Star Citizen: The Power of Intel® Optane™

Transcript — every row opens the video at that moment

1. 00:00:07

   good afternoon everybody my name is Alejandro Jose I'm a Solutions Architect for Intel Corporation I'll be one of your presenters today and I want to thank you for coming here the other two presenters will be here today with us is gonna be Sean Tracy from con pyram here's a technical director also from until we have Nathan predict he's a technical marketing engineer and well three of us will be your presenters and we're gonna be going over how by using interrupting technology cloud Imperium games was able to accelerate the game development and also kind of enhance the game experience so with that said let's get started on today's agenda so first of all we want to give you kind of a quick background on an overview of the findings from accounting periods also we're going to also give you a quick

2. 00:00:55

   introduction to what Intel optic technology is so you have a better idea after that the second point we're going to talk about is interrupted technology talking about the benefits what actually is and how has it helped other companies after that will go into the into the meat of the of the presentation of today which is actually going over how this technology helped improve all the different developments from for gaming and lastly we'll have it open for questions and answers so that's how Genda looks in here today so for now I'm gonna pass it over to Shannon 8 to give you the quick overview for the upcoming presentation so I'll kind of go through this a little quicker than I might otherwise so octane just a kind of level set everybody here obtained is basically Intel's implementation of new media new media

3. 00:01:43

   technology called 3d crosspoint the 3d crosspoint media makes up the foundation of the card then with an Intel controller and for more on top of that and then obtain is actually represents multiple products right so we're going to be talking mostly about the obtain SSD here the 900p there's also obtained memory there's some very strong similarities between the different products but they are intended for different usages this is mostly around the the 900p so you know what is 3d crosspoint what is obtained it's basically a new media technology a new a new way of storing data it doesn't work like NAND technologies work today it works in a different way and that brings some unique properties some unique advantages to it that you wouldn't get with the NAND device life or one thing from a programmers

4. 00:02:30

   perspective it's a bit addressable now this one actually works the 900p works a lot like a normal SSD you read and write to LBAs like you would with a normal is d but under the hood the controller is capable of actually addressing very small chunks of data relative to an end device which talks more in terms of pages we also have what you know one of the main properties here that will benefit we think game developers in the future the latency the access time to get to the data is somewhere in south of 10 to 10 micro second so it's about an order of magnitude below an end device so I think you see as we move forward with you know with the customized software that kind of really takes advantage that you'll see some real strong benefits that take advantage of that we'll see a little bit more on these last two bullets here we've got slides on performance consistency and the endurance those are both very

5. 00:03:16

   different than NAND devices and we'll go into that in more detail later I'm going to turn this over here now to Shawn he's going to talk about the the cloud imperium bit yeah so I'll just speak from here it will go into it a little bit more and just tell you what I'm gonna talk about anyways once we get to our section here and it's some of the things that we did to use multi-threading to improve some of the performance internal to the game engine itself and how we could do that to leveraging on NAND technology from Intel obtain things like multiple concurrent file reads at runtime improved loading times and ultimately just getting a smoother experience using the Intel obtained technology and then finally we'll go into how it's actually improved our build process as well as we're building you know up to 1.3 terabytes of source data at one time it

6. 00:04:06

   was real important that we had a real powerful storage device to be able to use it Thank You Sean so with that said we'll go over I'll go over quickly to more details on the in topic technology so as Nathan have said there are different types of products out there for different type of usages so the first one that have heard before the first ones come out with the Intel obtain memory which is pretty much an accelerating device so what it does accelerates your system once you pair it with a hard drive so you parrot that memory module with the hard drive to get like an SSD performance so that's kind of it does the usage for the Intel obtain memory the other two are storage devices so like a regular SSD or or hard drive well not like a regular because these guys are a lot faster more and more durable and lower latency the one the ones that

7. 00:04:56

   we have there is the 800 P which is more like up for like a day to day getting closer to the performance side and the one that will actually talk today and the one that cloud imperium to use is the 900 p which is designed for performance or endurance so that said why performance of endurance I think performance or endurance will benefit most of the most of these groups actually we're nerfing most of you guys who have heavy workloads so I'm talking to about the gaming developers media creation engineering workloads scientifical and medical research which are very heavy workloads that once you load them up on your system you run out of RAM and then you start doing what is called paging you start staring accessing the drive the storage many many times to get all that extra data so with the lower latency and how fast it

