Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
Nvidia announces Turing architecture
As I write this, the presentation is still ongoing. Nvidia has announced a 754 mm^2 die of their new Turing architecture with 14 Gbps memory. That's almost certainly GDDR6, and likely the "768 GB/sec" GPU that Micron promised Nvidia would launch early this year, which somehow shrunk to 672 GB/sec.
Apparently the heavy focus is on ray tracing, which is very different from the rasterization that has been mostly used for 3D graphics for about the last 20 years or so. The idea of rasterization is that you create a scene by drawing triangles one at a time. Each triangle gets broken into pixels, and then for each pixel of the triangle, you figure out whether it's in front of anything already drawn on that pixel (and discard it if not), and what color the pixel should be.
With ray tracing, you create a scene by drawing one pixel at a time. You basically cast a ray in some direction and see what it runs into to determine the color of that pixel. There doesn't necessarily have to be any triangles involved, though whatever your models look like, you need for it to be easy to determine whether the ray hits it or misses. Ray tracing makes reflections and partial transparency pretty easy to do, as you can just bounce the ray and continue off in its new direction to see what it hits next. It also makes shadows practical.
If you think you've seen reflections, transparency, and shadows with rasterization, you may well have seen a bunch of fakery that looks fine at a glance but is all wrong if you look closely. Ray tracing can do it right, and rasterization really can't. As image quality goes, ray tracing is better than rasterization, and it isn't close.
So why did anyone ever use rasterization? Because ray tracing is slow. Really, really slow. GPUs rely on nearby pixels on the screen having closely related work, whether warps doing computations together or getting global memory coalescence when reading in data. Ray tracing completely breaks that.
Nvidia's CEO is now making a big deal about how some nice looking demos that he's showing off are running in real time. While I believe his claims, it's also important to realize that, by computer game standards, the scenes he's showing off are pretty simple. The frame rate and resolution are also left unspecified, but likely not very high. And this for a card that will probably cost several thousand dollars. The cheapest card announced is $2300 for unspecified specs--and probably severely cut down. The top end model that is surely being showcased is $10,000.
For video rendering that needs to look excellent but doesn't need to happen in real time, this could be a big deal. Or it might not; that's a market that I don't really understand. But this is definitely Nvidia's attempt at getting render farms to swap out their Xeons in favor of Quadros.
What, you were expecting a GeForce card? The largest consumer GPU die ever was likely Nvidia's GM200, at around 600 mm^2 and used in the GeForce GTX 980 Ti. That was on a process node that was more than three years old by the time the card launched, so it was very mature. Even then, the bottom bin salvage part cost $650. There have been larger GPU dies such as the GP100 and GV100, but no GeForce cards ever launched based on them. If Nvidia made a GeForce card based on this Turing GPU, it would be the largest consumer GPU die ever by more than 25%. Selling those cards for $500 each is not a way to make money. Charging $1000 each makes it rather sketchy to call that a consumer card and might still not be that profitable.
Nor is it clear that scaling down this card would still give something useful for consumer use. As compared to a GeForce GTX Titan Xp, from Nvidia's own specs, it takes an extra 60% more die space to get 28% more TFLOPS and likely 23% more memory bandwidth, in spite of having the advantage of a newer process node. Scale it down to match GP102 die size and you likely end up with a slower card even on a newer process node.
For what it's worth, the new GPUs aren't available until next year, with a promised release timeframe of Q1 2019. Nvidia's CEO said in June that the next GeForce cards are still a long way off, and I think he meant it.
Edit: fixed memory bandwidth numbers. Not sure how I botched the arithmetic earlier.
Apparently the heavy focus is on ray tracing, which is very different from the rasterization that has been mostly used for 3D graphics for about the last 20 years or so. The idea of rasterization is that you create a scene by drawing triangles one at a time. Each triangle gets broken into pixels, and then for each pixel of the triangle, you figure out whether it's in front of anything already drawn on that pixel (and discard it if not), and what color the pixel should be.
