Wednesday, October 17th 2012
NVIDIA Kepler Refresh GPU Family Detailed
A 3DCenter.org report shed light on what NVIDIA's GPU lineup for 2013 could look like. According to the report, NVIDIA's next-generation GPUs could follow a similar path to previous-generation "Fermi Refresh" (GF11x), which turned the performance-per-Watt equation around back in favor of NVIDIA, even though the company's current GeForce Kepler has an established energy-efficiency lead. The "Kepler Refresh" family of GPUs (GK11x), according to the report, could see significant increases in cost-performance, with a bit of clever re-shuffling of the GPU lineup.
NVIDIA's GK104 GPU exceeded performance expectations, which allowed it to drive this generation's flagship single-GPU graphics card for NVIDIA, the GTX 680, giving the company time to perfect the most upscaled chip of this generation, and for its foundry partners to refine its 28 nm manufacturing process. When it's time for Kepler Refresh to go to office, TSMC will have refined its process enough for mass-production of GK110, a 7.1 billion transistor chip on which NVIDIA's low-volume Tesla K20 GPU compute accelerator is currently based.
The GK110 will take back the reins of powering NVIDIA's flagship single-GPU product, the GeForce GTX 780. This product could offer a massive 40-55% performance increase over GeForce GTX 680, with a price ranging anywhere between US $499 and $599. The same chip could even power the second fastest single-GPU SKU, the GTX 770. The GK110 physically packs 2880 CUDA cores, and a 384-bit wide GDDR5 memory interface.
Moving on, the real successor to the GK104, the GK114, could form the foundation for high-performance SKUs such as the GTX 760 Ti and 760. The chip has the same exact specifications as the GK104, leaving NVIDIA to tinker with clock speeds to increase performance. The GK114 will be relegated to performance-segment SKUs from the high-end segment it currently powers, and so even with minimal increases in clock speed, the chip will have achieved sizable performance gains over current GTX 660 Ti and GTX 660.
Lastly, the GK106 could see a refresh to GK116, too, retaining specifications and leaving room for clock speed increases, much in the same way as GK114, except, it gets a demotion to GTX 750 Ti, GTX 750, as well, and so with minimal R&D, the GTX 750 series gains a sizable performance gain over its previous generation.
Source:
3DCenter.org
NVIDIA's GK104 GPU exceeded performance expectations, which allowed it to drive this generation's flagship single-GPU graphics card for NVIDIA, the GTX 680, giving the company time to perfect the most upscaled chip of this generation, and for its foundry partners to refine its 28 nm manufacturing process. When it's time for Kepler Refresh to go to office, TSMC will have refined its process enough for mass-production of GK110, a 7.1 billion transistor chip on which NVIDIA's low-volume Tesla K20 GPU compute accelerator is currently based.
The GK110 will take back the reins of powering NVIDIA's flagship single-GPU product, the GeForce GTX 780. This product could offer a massive 40-55% performance increase over GeForce GTX 680, with a price ranging anywhere between US $499 and $599. The same chip could even power the second fastest single-GPU SKU, the GTX 770. The GK110 physically packs 2880 CUDA cores, and a 384-bit wide GDDR5 memory interface.
Moving on, the real successor to the GK104, the GK114, could form the foundation for high-performance SKUs such as the GTX 760 Ti and 760. The chip has the same exact specifications as the GK104, leaving NVIDIA to tinker with clock speeds to increase performance. The GK114 will be relegated to performance-segment SKUs from the high-end segment it currently powers, and so even with minimal increases in clock speed, the chip will have achieved sizable performance gains over current GTX 660 Ti and GTX 660.
Lastly, the GK106 could see a refresh to GK116, too, retaining specifications and leaving room for clock speed increases, much in the same way as GK114, except, it gets a demotion to GTX 750 Ti, GTX 750, as well, and so with minimal R&D, the GTX 750 series gains a sizable performance gain over its previous generation.
127 Comments on NVIDIA Kepler Refresh GPU Family Detailed
If they do that and start allowing voltage control again, they'll have a far superior product for people wanting to overclock. GPU Boost nonsense + no voltage control kept me away from the GTX 680.
And like-wise, the same applies to any tech site.
That is all. GK110 is un-released, nobody except nVidia employees and those that work at nvidia board pertners know anything about it, and none of them can comment due to NDA.
So anything, anything at all about it...is questionable.
Heck, it might not even actually exist, and is only an idea.
Post a pic of a GK100 chip, full specs and everything else officail form nvidia, and I'll stop my speculation.
Otherwise, if you don't like my post..that's just too bad. the report button is to the left, if you like.
you cna say asll you like that it was planned, you have no proof, and neither do I. And neither of us, if we did, could post it. So I can think and post what I like, and so can you. It's no big deal...only you are making it a big deal that I do not agree with this news.