8. 00:05:46

   is you're able to do this with all this workflow so those are kind of the work those that we're looking for and we think that this drive is a really good test solution once you're working with demanding with them ending work looks for example I've talked to someone earlier today that they have this huge picture of I think was 16 K and high PPI so high a per pixel pixel per inches and they had to actually chop that picture in several stages in order to be able to process it so the idea is by using this drive you nationally have to do that you can actually load that whole picture and actually be able to work with it so those are kind of the usages that we see and also new possibilities why new possibilities because this is a brand new technology so there's so many things that US and Intel can think but if we

9. 00:06:34

   actually you guys help us to develop new technology see new possibilities we don't know what's out there so that's one of the reasons we actually are working with a client peering games because they have the know-how on the game development and we have the know-how in the hardware so they can actually come up with with a better ideas or more ideas or newer ideas that we haven't thought about so that's where we also look into new possibilities and with that I will pass it over to my friend Nate's there we'll talk you a little bit more of the benefits and why this also it's a good solution for you guys okay so you know at the end of the day I think you know what this is gonna come down to is you know why should I care right why do you guys care about this this is all you know hardware and I assume a lot of you guys are software developers here right so in order to explain why I think you should care I

10. 00:07:21

    have to kind of go into this concept of what's called queue depth right and I don't know how much this will be familiar to you so apologize if this is repeating something you already know but all the i/os that go through a system go through a storage queue and in the case of SATA you've got a single queue and that's generally not that big of a deal because it drives relatively slow but when you move into an nvme or PCIe device with nvme now you've got a you've got a queue for each one of the cores in your system right so if you've got an eight-core system you've now got eight queues and if you think about that the number of i/os that are in the system are actually fixed right and so now instead of putting all eight of those or all you know all let's say 20 of those IOT's into a single queue now you're spreading those out over eight queues and what happens is that lowers the depth of the queue right the queue now has only a couple entries each and it each queue has a couple entries each

11. 00:08:09

    rather than one deeper queue and you know why is that important if you actually look at a datasheet for an SSD those data sheets are gonna are gonna list performance data right they're gonna list sequential data and they're gonna list random data the random bid is very important because most of the workstation workloads are actually highly random and we'll see more about that in just a minute but the other important thing here is that the numbers that they'll quote are usually in a ops you know you might see a number like 300,000 I ops 400,000 I ops those numbers are almost always quoted at queued up 32 or above so if you actually look at this profile that we've taken with a number of common applications right these are applications that you you know you probably use before visual studio maybe side effects for rendering if you look at a profile in a workstation environment those key adepts actually

12. 00:08:56

    stay below 5 80% of the time more than 80% of the time right so these high 300,000 IAP numbers that you keep seeing quoted on these data sheets you'll never see that on a workstation you might see it on a server where the Dacula are able to get a little bit deeper but workstations operate at very low cute ups so one of the things that that makes octane unique you know and we'll see some more specific data in a few minutes is the queue depth the load queued up performance is actually you know very close to a high queued up performance in other words you'll get to the high performance aspect of the drive quickly and you'll stay there across the entire range right and what that will translate into for a user is it's usually going to be something like responsiveness right that millisecond delay that you get when you're trying to open something it goes away and you get used to that I mean I've been using one of these drives myself for quite a while and you get

13. 00:09:44

    used to it when you go back to another device it actually it's hard to go back so let me yeah let me show you a little bit in terms of actual performance and metrics at these key depths right we see here the random read and random write these are you know again in a workstation environment most things are random and you'll see if you look at the actual random read which again most workstation environments are mostly read you'll see up to a 6x performance differential right and that's over a very high performance nvme NAND device so a huge difference in terms of usable performance right not high-end you know it's it's not like you got a sports car and you're driving around in a school zone all day right it's something that you can actually see and feel and in your day to day operation this because of the way rights work you'll see the numbers not quite as

14. 00:10:31

    high of a differential they're at the low key depth but you can see as you get up to the higher key depths there it starts to taper off so even when you do start writing a lot the drive continues to accelerate where an and device may just taper off and stop okay so putting it all together most people aren't just reading or just writing all of the time they're usually doing a combination though too so what this graph represents here is we've taken a queue depth of four which is I think you know a relatively decent number and we've held that steady as we as we varied the read from 100% to 0% right and if you look at an end of that number that that actual throughput of data there even though you might see these numbers like three thousand megabytes a second you can see the actual usable speed here is closer to two hundred megabytes a second and it

