[quotemsg=15383008,0,328798][quotemsg=15382479,0,925801]This will be facing the 14nm Nvidia X1 replacement chip later this year for xmas, so already going to be behind. I'd expect Nv to double their 512 FP32 and 1024 FP16 numbers of X1 when they go to samsung in Q2/Q3.[/quotemsg]Really? What other chips have doubled between 20 nm and 14? When have GPUs ever doubled in a single generation (if you compare cards with equivalent TDP)?
[quotemsg=15382479,0,925801]It's far easier to port to the exact same chip on a PC where you made your game, to mobile that uses a chip with every feature of the big brother, just scaled down. IE, devs have been using maxwell for ages (in gpu terms anyway...LOL), so they don't have anything new to learn to port a pc game to X1[/quotemsg]I doubt this. While X1 might support DX12 & OpenGL 4.5, Android doesn't. That means a pretty substantial port. Also, the parameters of any mobile chip are going to differ from its PC and console cousins, which will require significant tuning (as will the smaller screen size). And NV doesn't have such a big chunk of the mobile market that developers can specifically optimize for it - most will be aiming for a far lower common denominator, meaning much of the X1's power will probably go to waste.
TLDR for some, but I'm sure stock holders in some of the companies would want to know this stuff.
You're talking desktops, I'm talking SOCS. Desktops have many things that used to be doubling etc that are now just a few mhz more etc of the same memory, pretty much same channels etc. Early on in discrete you had doubling of bandwidth via new faster ram or more channels, along with new gpus (just like we have in mobile's early revs up to now). Now they can't simply add channels cheaply etc on discrete without blowing up cost. The low hanging fruit isn't there today, but it is on mobile still.
28nm K1 is doubled by 20nm X1. This has been going on for pretty much every rev from many socs when shrunk (especially with help of surrounding better stuff). In this case 3x fp16, close for fp32, 2x perf/watt etc. The next page here shows manhattan offscreen 1080p 63fps vs. K1's 31fps. A double right? Not sure why you wouldn't think you could double the gpu at 14nm+finfet on top. I expect a 14nm X1 replacement to come with double the cores and probably faster mem, larger cache, faster clocks on gpu etc (for Qcom, apple etc too). When you put that all in a device you'll get 2x fps most likely in many things in gaming.
K1 vs T4. Double in 3dmark.
Unreal 4 already runs on K1, so any game running unreal 4 can easily be ported to K1/X1 etc. The same can probably said for Qcom etc, as Qcom has demonstrated unreal 4 on a ref tablet a while back last year.
3dmark 1.2 unlimited again 31K for K1, vs. 16K for T4 in shield handheld and that is overall, where the gpu test is 36K for K1 and 17K for T4 (over double). Same page basemark dunes more than double, Hangar benchmark is just over a triple offscreen! T-REX HD is 68fps for K1 to 13.7fps for T4...LOL. IF you look at T4 vs. T3 you see much the same. You could likely do the same on Qcom chips. Apple (imagination) has been doing the same from rev to rev.
Shield vs. Nexus 7 (T3 version, Nexus 7 2013 in there too, but look at T3 nexus 7). 19fps vs. 3.5fps in 3dmark ice storm and the graphics score is 7x higher for T4 vs. T3 here. Even the physics side was over double. Ice storm extreme a good 6.5x faster also. Basemark X offscreen shows the same at 4x etc for T4 vs. T3. Way over 2x faster. gfx bench shows the same a page later at almost a triple again.
Why would you not expect a die shrink + finfet on X1 to add another smx? I'm guessing 512 cores vs. 256 of x1 for the xmas chip. Gpu perf has been exploding for every rev from all vendors due to the low hanging fruit. Much like we used to have years ago on desktop. Huge returns from shrinks on mobile still due to how much of a difference it makes to shrink a chip (and more stuff) when dealing with <10w tablets etc. Check each rev of the ipads you'll see the same from apple.
Ipad4 (32nm a6x) vs. ipad3 (45nm a5x). Once you get to REAL benchmarks (not synthetic crap) like Egypt HD it's a double from 21fps to 40fps for ipad4. Same story, both running same res.
"Despite sub-2x gains in a lot of the synthetic tests, Egypt HD shows us what's possible in a simulated game: the new iPad is roughly twice the speed of the previous gen model when running at the panel's native resolution."
Same benchmark in 1080p offscreen shows 25fps ipad3 vs. 47fps ipad4 (almost a double here too). Do you read the reviews or what? I could run through qcom revs but you should already get the point. Note apple did it with a decrease in die size also. A5x 165mm^2 vs. A6x 123mm^2. Also note same anandtech review they improved battery while doing it. Again, this shows the point. They shrunk the die size of the soc, got doubles in all the gaming stuff, and improved Bat-life. Both ipad 3/4 have 42.5whr batteries.
