Gadgetory

it's time to revisit PCIe bandwidth testing we're looking at the point at which a GPU can exceed the bandwidth limitations of PCIe gen3 slots particularly when in by eight or by sixteen modes this comparison includes dual Titan V's testing and by eight and by sixteen configurations pushing the limits of the one gigabyte per second per Lane bandwidth restriction of the PCIe slots we bought the first Titan v then YouTube channel bit speed trip and said to us the second one and we're now able to spread more load between two PCIe slots than we could before before getting to those this coverage is brought to you by EVGA and their 10 ATT is c2 which we've recommended fairly highly for its build quality and the icx sensors which are kind of fun to play with you can check our full sc2 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 the last time we did this test was on 1080 ice and we only did a single card with a single card in one slot between x16 and x8 we saw no more than generally a 1% difference between the two not that exciting basically the same we now have Titan V's which are considerably more powerful in certain applications that can leverage them and we also have to Titan V's so we can really only leverage this in one application right now and that's ashes of the singularity as far as games go now there are other applications like GPU PI where you can leverage them in tandem but it's not quite the same workload PCIe bandwidth consumption is going to very heavily on what's being used to test it for example the folks who do GPU mining often extend their cards via a riser cable to a by one or by four slot generally by one if you're doing that you can tell already that that particular mining application doesn't really care about the bandwidth of the PCIe bus so it will depend - is tends to care as we've seen in the past Asha's also supports explicit multi-gpu via DirectX 12 and that means we can use two cards without a bridge between them and have them each on the slots and they will commute via the PCIe bus rather than an external bridge this puts more load on that PCIe bus and between two cards of this power class we would expect that it might start hitting bandwidth limitations for really the first time we've seen almost ever on PCIe gen3 with two cards so that's what we're testing today together I don't have to speed on the PCIe specs versions four and five have been fast-tracked but aren't here yet and won't be for any reasonable period of time the PCIe gen 2 did about 500 megabytes per second per Lane Gen 3 does about 980 5 megabytes per second per Lane so we'll call that 1 gigabyte and Gen 4 will do one point nine seven gigabytes per second per lanes or call it two gigabytes so again previously no more really than a 1% difference with this type of test but now we're looking at how our performs here unfortunately again really only one game works for this we tried several so we tried total war Warhammer DirectX 12 with all the different eight betas that it has and that does support explicit multi-gpu but it will not detect both Titan V's and let them run without a bridge we also tried hitman and Deus Ex mankind divided both of which are explicit multigp ready neither of which saw both Titan V's and we do the same for rise of the Tomb Raider which is not a particularly great benchmark title to begin with but that's fine because it also did not see two Titan V's so we were lots of dashes and potentially synthetics further on this point there's a potential limitation in that if you wanted to bridge these you'd have to use env link a lot of you asked in our teardown hey what's that connector at the top of the card that doesn't look like ass alive fingers you're right it's env link it's a different protocol that Nvidia put out not too long ago about a year ago and has greater bandwidth if you want an NV link bridge it's $600 and we're not even sure if it's supported on this card right now because we don't have one because it's $600 so we're testing two of these explicit multi-gpu in ashes let's go through the numbers starting first with our dual of versus a single Titan B benchmark we can look at this more of an SLI versus single card test for starters I got two and one we overclocked the cards equivalently in each setup then tested with 16 PCIe lanes available the cards tested at 63 FPS average for the dual configuration with lows at 25 and 22 for one percent endpoint 1% low frames I miss respectively the single Titan V configuration operated at 37 point 6 FPS average immediately establishing that we are seeing at large noteworthy gains unfortunately ash's is where they stopped as we were unable to get the other games to detect both cards the stock cards compared at 33 FPS average versus 57 FPS average which is a noteworthy gain of 72% either way we're already seeing at least 68 percent gains from adding a second card which is significant and no the Titan V is not a gaming card and you shouldn't do this and it's not a gaming architecture but it's still scaling in games and that gives us the grounds to test whether the PCIe differences will scale moving on then to the PCIe bandwidth limitation testing our charts now focus on by 16 and by 8 configurations both overclocked to push the limits of the PCIe clock as it transacts between two devices we're measuring 63 FPS average for two cards with an X 16 interface and measuring 56 FPS average for two cards with an X 8 interface that's about a 12 to 13 percent performance improvement with PCIe gen3 by 16 and is significant in illustrating that we're nearing the end of Gen threes bandwidth which is 15.75 gigabytes per second for 16 lanes PCIe gen 4 will push about 2 gigabytes per second per Lane and will resolve any potential issues with these cards but NVIDIA is also using again it's $600 and be linked bridge to solve for this except not in games but in real applications that are intended for the card as for single cards we did not notice any appreciable differences between the single Titan v configurations and by 8 versus the single Titan V and by 16 it appears that we start encountering the most issues when transacting across new cards in explicit multi-gpu the question becomes whether or not you'd notice this in other applications not every application cares about PCIe bandwidth my name is a great again or GPU miners used by once lots for all their work it's possible that application is more tailored for the Titan V would not run into this issue as the two cards may not even need to talk to each other like they deal with this game if you use these types of devices for their intended scientific purposes please let us know and we can do more bandwidth testing just let us know what application do you use because we're not experts in that area and that would give us a good starting point the takeaway then is that this is really interesting information we're getting data that shows that finally after a couple of years of writing this test every time a new high-end card comes out we're seeing some performance Delta when it comes to two of these transacting across the bus it's one game that's not representative of the whole but for purposes of our audience gaming looking at a next-gen architecture Volta we can see that it's starting to approach the limitations of PCIe gen3 now how much that matters will largely be determined based on what Nvidia does with this card in its gaming cards in the future because Volta as it stands today is probably not going to make it into the next 1180 or whatever they call it so we'll see how that changes things and see if it stays the same but for now it looks like we're actually approaching those limitations which is pretty cool because we haven't been able to say that before 12% difference is a big one Nashes so that's all for this one as always you can find more information in the article below or just check the channel for all of the other videos on the titan v's subscribe for more you can go to store it I can't resist dotnet slash mod matte to preorder this mod matte or catch our upcoming recap of the livestream we did where we tried to take top 10 world records with Titan VIII overclocking on a 79 80 XE that's all for this one I'll see you all next time