15. 00:11:19

    stays relatively in that range until you get very close to the full write performance then it starts to ramp up sharply again if you look at what obtain can do here it starts out at a high level and it just kind of stays there right so it doesn't really matter you can read you can write you can interleave the tude the performance will stay generally the same and then the last bit I'm going to talk about here too is endurance so when again when you start talking about SSDs you have to worry about right wearing the cells out right you don't have that problem with hard drives for the most part but on an SSD the cells will eventually go bad and a lot of drives will have spare and like but if you look at the actual data here this is this is warranted performance right so these SATA SSDs here nvme SSDs on the right hand side this is what the vendor will actually warranty the drive to allow you to write to them before they're out of

16. 00:12:07

    warranty essentially right that number of varies anywhere from 150 terabytes up to about 400 terabytes whereas the obtains you know again orders of magnitude higher than that so practically speaking you can write one of these op tain drives up to 22 times alright sorry you can write one of these nvme or ananda devices up to 22 times and replace them 22 times before you've worn out a single one of these obtained drives so you don't have to worry about the drive disappearing you don't have to worry about the cells going bad the drive is is guaranteed to work out to this up to this terabyte written data that you see here okay so in terms of what that means for actual applications alejandra we'll talk more about that now appreciate it so staying in the on the theme of how did

17. 00:12:57

    these translate to you who have he who here have spent a lot of time where you kick off a render or a simulation compilation and you just okay I just kicked this off and I'm gonna go now and you know spend go get some food good coffee go play around no sore fight and then your supervisor to ask you like what are you doing why you're not working or like well I'm just waiting for compile so I'm sorry to tell you that with this obtained drive this sort plate fight time haze we're gonna actually reduce it and at the end of the day that actually makes your life easier because when you will go ahead and kick off the compiled in a Friday you're worrying throughout the whole weekend it did to compile their work that it crash and you have to wait for such a for such a long time a period to see if it happens but now with the with this type of storage with this technology Intel obtained

18. 00:13:44

    technology you actually can do that a lot quicker and we have seen that through rendering through a loading tumbler projects also now the ability to to run a larger workloads and so that actually helps you to get all these done faster so how does it translate to you you get things done faster you get more time for you or more time for each iterate on any changes that you have to do to your project to show a couple examples of that quickly is we have also worked out with company called digital dimensions which they implement in the technology and they sell for the one percent faster rendering time also by using these drives they also incremented the processor usage by seventeen percent which means that you just spend this this amount of money to buy a processor there's actually sitting idle waiting for more information just because the current storage is not fast enough

19. 00:14:33

    defeated but now that we have the Intel technology that is able to feed it faster that you have improved the its utilization by more because now he can get more information and he can get going and do the work faster also we work with side effects and they have seen by the scene under rendering they've seen an improvement of the 2.7 X so that has also helped them a lot when it comes to particle rendering which is the when you're doing all these calculations over you doing for example a maelstrom of water and try to see how the physics and the interaction whichever particle goes goes with it so it went from actually 17 hours the seven second clip of a maelstrom from 17 hours to taking to render to actually 6.3 hours so that's a big improvement from them so if you want to make any changes you don't have to wait I guess two days you can you just do it and you guys

20. 00:15:20

    should do to two rounds of that rendering in one day and do any changes they need to do but those are just some examples that we have with the real example that we want to share with you guys in the important one and why we're here is how has clone Imperium seen any advances or on there gaming development by using this technology and for that I'll hand it up with you Sean thanks a lot guys so us at cloud Imperium games making star citizen as well as squadron 42 it was really nice to be early adopter it's partnered up with Intel because we really got a chance to be first boots on the ground in terms of trying out some different things seeing what we could actually do to leverage the technology so the first thing I want to talk about is kind of

21. 00:16:08

    the end user optimization what the players are actually seeing for us and and we found out very very quickly that performance is the key to a good player experience and to get that performance we have to do some things internal the game engine and one of the big ones of course is multi-threading now this isn't a multi-threading talk nor is that really about CPUs but I do want to explain some of the really core concepts of it and really multi-threading is a concept that we always have more than one object to call you've got more than one object to code more than one particle emitter be updated you've got multiple files streaming in at once you've got multiple pieces of game code going on at one time you've got entities with different components on them that you need to update per frame some that you don't need to update to render an actual frame so these concepts actually resulted in two really key systems that are internal to our engine one of them is the batch updater and the second one

