on there's some wine in here if you can even hear that I think that's from the Titan XP I don't know why there's so much coil line on that particular card that's incorrect all right so we got to redo the drivers it looks like cool man just shutting the thing down it's at 300 Watts right now that's the i9 CPU for UI 9 + 4 GPUs and Idol almost 200 watts other installations please work this time that president we've gotten before finally drivers are working a little too well I haven't restarted yet I need to but we've got 1080i 1080i Titan XP and quadra they're all detecting QP 104 TP 102 to be 102 keep the 102 they're all detecting correctly see if it has thermals for any of these I probably need to restart for those I guess ok so good start so I rebooted and got GPZ open it is detecting everything now so you can see Titan XP is this is all idle we're at 36 see we'll see how it does when it's actually under load that's probably the hybrid card it's at 27 see this is probably the FTW three card at 35 see the fans aren't even spinning and here's the P 4038 C so what we're going to do is kind of monitor these and keep an eye on those during the initial render testing and then once it looks like it's safe thermally we'll go ahead and let it rip and I might also interestingly open up a separate instance of blender and use some of the CPU power so one thing we could do to make use of the i3 Seaforth I three CPU the high nine CPU would be we could take two to four of the threads from it and let blender use the other 16 or 18 thread keep a couple of them open just to make sure that as GPUs are still fed properly and then the remaining threads can render and we previously found in our eye 9 review that the with all 20 threads available it was out rendering CUDA so that would be pretty cool to actually be able to use that capability plus the GPUs okay there we go Titan xpp 4000 hybrid FTW three so open the render file for those of you don't know what we're talking about when we say production workloads or production GP or CPU this is what we're talking about it's during something like an animation so we get user preferences computer compute device for cycles it's going to be CUDA we select all of them that's pretty crazy actually see all those so we're going to use all four GPUs save settings and then we can open a separate instance for the CPU if we want over here we've got image editor is the display but the important parts are feature set experimental are supported we're going to try experimental device GPU compute rather than CPU we're focusing on GPUs today and then down at the bottom I'm going to make sure it's optimized for at 256 by 256 tile sizes this should be less relevant going forward but as I understand it today that will speed things up for us and then for sample count or actually for resolution we're at four we should be at 4k 3840 by 2160 so 3840 by 2160 at a hundred percent that's 4k then we have sample count we're at 500 which is production-ready sampling and then we've also got a thread count for the cpu and here so you can actually see that it's got 20 threads right here from the cpu we could fix this and tell it to do like 19 or something for now I'm gonna leave it on auto tech so this is our GP render window anyway so think we're good to go let's let's start with just one frame and see how it does and then we'll try more from there that's pretty crazy nice I used to see in that like if you don't do any kind of rendering and this probably doesn't mean as much to you but to see four tiles at Tuesday six by three to six each going at the same time is pretty exciting so we get a CPU going to and that would be twenty more tiles or whatever we assign to it at a at probably sixteen by sixteen for optimal speed on a CPU but on a GPU we can get away with this let's keep an eye on em per atures are doing so far Titan X P 58 c P 4060 C not bad really hybrid is no problem at all hybrid the hybrid sandwiched between everything doesn't even care at 35 as the w3 is a 53 and it's fans are still off by the way so we're doing okay here so far it's a power consumption from the wall is 700 plus watts right now and if there's anything interested in here for you so our maximum power consumption so far 755 minimum 100 and then we've got bolt amp readings but yeah so we're at 600 700 plus and this is going crazy fast 780 equals x percent of 1550 so we're at 57 percent utilization on the power supply right now it can take a bit more and we've got fan speed zero decent on all these things 71 see on a Titan 74 on the P 4000 let's help that one out a little more okay so we're going to make this better I'm going to get some more fans on here all right 70 and 72 for those first two DPS right now that'll feed air to the blower fans which will then push it out the hole that's warm oh crap push it out the back what's the I don't have a thermocouple nice to me so let's just take a spot reading of like the furniture behind it 35 see for the shelf behind the video cards which is not meant to be scientific it's just meant to be like man that's pretty warm considering we're at 35 C on the shelf and room ambience is a about 23:36 C on the Shelf now full 37 it's going up so that means this is pretty cool that means the fans actually helping get more air in there so we can force it out the back 38 on the park near the Titan XP all right so we are at pretty HD just on furniture behind me is saying the GPUs are now steady at 68 for the Titan XPS are at 68 here and then 70 for the P 4043 67 and this just looks like it's still rendering along yes it's still going so good there our frequency is dropped a bit on the Titan XP I think we can pick that up temperatures dropping a lot Titan XP is down to 60 now so that's going to free up power let's give it some nice drama core to 670 43 66 everyone's pretty happy there's our memory consumption so here's a if anyone's curious memory consumption on these things platon XP is at two point three gigabytes use our animation just let them use