With ray tracing, you create a scene by drawing one pixel at a time. You basically cast a ray in some direction and see what it runs into to determine the color of that pixel. There doesn't necessarily have to be any triangles involved, though whatever your models look like, you need for it to be easy to determine whether the ray hits it or misses. Ray tracing makes reflections and partial transparency pretty easy to do, as you can just bounce the ray and continue off in its new direction to see what it hits next. It also makes shadows practical.
If you think you've seen reflections, transparency, and shadows with rasterization, you may well have seen a bunch of fakery that looks fine at a glance but is all wrong if you look closely. Ray tracing can do it right, and rasterization really can't. As image quality goes, ray tracing is better than rasterization, and it isn't close.
So why did anyone ever use rasterization? Because ray tracing is slow. Really, really slow. GPUs rely on nearby pixels on the screen having closely related work, whether warps doing computations together or getting global memory coalescence when reading in data. Ray tracing completely breaks that.
Nvidia's CEO is now making a big deal about how some nice looking demos that he's showing off are running in real time. While I believe his claims, it's also important to realize that, by computer game standards, the scenes he's showing off are pretty simple. The frame rate and resolution are also left unspecified, but likely not very high. And this for a card that will probably cost several thousand dollars. The cheapest card announced is $2300 for unspecified specs--and probably severely cut down. The top end model that is surely being showcased is $10,000.
For video rendering that needs to look excellent but doesn't need to happen in real time, this could be a big deal. Or it might not; that's a market that I don't really understand. But this is definitely Nvidia's attempt at getting render farms to swap out their Xeons in favor of Quadros.
What, you were expecting a GeForce card? The largest consumer GPU die ever was likely Nvidia's GM200, at around 600 mm^2 and used in the GeForce GTX 980 Ti. That was on a process node that was more than three years old by the time the card launched, so it was very mature. Even then, the bottom bin salvage part cost $650. There have been larger GPU dies such as the GP100 and GV100, but no GeForce cards ever launched based on them. If Nvidia made a GeForce card based on this Turing GPU, it would be the largest consumer GPU die ever by more than 25%. Selling those cards for $500 each is not a way to make money. Charging $1000 each makes it rather sketchy to call that a consumer card and might still not be that profitable.
Nor is it clear that scaling down this card would still give something useful for consumer use. As compared to a GeForce GTX Titan Xp, from Nvidia's own specs, it takes an extra 60% more die space to get 28% more TFLOPS and likely 23% more memory bandwidth, in spite of having the advantage of a newer process node. Scale it down to match GP102 die size and you likely end up with a slower card even on a newer process node.
For what it's worth, the new GPUs aren't available until next year, with a promised release timeframe of Q1 2019. Nvidia's CEO said in June that the next GeForce cards are still a long way off, and I think he meant it.
Edit: fixed memory bandwidth numbers. Not sure how I botched the arithmetic earlier.
Post edited by Quizzical on
![[Deleted User]](https://images.v-cdn.net/banned_large.png)
Comments
I don't expect Nvidia to announce a next generation consumer GPU well ahead of time. They're already ahead of AMD in the efficiency metrics, so they want people to buy now, not wait. They'll try to keep hard information about the cards secret as close as they can to launch. And there's no need to tie the launch date to a conference.
You make announcements well ahead of time when you're behind. If you expect that someone who buys today is going to buy from your competitor, you want them to wait. See what Nvidia did with their launches of Fermi and Kepler, for example. Or more recently, AMD with Vega. But if you expect that someone who buys today is going to buy from you, you want them to just buy it and not wait.
As for RTX GEFORCE card announcement it’ll be a week from today at Gamescom. It’s clear it’s going to be the RTX 2080
And even if they do launch consumer cards based on Turing (or Volta), it probably won't be on a chip with a die size of 754 mm^2, nor the 815 mm^2 of GV100. Consumer graphics cards don't need a bunch of silicon dedicated to double precision (in Volta), ray tracing (in Turing), or half-precision matrix multiplication (in both). Maybe they'll offer a Titan T for $3000 as a companion to the Titan V of that price tag, but that's not really a consumer GPU.