I can't tell you how many double from 20nm to 14nm because we have no examples yet...LOL. But to assume it won't happen after I've shown you repeatedly it happens each gen, is kind of silly. We're not just talking gpu here though, as they are surrounded by many things that upgrade per rev also (faster mem, clockspeeds etc that all help out each new wave of soc/device revs). The only way I see this not happening for xmas chip (after x1) would be that it's a shrink of the same tech and all the revs before were quite a bit different also for the most part (more features etc as anand points out). But I still say they'll just put 2xSMX etc. I do think it will slow down some now (much like desktops) at 10nm etc but we'll see if they can do 4xSMX (etc for everyone else) or not. If 10nm versions are pascal or something then that might not be relevant and they'll get a double maybe some other way (not directly related core#'s). Either way, so far my point is proven by many soc makers. Until I see a rev NOT double you're probably wrong
It has ALWAYS happened so far (and far more than double in many things gpu on socs as shown by NV's revs from T3-T4-K1-X1).
Porting isn't hard. Also NV noted that with all of the ports they've been involved with so far they took less than a few weeks and had SMALL teams doing it and they commented most of the team's time went to mapping the gamepad functions not the porting of code (for portal & Half-life 2). You have OpenGL ES3.1 + AEP already, not sure why you wouldn't think higher levels would come soon. When all the chips can do OpenGL 4.x you don't think there will be an AEP rev2 or something (or a later version after lollipop) than will support these features? You do know ES 3.1/3.2 (surely coming soon for mobile) is a subset of it's big brother 4.x right?
Gameloft etc are already amping stuff up. Games don't look ugly on android these days (or ios).
Why do you think NV is releasing tools/debuggers for OpenGL 4.x apps on ARM socs?
"Tegra Graphics Debugger 1.3 with added support for OpenGL ES 3.1 Android Extension Pack, OpenGL 4.3 and 4.4 applications."
Sorry, I just don't see your points. Google wants gaming on android and DX/consoles dead (hence the AEP, and likely full opengl coming soon), and all the socs do what I claimed noted above many times. Just like Wintel wanting windows/Intel everywhere, google wants android/ARM everywhere. As chips move to full opengl (NV isn't going to be alone for long) so will games on android. Then again if ES3.2/4.0 etc can do everything OpenGL 4.x can do there's no need, it's still an easy port just a different way. The whole idea here is to get desktop stuff on mobile, then assault desktops. Period. Games first, then apps, then full blown ARM PC's with discrete gpus etc. Valve (hates the same stuff, DX, Windows, MS store) and others will help the mission. X1's power (and other upcoming socs) won't go to waste as Unreal 4, Unity 5 etc games amp up on mobile.
Unity5 info, directly aimed at many platforms.
The whole point of both engines (and others coming I suspect) is easy cross-platform stuff.
"Unity 5 brings a wealth of new capabilities and features enabling all developers, from the smallest indie teams to the largest studios, to create amazing games and interactive experiences, and painlessly deploy them to almost any platform."
That's in the 1st paragraph. Clearly this is the point. Unreal4 has same story. Again, easy deployment to everywhere is the point of both engines. Porting older games (such as Half-life 2, portal etc) is easy also (as noted weeks!) since they can do this stuff in 3.1 ES and more with AEP. No doubt more can be done faster as they move to 3.2 ES or 4.0 (glNext? Maybe no 4.0, just everyone goes glNext?)
Serious Sam 3 running on K1 talk from devs:
"Yes. We were even more amazed than you are. That's the first version of SS3 for Android. And it would not have been possible without this Tegra miracle.
This chip looks like something that will knock everyone's socks off. We couldn't believe how good it is. It supports full OpenGL4, and we practically directly ported the entire PC game to it, along with PC content. It was not cut down like e.g. for Xbox 360.
As an additional point, when we compiled the code without errors and ironed the glitches in initialization code, as soon as it started the graphics correctly for the first time, it went right through the intro sequence, and into the game. Without any noticeable visual errors. I think I can safely say this is the first time ever. Normally, we'd have a few weeks of unexplainable black screens, inverted colors, garbage in textures, broken shaders, etc, until it looks ok visually. With K1, this just worked."
EASY. Developer bragging about it here. This is K1...NOT X1...LOL. This is a DEV talking, not me and that was OVER a year ago. X1 will certainly make stuff even easier, and I think we'll see much more as everyone gets to this level later this year or next year. I'm thinking they're holding it for the next chip and a shield handheld update probably in a few months (GDC announcement maybe? who knows). You'd need a big release or two to show off. Note the dev (AlenL working on SS3) says this about OpenGL in response to a guy asking about if it runs that or ES:
"KRON, the engine works on both full OpenGL and the ES. Just that the latter is limited in quite a few aspects. This is the main advantage of K1 - support for full OpenGL."
It doesn't sound like the game is limited by android. He kind of sounds like it's full openGL. Again X1 just ups the game even more. The game could also be delayed by waiting for android stuff though, as he notes commenting to another user linux had an issue with the game until all 4 cores could be 100% on, and was patched. He expects the same for android soon. He notes graphic level on SS3 with K1 was running on medium. I suspect X1 can do high at least and who knows if 14nm can hit ultra (depending on res of device I'd guess etc).
From AlenL post 56 (notes also T4 backporting may not happen, too little power):
"(*) Note though that nvidia's drivers are top of the line. I expect a lot of driver problems with other vendors and that may be a very limiting factor."
We'll see how long it takes for NV to take over mobile gaming devices
Devs go where it's easy to port first. We are just beginning, and as I said before NV has major advantages with features now and devs, and as the dev says, DRIVERS that work! Note the post from a guy asking if OpenGL4 is on android (post 70/71) and the devs answer:
"The version we have, yes. But who knows what each vendor puts in. That's why we also support GLES3 and also GLES2 if some additional extensions are present."
Talos principle uses Serious Engine 4 (SSam3 uses 3.5 IIRC).
Check out page 19/20. Tim Sweeney's (EPIC unreal guy) comment in there too, saying easily run PC stuff on mobile from here forward.
Trine2 K1 dev talk, and comparing versions of PS4, PS3, Xbox360 etc. Note levels mentioned as left off on xbox360/ps3 but included in K1 because optimizing for consoles was tougher (so they gave it up, not so on K1, it runs just like PS4/PC versions he notes). And one more comment here about the porting:
"The studio had a headstart of sorts owing to the proliferation of platforms Trine 2 already supports. The Kepler graphics architecture in Tegra K1 is derived from a full desktop part, and supports every major graphics API. Android ports typically need to be down-converted to the more restricted OpenGL ES - Trine 2 is a pure conversion from desktop code - this is the "secret sauce" in the latest generation of mobile graphics processors, the path that makes mobile a viable target for a huge library of existing titles."
So kind of sounds like if you have the engine/driver/chip support you run FULL OPENGL right? The next paragraph the dev says they used the mac/linux OpenGL rendering as the base. It also doesn't sound like "SIGNIFICANT TUNING" is required. I mean they gave up on monster levels in Xbox360/PS3 versions, but NOT K1. It's all in there. Further:
"Indeed, Trine 2 on Mac OS X and Linux used OpenGL for rendering and that renderer served as the basis for the K1 port."
Not quite sure if I'm getting his statement correct here, but clearly the future won't be scaled down to ES. They clearly say that is the point of K1 forward (from both devs, trine and serious sam guys).
"Nvidia has stated that Half Life 2 on Android is a port of the Linux version with OpenGL support to OpenGL ES, and based upon a casual playthrough of both Half Life 2 and Portal, it’s not immediately obvious that there are any issues with the engine port itself. In fact, it runs quite well."
“Through our close collaboration with NVIDIA, Epic’s Unreal Engine 4 ‘Rivalry’ project demonstrated at Google I/O shows what’s possible when PC-class gaming technologies and performance are brought to mobile devices,” said Tim Sweeney, founder of Epic. "In under three weeks, Epic created the “Rivalry” demo. How could something so visually complex be created in such a short period of time? As it turns out, Epic had previously created the scene to show off UE4 engine features. However it was made for a DX11-class PC. Working together, the Epic team and NVIDIA dev tech engineers were able to port the demo, along with new, original content, to Android and AEP. The “Rivalry” demo Google showed was actually running the same high-end desktop rendering pipeline."
The mobile gaming future is bright, any way you want to slice it. I used NV as an example here, but make no mistake others will be here too, it's just a matter of when, and then gaming on mobile explodes for real. You won't be able to claim it's wasted power when everyone can do it (NV device or not).
glNext coming GDC this week (mar 5th). With Neil Trevett being president of Khronos (Creator of OpenGL ES work group there also, so definitely a mobile guy) and head of Mobile Ecosystems stuff at NV, I'm guessing this will be the version to truly merge mobile/pc etc. So the next version of android (or the one after it?) will probably run this. Everyone will join on this no doubt (Qcom, Apple etc all will use it). You don't want to be left out of the best gaming content.
Watch out DX12. Valve will push this for linux and probably a steamOS for android at some point (or NV will, no brainer for both sides). One more way for Gabe to stick a knife into Bill Gates/DX12/Windows/MS Store
I think we're done here