upon a memory we're at 1.3 over here - and - it is using I mean it's using a pretty good amount it's at least 2 gigabytes per tile is what we're looking at which honestly I think we might be able to step up the tile size if we wanted but that's doing ok for now and that fan really helped out the temperature is our Titan XP and our slightly overclocked to that 53 P 4068 instead of 70 that's pretty cool just getting a fan up that really improved in significantly we're almost done with the first frame exams six amps 12 volts okay is that it that's it rendered time 14 minutes 22 seconds so we have gone down from 50 oh we have saved ourselves well let's see over the if you're at roughly 50 to 60 technically so let's just split it and call it 55 355 420 frames previously that would be 6600 minutes or if you prefer 110 hours versus now if we let's just make some assumptions here and assume that it is going to be fairly linear let's go ahead and call that just let's call it a with that 22 let's call it fourteen point two five and multiply that by 120 frame divided by seconds in two minutes era in two hours and it goes to twenty-eight point five hours so twenty-eight point five versus 110 hours it's our difference power draws don't really care about right now because it's more important to me frankly to get the render done and not use our primary production machine which here's the thing okay so even though it says 110 hours for the first one it's not actually 110 hours because what's going to happen is we start the render and then it has to stop like every 12 hours so that we can use the machines of do videos to render videos edit videos can't run blender during that time so it's not 110 hours it's more like probably a couple weeks and this going down to twenty-something hours less than thirty means I can set it up run it walk away from it come back in the day and it'll be done now here's the thing I guess well first of all we need to do some small Corrections on this I guess a little blown out but that's still pretty representative of the time required now the next thing we're going to end up doing is running a cpu render at the same time and let's just give it 18 threads and what's best here sixteen by sixteen sixteen by sixteen Andrew tells me for faster rendering with CPU so let's just get a rough estimate of this one so CPU render oh well there we go eighteen threads ripping through it see this is where the sixteen core thirty-two thread thread referred GPUs will be really interesting is in an application like this so again this is what a production machine looks like this is what people mean when they say production task is cutting off four times the time required to render something that's huge and why I'm also interested in the future Intel 18 core chips and the future and the sixteen core chips but anyway so right now we're going on CPU I 970 nine hundred X what's our power draw here 320 Watts set up quite a bit from idle the GPS are doing nothing right now 14 amps by 12 volt 168 Watts going to the CPU and the memory channels we're definitely against the TDP window there but so what we're going to do now is walk away from this let it finish it looks like it blender thinks it'll take 45 minutes but it's going down rapidly so we're going to walk away from it CL on it takes and come back and then decide if it's best to render on the CPU and on the GPU simultaneously if it is we'll let that go and then I hope that cuts to sound like 20 hours that would be great alright so we came back after 47 minutes and it's still going so despite the 300 X actually being okay and our blender benchmarks it's struggling here and part of that is a newer version of blender where the GPS are just being utilized better so this it looks like we'll take another 20 plus minutes to finish which puts us at over an hour for 18 threads so CUDA is just blowing the sin away OpenCL I would do despite some of the blenders previous issues and OpenCL it's gotten better so basically if we let this thing run for the entire 20 to 30 hours that the GPUs are running it will out it'll save us like they'll save us a couple frames ultimately if you look at the difference from either end it's the difference of like maybe a few hours natan so what we're going to do is not burn an extra 200 to 300 watts on the CPU and the memory and everything else and just run the GPS only so it'll be 600 to 700 watt draw and CPU won't be in use other than to feed the frames and then we should be able to finish the thin in 28 and a half to 30 hours and we'll have a new intro animation for you which I don't think will make it in this video because it actually does need to be rendered but keep an eye out because it'll be a probably silent change over to the new one and those of you who saw this video will know where it came from so that's it for this one hopefully this gave you some insight as to what actually happens with a production setup there are plenty of other production applications this is just one of them but it gives you an idea of why you would use one card versus another or why it actually matters to save minutes off of rendering in this case 30 plus 40 minutes yeah we're down to 14 minutes from 50 to 60 so if you're saving 40 minutes on a render again the real-world example that we've given you here today is a difference of weeks of rendering on one machine or straightforward just 110 hours 120 hours 230 enough time to let you render something else or increase your quality settings or whatever it's pretty cool stuff you can subscribe for more as always patreon.com slash gamers and access valves out directly and we haven't officially announced yet but there's a new shirt you can click the link of the description below it'll go straight to the new shirt or if you go to store documents next on that or gamers Nexus squarespace.com you'll get to the shirt page and you can pick up a new anniversary edition shirt which we'll talk about more in a separate episode thank you as always I'll see you next time