If you take all that out, would Volta and Turing still be different from each other? Would they be different from Pascal? Maybe they would, but I don't know. We still don't have very many details about the architectures. And even if we did know all about them, that wouldn't tell us what would be in GeForce cards based on the "same" architecture. Some benchmarks make it clear that GP100 is very much of the Maxwell/Pascal architecture, but others make it look more like GCN/Polaris than consumer Pascal.
Ray tracing is not coming to games anytime soon, outside of the possibility of a handful of optional effects in a handful of sponsored titles, unless you count pre-rendered lighting effects not rendered on your GPU. Games for which neither a GeForce GTX 1080 Ti nor a Radeon RX Vega 64 meet even the minimum required specs aren't going to have much of a market anytime soon. Trying to make a game offer both ray tracing and rasterization is so outlandishly expensive that no one is going to do it.
Considering right now the dozens of titles with specific GameWorks implementations, HBAO+, HFTS, VXAO, PCSS+, TXAA, Ansel, HairWorks, WaveWorks, TurfWorks .. etc. NVIDIA will just adapt the Ray Tracing stuff into new effects that will be implemented just like the others: RT AO, RT GI, RT Shadows .. etc.
Unlike compute and AI, raytracing is entirely about graphics, and so it would make no sense to castrate or remove it from Geforce cards.
NVIDIA released a teaser for the Geforce variants right after the SIGGRAPH event, showing some players chatting online, the player names have hidden meanings and hints:
RoyTeX = RTX
Mac-20 = 20 Series
Eight-Tee = 80
NVIDIA RTX 2080
Also one of the players is named: Not 11
There is also the phrase: give me 20
Date of the launch is 20 August 2018, the numbers of the date appear in a specific order that translates into 2080 right at the end of the video in how they pop up in sequence
https://youtu.be/F7ElMOiAOBI
https://www.youtube.com/watch?v=ms7HQ7rckpA
RTX 2080 - TU104 - August - 50% faster than a 1080
RTX 2070 - TU106 - September - 40% faster than a 1070
GTX 2060 - TU116 - November - 27% faster than a 1060
GTX 2050 - TU117 - November - 50% faster than a 1050ti
Is it aimed at professional non-realtime rendering of images?
This is what this particular announcement and timing is all about.
AMD stock just advanced greatly versus Intel recently. nVidia doesn’t want to be ignored.
With Pascal, I forcefully said that it would be a paper launch, or at best a soft launch, because there simply weren't enough GDDR5X chips in the world to have a hard launch of anything besides a low volume part that doesn't matter. In contrast, GDDR6 has been in volume production for a number of months by now. Similarly, TSMC's 12 nm process node has been around for a while producing GV100 chips of the Volta architecture.
Look at the timeline for Pascal-based GeForce cards, for example:
GP104: May 2016
GP106: July 2016
GP107: October 2016
GP102: March 2017
GP108: May 2017
That's about a year to get their lineup out. Or let's look at Maxwell:
GM107: February 2014
GM204: September 2014
GM206: January 2015
GM200: March 2015
That's over a year. And Nvidia had far more incentive to hurry both with Pascal (die shrink from 28 nm to 16 nm) and Maxwell (massive architectural improvements over the very broken Kepler architecture) than they do with Turing.
That's also far too small of a performance spread to justify four dies rather than two or three plus salvage parts. That sort of blunder is a pretty sure sign of a random person with Internet access making things up, not an actual leak. Even if Nvidia does launch four new GeForce cards over the course of the rest of this year, it's far more likely that it would be two dies and salvage parts.
If Intel's foundry advantage is gone forever, should they really be worth much more than AMD? Even if it's just a brief blip and Intel will recover, having nothing remotely competitive to offer in servers--their core market and one they've long dominated--for over a year means that someone other than Intel is going to make a ton of money in that time, and AMD seems like the most likely beneficiary. If AMD gets considerable market share for a while, there's no guarantee that Intel goes back to dominating the server market with 99% market share and monopoly pricing power even if their fabs do reclaim the lead.
Nvidia, in contrast, didn't just fall on their face the way that Intel did. As both Nvidia and AMD rely on external foundries and can readily buy from exactly the same foundries, it's unlikely that Nvidia will get a large fab disadvantage as compared to AMD--or a large advantage, for that matter.
I'd bring up there are 3 main parts on that die for the first time ever (a new architecture) including RT core plus the Turing GPC is two SMs larger than a Pascal GPC, and it's obvious a cut down version will happen for consumer gaming.
My guess for RTX 2080 sizing (and with the 1080 at 314mm² ) mine would be around 400mm² and that is going from 16nm to 12nm. Not massive shrinkage feasible, but that is why we got new architecture, the die design is going to be different that before. I also think sub 200W too.
The RT core is specifically there to accelerate Ray Tracing, and at its core is mostly FP math. An average shader core can do that with many FPs but it’s not accelerating. The marketing we’ll see is Nvidia’s ace up the sleeve which is dedicated logic the accelerate.
GP104: May 2016
GP106: July 2016
GP107: October 2016
GP102: March 2017
GP108: May 2017
That's about a year to get their lineup out. Or let's look at Maxwell:
GM107: February 2014
GM204: September 2014
GM206: January 2015
GM200: March 2015
That's over a year. And Nvidia had far more incentive to hurry both with Pascal (die shrink from 28 nm to 16 nm) and Maxwell (massive architectural improvements over the very broken Kepler architecture) than they do with Turing.
You've failed to mention that this year they haven't had a reason to launch anything, there has been no competition from AMD. In the year since VEGA there has been no material gains drivers nor game development that have pushed those cards performance much past anything that the initial tests were done with.
But THE biggest thing not mentioned for launching earlier is what we saw with cryptocurrency. The incentive has been to wait because of that and with AMD being irrelevant. That is a hell of a long time to prep and design and fab a wicked card.
It's entirely about perception, public opinion, and nVidia wanting to square their GPU tech off against traditional CPU marketplaces - like raytracing.
As far as nVidia not doing a paper launch of GeForce - it's very much true that doing so isn't necessary from a competitive standpoint. AMD still isn't threatening Pascal right now, let alone whatever comes out next. But if nVidia needs a stock bump - you can bet that the first thing that will come out will be a press release for a paper launch, and nV's stock will go up 7-10% on just the news alone.... so there is very much at least one reason that nV ~could~ choose to do a paper launch.
I mean, otherwise, why else would they have just paper launched this Turing Quadro? They haven't even done anything significant with Volta yet.
With that said, I think NVidia will tell us about new GeForce cards soon.
When making several dies for a given architecture, you generally want each to have around double the performance of the next one down. You can then use salvage parts to fill in the gaps. For the Pascal generation, Nvidia has dies with 30, 20, 10, 6, and 3 compute units. For Vega/Polaris, AMD offers dies with 64, 36, 16, and 8.
The rumor you site claims four dies with the top one only 3-4 times the performance of the bottom one. I don't know if that many dies that many GPU dies with such a small performance spread have ever been launched by a single GPU vendor so quickly, but if it ever has happened, it was surely many years ago when the design and mask costs per die were vastly cheaper.
nVidia starts taking sales from themselves as soon as the roadmap for the next generation gets leaked, and it only gets worse as the rumors keep flying. I mean, how many posts on this forum alone have recommended waiting at least in part because the next generation is due out "any time now"? nVidia hasn't done anything official yet - and yet people are still holding off.
nVidia is all about the market valuation, not necessarily balance sheet profit. They don't always align or overlap exactly.
The engineering effort to make an entirely new generation for modest gains would be better put toward 7 nm parts that won't release until the TSMC 7 nm process node is ready. If they can get 50% more performance in the same size and power as before, then making new chips makes far more sense. I'm skeptical that they can get that kind of gains, however.
It's also possible that they'll launch new bins of old parts, perhaps aided by metal layer respins. Take the GeForce 1000 series, add 100 MHz and maybe disable a little less on salvage parts, and call it the GeForce 1100 or 2000 series. See, for example, what AMD did going from the Radeon RX 400 series to the Radeon RX 500 series, or what Nvidia has done in the past on various occasions, most notably going from the GeForce 400 series to the GeForce 500 series. That doesn't incur the cost of completely redoing your entire lineup.
From someone who claimed to of worked on Turing - with a google translate - https://www.zhihu.com/question/290167656/answer/470311731
GP102 on TSMC "16 nm": 11.8 billion transistors in 471 mm^2 for 25.1 million transistors per mm^2
Unnamed Turing on "12 nm": 18.6 billion transistors in 754 mm^2 for 24.7 million transistors per mm^2
Transistor density going down is not the sort of thing that you expect from a full node die shrink. Regardless of the names, "12 nm" is just a slightly tweaked and more mature version of the previous "16 nm" process node. It's not actually a die shrink.
I have no doubt that they can accelerate ray tracing by creating some fixed function logic for it. The problem is that that fixed function logic will be unused on consumer GPUs. Maybe they can justify it to avoid the costs of a major redesign between chips if it's only a few mm^2 for the entire chip, but not if it takes a lot of space.
Something in Turing that isn't there in Pascal is taking a whole lot of space. Consider:
GP102: 3840 shaders in 471 mm^2 for 0.123 mm^2 per shader
GV100: 5120 shaders in 815 mm^2 for 0.159 mm^2 per shader
Unnamed Turing: 4608 shaders in 754 mm^2 for 0.164 mm^2 per shader
If all the extra stuff in the Quadro RTX cards is still there in consumer cards, they'd need to increase clock speeds by 1/3 just to break even on FLOPS in a given die size. The tensor cores probably take a lot of space, but I have no idea how much that is versus the ray tracing junk. That Turing takes more space per shader than Volta on the same process node, even though Volta also has the tensor cores as well as double precision logic that Turing lacks, makes me think that the ray tracing stuff isn't small.
And don't think that games are suddenly going to switch to ray tracing so that they can use Turing efficiently. If a game relies on ray tracing rather than rasterization, then you'll end up with a game that can't run on a GeForce GTX Titan Xp or a Radeon RX Vega 64. By the time things have advanced so far that it's acceptable for a game to be unplayable on today's high end, any cards that launch later this year will be thoroughly obsolete, too. Even if all future GPU architectures are built primarily for ray tracing and not rasterization, it will be several years before games can really use it without killing their potential market size.
Once 7 nm is ready, of course you launch new parts on it as soon as you can. That Nvidia is launching a 754 mm^2 die on 12 nm now tells me that they don't think they'll be ready to launch chips with giant dies on 7 nm anytime soon. But just because they can't make a 600 mm^2 die on 7 nm just yet doesn't necessarily mean that they can't make a 100 mm^2 die.
For what it's worth, AMD showed off working a working 7 nm GPU in June, said it was already sampling, and would launch by the end of the year:
https://www.anandtech.com/show/12910/amd-demos-7nm-vega-radeon-instinct-shipping-2018
I don't know how quickly Nvidia will get to 7 nm, but once they can do 7 nm GPUs with acceptable yields, they will. And there, unless the process node is broken or something goes completely haywire, the gains available will be plenty large enough to justify a new generation.
Yes TMSC 12nm is just a refinement of TSMC 16nm but we’re talking maturity now with great bins. That small refinement will also bring some efficiencies. Who knows where the RT cores will reside. The RT cores are likely built into the SMs, they also might also reside in a separate area, like the center block.
The Turing architecture itself introduces the ability to run floating-point and integer workloads in parallel, which should help improve other aspects of performance. That will be a carry over for sure.
Btw a consumer AMD 7nm that will be a high end gaming competing card will be over a year away. Don’t bet any hopes on anytime sooner.
7nm AMD workstation card towards end of year, gaming on 7nm still long way off ie not 3 months away.