22. 00:16:56

    is the background workers so batch updater what's a batch updater so a batch updater is something that we have actually put together and it's a system composed of batch workers and this is probably not a unique concept to b2b to hear about really and batch workers we found out a little bit like rabbits especially for the game coders you get a couple of these you learn how to use them and all of a sudden you've got dozens of them and you've got to try to be wrangling them all across the all across the code so these are used to update batches of the same object that's distributed over multiple CPU cores what's really important to us being a PC game is that we're developing for a bunch of different platforms and depending on the users they might have a better CPU they might have a couple more cores they might have a couple more

23. 00:17:43

    threads that they can actually throw at the game engine versus another user so we wanted to be able to be as scalable as absolutely possible so it's really it's really useful in combination with an entity component system and a lot of engines have already moved this way and a good example is unity has always kind of been the set up in this kind of method on Rails going that way as well as lumberyard and really what it does it allows you to relatively easily parallel eyes of your game code without your game coders necessarily understanding the the underlying process that's going on there so in terms of the batch updater we use this for frame dependent work so anything that we need computation to show a frame in in in in that method now on the other side of it we have our background worker system so the system

24. 00:18:32

    this one is designed around single job job objects and the idea is to run the background worker systems whenever the batch workers run out of work so when you're going yo buddy and get back to work we can actually do these batch background workers so this is where we actually begin to be able to leverage the technology of NAND SSD or Intel octane technology so these background workers are really tightly coupled with our low-level i/o system and it specifically gives us some pretty big benefits that are not only user facing but also facing for the game developers so how does that work we couple these background workers to the i/o system for Intel 3d NAND SSD and Intel obtain SSD tech so we can read multiple files in parallel brock's using asynchronous i/o versus a more serial access pattern that

25. 00:19:19

    you would get in a platter drive sort of this file by file by file and to give you an example some of our we've got spaceships in the game these are you know it's 300 megabyte asset that's just made up of a bunch of different files at one time so anytime that we have a player come into a system we've got to load all these ships anytime we have a player come close to the planet we've got to load all the outposts on the planet giving them a sort of stutter free experience is a pretty important thing for us so basically when a file block finishes a free background worker wakes up it decrypts and decompresses said block even if it's not fully in memory yet so this is usually where you know the theory of game load times being improved by obtain SSE if you don't do something like this you're actually still going to run into a bottleneck so do the number of parallel files we can work on multiple file blocks in parallel it gives us up to for 16 kill the bug bite

26. 00:20:09

    block blocks in flight at one time so in the end we get four blocks in flight across n number of cores a number of processes in parallel so again that's a lot of just you know quick explanation of how we're using this but what does that actually do for the gamer and that's what this is more about well what it does is it improves game loading times now in star citizen we actually only have one load you load the universe and the universe is a pretty big place there's a lot of stuff to load within there so one of the examples that I've got up here we had a 7200 rpm platter drive and we loaded the universe on it and then we tried the exact same thing on an Intel obtain 900 P and the load time was actually 130 seconds faster which is significant improvement in that one load and now against an SSD SATA the

27. 00:20:58

    improvement was a little smaller but absolutely still there so four and a half second improvement to load time so again having the multiple concurrent file reads during streaming or loading made this a lot faster so moving on from this just showing an example of that actually happening now the two on the left-hand side of this video are the Intel obtain 900 piece so that's look this is actually our load screen it's not the most exciting of videos I understand seeing a load screen and you'll see this little pink placeholder sort of load time graph that we had and this is the SATA SSD on the top right and the 7200 rpm platter drive on the bottom right so after about thirty some-odd seconds we get our nice load so

28. 00:21:48

    the Intel obtain 900 feet that's our one load so that's how our players are experiencing that immediately shortly thereafter we've got the SSD inside a configuration that's gotten loaded up now and we're still waiting for the platter drive now I'm not going to make this whole video run I wonder how I can speed this up here it's right there I'm not gonna insult everybody telogen so take three minutes to load but after about three minutes you see the 7200 RPM come behind there's no plate play blast I didn't want to have you guys sit through the whole thing point being is that you've got it in

29. 00:22:36

    about 180 seconds faster which is a huge deal for the players especially so now how did actually benefit game development and actually working day to day there was already some pretty big improvements right out of the box so one of the big ones was real word that Oh copies are faster than ever a lot of the version control systems now like it or even perforce have these parallel sync options and you can really abuse those with this kind of technology you can also use this with internal tools so we have an internal tool that we called copy build and it's basically how the artists how the program was how everybody pulls down 65 gigabytes of packaged data after a build process has run so one of the sort of graphs up here is how much faster was that build copy so again this is about sixty five gigabytes being copied down onto a developer's PC on a 7200 platter drive mechanical drive rather it's faster by

30. 00:23:26

    288 seconds considering you do this four five six times a day really depending on the day that's a huge improvement and again any downtime just reduces that amount of swordplay like Alex was talking about on and in on an SSD and in SATA configuration it's a faster by about 76 seconds this is again still a big improvement that you're pulling down these builds multiple times throughout the course of the day so again there's a very very real-world example the other one and this wasn't something that it was necessarily expecting but ended up benefiting our artists especially the facial animators quite a lot large source file export was massively sped up so one of the things that we do we have some Triple A actors within our games so the data that we use to represent these guys within Maya is pretty massive we've got really really high density meshes

31. 00:24:15

    and some super super high density animations so one of these exports can be up to seven gigabytes of uncompressed Calado data this is all being written to drive sometimes it's going to take about four and minutes to export so we got that four and a half minutes just using an Intel 900p down to just a couple seconds so for artists to be able to iterate on this kind of stuff it was it was incredibly important to them and again going from you know the the minutes down to seconds for their sort of iteration loop was super beneficial and one other one that I wasn't necessarily expecting was our frame by frame captures this is something we do sometimes when we want to really really consistent 60 frames a second we're outputting per frame a JPEG or a PNG onto disk and usually the bottleneck when you're doing that as you're playing the game is actually writing to disk and we we slow down the

32. 00:25:02

    game to a state that it can actually keep up so when we did this usually when we play it on you know like a 7200 rpm kind of drive you're playing it in slow motion it's like 510 frames a second and you're very slowly capturing this and most of us have put these kind of things together for demos or for marketing content or something now being able to do it on an Intel 900p this was almost playing at real time so it was much more 24 30 frames a second so I could play the game quite normally and still have a lossless capture done written straight to disk so this was a really nice thing too lastly we benefited pretty massively from changing our build configuration on our build machines to use the Intel 900p so one of the things that we have to do just before we build all of our assets

33. 00:25:50

    is we actually have to synchronize all those assets and I talked a little bit about this earlier but the actual data that the build process has to do is much much more than an artist will have to take a look at so right now in star citizen we have 1.3 terabytes of data to sync per branch this is crazy this is a lot of data this has done multiple times depending on content needs so sometimes I'll change a preset in our resource compiler and it requires of rebuild all of our source assets so they'll have to pull that down onto the build machine and then all those assets will need to get compiled we do this for every major release and considering we've release every three months that's kind of a big deal and anytime we change data formats so we got a speed improvement pretty pretty dramatic actually and just before I go into the speed our current

34. 00:26:37

    configuration was 24 SSDs in out of raid six on on a super box that we have in Austin and Nate had come out worked with our build guys a little bit and we got our data sank down quite a lot more so right now in the current configuration it was taking about six hours and 13 minutes to bring down 1.3 terabytes of data once we change it out to our opt-in configuration which instead of 24 SSDs we actually hooked up just for Intel obtain SSDs and vrock raid configuration so for Data Sync this only took three hours in 30 minutes so we basically went twice as fast this again is a huge deal because when the build guys are coming in and they have to do a full data sync of this they're either going to have to do it on the weekends they're gonna have to do it

35. 00:27:24

    late at night sometime when it's actually not going to block development or it's not going to block our fast code builds from coming up so they get a lot more sleep and we're able to actually use a lot a lot less drive space in our in our servers so so really to kind of summarize everything we saw some end-user optimizations faster load times more smooth streaming during gameplay we have some optimizations directly for game development that's file sinks source file exports as well as capturing and then we were able to accelerate the data build process and all of this we pretty much did out-of-the-box there was very very little special setups that we needed to do to get this benefit right away so really it opened new doors for us for creating these novel experiences for our players we can load bigger data sets and we can load them faster and

36. 00:28:13

    smoother than we can on different drive configurations so really allows us to improve long-standing systems and tools to take full advantage of the Intel obtained technology thanks for watching for the latest and greatest in star citizen squadron 42 you can subscribe to our Channel or you can check out some of the other shows and you can also head to our website at www.uvu.edu/library