The rest of us is just saying that it is entirely posible and probable that GK100 existed and was simply delayed or slightly redesigned into GK110, in a move similar to GF100 -> GF110. The proofs although rumors, are out there and have been there for a long time. Rumors about chips don't always end up being entirely true, but there's always some true to it. GK100 was mentioned many times. GK110 DOES exist. 2+2=4
All in all Nvidia has already shipped cards based on the 7.1 b transistor GK110 chip, so the notion that such a chip cannot be made is obviously false.
To each his own I guess.
You can find info just as easy, too.
And becuase of this, I do think nvidia knew long before AMD's 7970 release that GK110 was not possible(which is when that news of GTX680 being a mid-range chip), and as such it wasn't meant to be GTX680, ever. Is GK110 the ultimate Kepler design...sure. but it was NEVER intended to be released as GTX680. It was always meant as a Tesla GPGPU card.
Liekwise, AMD knew that Steamroller...and excavator were coming...and that they arre the "big daddy" of the Bulldozer design...but that doesn't mean that Bulldozer or Piledriver are mid-range chips.
If GK100/110 was so unfeasible as a gaming card that it was never meant to be one, they would design a new chip to fill in that massive ~250mm^2 difference that exists between GK104 and GK110, instead of using GK110 as the refreshed high-end card. GK110 being an HPC chip wouldn't have so many gaming features wasting space either.
EDIT: SteamRoller, etc. Worst analogy ever.
I mean, it's like dropping the hot clock. They knew they had to.
Keeping within the 300W power envelope with the full-blown Kepler design was obviously not possible, proved by Fermi, IMHO.
Jen Hsun said "The interconnecting mesh was the problem" for Fermi. That mesh...is cache.
And gaming doesn't need that cache. But... HPC does. Gaming needs power-savings, and dropping the hotclock and lowering cache and DP lowered power consumption enough that GTX680 is pretty damn good.
GK104..was that chip you just mentioned.
HPC is more money. WAY MORE MONEY. So for THAT market, yes, a customized chip makes sense.
See, Fermi was the original here. GF100 is the original, NOT GK100. or GK110.
If nvidia started with Kepler as the new core design, then I would have sided with you guys, for sure, but really, to me, Kepler is a bunch of customized Fermi designs, customized in such a way to deliver the best product possible for the lowest cost, for each market.
You may think the Steamroller analogy is wrong here, but to me, that is EXACTLY what Kepler is. And you know what..nVidia says the same thing, too. :p
The hotclock to me, and the lack of DP functionality, says it all. hotclock lets you use less die space, but requires more power. DP functionality also requires more power, because it requires more cache. Dropping 128-bits of memory control..again, to save on power...
If the current GTX680 was meant to be a mid-range chip, after doing all that to save on power, damn, Nvidia really does suck hard. :p
I speculate it wouldn’t or Nvidia would've not put in place such restrictions if there wasn't good reasons. Will they still have it that way for next generation? Yes, almost assuredly but at that point better TDP and improve clock and thermal profiles will mean there's will be no gain over operating at an exaggerated-fixed clock. I think for mainstream both sides will continue to refine boost type control. It provides them the best of both worlds, lower claimed power usage, while the highest FpS return.
It's allways entertaining when fanboys get butt hurt over peoples opinions, the fact is a stock reference 7970 is slower than a stock reference 680, I know this hurts you to accept this fact but it is true. As for the Ghz edition, compare it to a 680 that is factory overclocked and the result is thew same. My 680 classified walks all over any 7970, so :cry::cry: less and check facts more.
:rolleyes:
The 7970 gets beat by the 680, sure, but the pricing has updated itself to reflect that now - the 7970 is priced about the same as the 670, and the GHz edition priced around the 680.
If we go by similarities, as in they look the same to me with a few tweaks we can go back to G80 days. Same on AMD side. But you know what? They have very little in common. Abandoning hot-clocks is not a trivial thing. Tripling the number of SPs on a similar transistor budget is not trivial either and it denotes exactly te opposite of what you're saying. Fermi and Kepler schematics may look the same, but they aren't the same at all.
As to the rest. It makes little sense to think that GK104 is the only thing they had planned. In previous geerations they created 500 mm^2 chips that were 60%-80% faster than their previous gen and AMD was close, 15%-20% behind. But on this gen they said: "you know what? What the heck. Let's create a 300mm^2 chip that is only 25% faster than our previous gen. Let's make the smallest (by far) jump on performance that we've ever had, let's just leave all that potential there. Later we'll make GK110 a 550 mm^2, so we know we can do it, and it's going to be a refresh part so it IS going to be a gaming card, but for now, let's not make a 450mm^2 chip, or a 350mm^2, no, no sir, a 294mm^2 and with a 256 bit interface that will clearly be the bottleneck even at 6000 MHz, let's just let AMD rip us a new one..."
EDIT: If GK110 had not been fabbed and shipped to customers already, you'd have the start of a point. But since it's already been shipped, it means that it's physically posible to create a 7.1 b chip and make it economically viable (the process hasn't changed much in 6 months). So like I said something in the middle, lika a 5b transistor and/or 400mm^2 would be entirely posible and Nvidia would have gone with that, because AMD's trend has been upwards in regards to die size and there's no way in hell Nvidia would have tried to compete with a 294mm^2 chip, when they knew 100% that AMD had a bigger chip AND they have been historically more competent at making more in less area. Nvidia can be a lot of things, but they are not stupid and would not commit suicide.
The fact you can't ignore that bit, says something.
What, I cannot speculate myself?
And when you can't attack my points, you go after my character? lulz.
As if I want to be the source of rumours. :p Yes, I want to be a gossip queen.
Like, do you get that? I'm not the one that posted the news...BTA didn't either...he just brought it here for us to discuss...
These same sites you trust, get it wrong just as often as right. Oh yeah, Bulldozer is awesome, smokes INtel outright..yeah..that worked...
HD7990 form AMD in auguest....but it was Powercolor...
Rumours are usually only part-truths, so to count them all as fact...is not my porogative. :pWell, that's just it. This is complicated stuff.
I am not saying at all that GK104 was the only thing...it isn't. But GK110 was never meant to be a GTX part. Kepler is where the Geforce and Tesla become truly seperate products.
And yeah, it probably did work exactly like that...300mm2...best they could get IN THAT SPACE, since this dictates that they can get so many chips per wafer. You know, designs do work like that, so they can optimize wafer usage...right?
Could just be reading them wrong, I suppose, but I think not.
Anyways, rumors are rumors, but they exist for a reason, and this particular family of rumors has been around for almost a year now... plenty long enough to indicate there's something to it.
Enough debating about rumors for me, though.
Looking at the whitepaper, anyone who knows a damn about GPUs can see that GK110 has been designed to be a fast GPU as much as it's been designed to be a fast HPC chip. Even GF100/110 was castrated in that regards compared to GF104, and G80 and G9x had the same kind of castration, but in Kepler the family where "Geforce and Tesla become truly seperate products." they choose to mantain all those innecessary TMU, tesselators and geometry engines.
- If GK104 was at least close to 400mm^2, your argument would hold some water. At 294mm^2 it does not.
- If GK104 was 384 bits, your argument would hold water. At 256 bit, it doe not.
- If GK110 didn't exist and had not released 6 months after GK104 did...
- If GK110 had no gaming features and wasn't used as the high-end refresh card...
- If GK104 had been named GK100... you get it.
And no, I do not agree with the summation that GK110 was intended to be a "fast GPU". The needed die size says that is not really possible.
But, since it's for HPC, where precision is needed over speed as a priority, that's OK, and lowered clocks, but greater functionality, makes sense.
However, for the desktop market, where speed wins overall, the functionality side isn't so much needed, so it was stripped out. This makes for two distinct product lines, with staggered releases, and hence not competing for each other.
I mean likewise, what do all those HPC features have to do with a gaming product? :p
And as cadaveca said, the info we have right now is just rumors and speculation, lets just wait and sooner or later we will all know for sure.
www.techpowerup.com/170096/NVIDIA-Maximus-Fuels-Workstation-Revolution-With-Kepler-Architecture.html
Their Maximus platform is composed off GK104 based Quadro and GK110 based Tesla cards. So I think that you're missing much more than I.
And oh, I don't doubt there will be a GK110 based Quadro, but it's not been announced yet afaik. I've only heard about them in the same rumors as the GeForce part so... ;)And yet it all points to Nvidia using it. And in the past they have used chips of the same size and quite successfully.
And an HPC chip has never been profitable on it's own and I don't think it is right now either.
For me, Medical imaging is part of the HPC market. Precise imaging isn't needed just for medical uses either, anything that needs a picture that is accurate, from oil and gas exploration to military uses, all fall under the same usage. Both Tesla and Quadro cards are meant to be used together, building an infrastucture that can scale to consumer demands, called Maximus. If you need more rendering power, say for movie production, you got it, or if you need more compute, for stock market simulation, that's there too, so I fail to agree you've posted much that agrees with your stance there. Nvidia doesn't build single GPUs...they build compute infrastructure.
Welcome to 2012.Did you read that press release?
:p
I mean, that whole press release is nVidia claiming it IS profitible, or they wouldn't be marketing towards it. :p
In fact, that press release kinda proves my whole original point, now doesn't it? ;) GK104 for imaging(3D, Quadro and Geforce), GK110 for compute(Tesla).
Like, maybe I'm crazy...but...well...whatever. I'm gonna play some BF3. :p
i could care less about your classified 680s blah blah i still had my card months before 680 was available and enjoying roughly the same performance.
Simple fact is if i want to be a dick and pull useless numbers the 7970 holds the World Record for 3DMark 11 Extreme, Heaven Extreme, among others
when both cards are clocked they perform the same, they excell in certain games over their rival and vice versa
AvP favors AMD
BF3 favors NVIDIA
etc
And I'm not confusing anything... the 7970 and 670 are about the same, and the 7970GE and 680 are about the same. If you overclock the 7970 or 7970GE, they'll match an overclocked 680 - trading blows depending on the game/test and overall being about the same.
I know you have to defend your silly purchase of an overclocking-oriented voltage locked card, though :laugh: