GTX 780 rumors

[Lepton87]

I don't think AMD made any decree, but they have left board designs up to partners in the past. Hopefully reference models stay as robust as they've always been, but I was just hypothesizing and I certainly don't hope AMD heads in that direction. I would never waste my money on a $500 card that doesn't have overvolting capabilties unless there were no other options out there. Even then I'd probably grab the card that the guys at XtremeSystems or somewhere figured out a solder mod for and do that.

In the end any new part from either company will be competing against my 7970 @ 1.35GHz for my dollar. I wouldn't consider a card unless it was at least 25% faster at stock, so we're talking close to GTX 690 performance (which is about 33% faster than what I currently have), and then some ample overclocking room on top of that. While I'm optimistic about what both companies can put out, that's a high bar to reach without a die shrink. We'll see. :thumbsup:

IMHO it's not going to happen on 28nm. I also want a single card that outperforms my current cards in all circumstances preferably by a healthy margin, but I'm afraid that bar is set far too high for 28nm. I hope I'm wrong, I want to finally ditch multi-gpu quirks. Just look how unimpressive the refresh was on 40nm. It seems both you and I want what usually comes with a die shrink without that process shrink. Usually after die shrink flagship cards perform on par with the past dual cards or not even that. At launch 5870 couldn't quite match 4870X2. GTX295 to GTX 480 GTX was about the same performance. Then 7970 couldn't match 6990 at stock with launch drivers, but OC headroom more then made up for it. Right now 7970GHz is faster then 6990, but it took driver maturity and clock increase. GTX680 still can't match GTX590, but it always seemed to me like a clocked to the roof mid-range card, so no wonder why it can't match GTX590. So hoping for GTX690 level of performance on a single card at 28nm process is wishful thinking. I really want that to happen but if the history is of any indication it won't happen.

 

[bunnyfubbles]

Those nominal/boost clocks are far too close together to be believable.

and yet its the same difference in clock rate as the 680...so how is it any harder to believe now when it has already happened?

really, the most unbeliaveble part to the clock rate is the fact that these rumors suggest a much larger GPU chip, the same 28nm manufacturing, and then a clock rate increase on top of that.

 

[bunnyfubbles]

I don't think AMD made any decree, but they have left board designs up to partners in the past. Hopefully reference models stay as robust as they've always been, but I was just hypothesizing and I certainly don't hope AMD heads in that direction. I would never waste my money on a $500 card that doesn't have overvolting capabilties unless there were no other options out there. Even then I'd probably grab the card that the guys at XtremeSystems or somewhere figured out a solder mod for and do that.

In the end any new part from either company will be competing against my 7970 @ 1.35GHz for my dollar. I wouldn't consider a card unless it was at least 25% faster at stock, so we're talking close to GTX 690 performance (which is about 33% faster than what I currently have), and then some ample overclocking room on top of that. While I'm optimistic about what both companies can put out, that's a high bar to reach without a die shrink. We'll see. :thumbsup:

considering nVidia limited their first round of Kepler to a chip smaller than their midrange chips during Fermi, they have plenty of room to expand regardless of being stuck on the same manufacturing process

 

[3DVagabond]

considering nVidia limited their first round of Kepler to a chip smaller than their midrange chips during Fermi, they have plenty of room to expand regardless of being stuck on the same manufacturing process

Seems that making a chip any bigger on 28nm though has been difficult for them. We aren't sure how much bigger a chip they can actually get manufactured and be commercially economical.

 

[bunnyfubbles]

Seems that making a chip any bigger on 28nm though has been difficult for them. We aren't sure how much bigger a chip they can actually get manufactured and be commercially economical.

the difficulty with 28nm didn't stop them from making one of their biggest ever chips with the GK110, and while its true GK110 is exclusive to HPC, they still get the experience of working with 28nm on such a scale without the consumer space risk

and 28nm hasn't stopped AMD from making a chip that is significantly larger than their average flagship with Tahiti, and rumors from AMD's side suggest they're going to go even bigger for their next round as well

thus far, all the evidence suggests that nVidia has been conservative to mitigate their risk in order to avoid another consumer Fermi fiasco, and nothing to suggest they're avoiding going big altogether

It wasn't long after the release of GF100 and GF104 that they released GF110 and then GF114 and all the woes of 40nm were pretty much gone. I don't see why we wouldn't see the same with 28nm, especially when nVidia avoided the issue in the first place and even has started to distance their HPC parts with their consumer ones, further mitigating process fab risk.

Whats more is that the rumors of the GPU to go into a GTX780 suggest it would be closer to 5B transistors than to the 7+B in GK110, so its not like they're going completely insane here with size.

 

[3DVagabond]

the difficulty with 28nm didn't stop them from making one of their biggest ever chips with the GK110, and while its true GK110 is exclusive to HPC, they still get the experience of working with 28nm on such a scale without the consumer space risk

and 28nm hasn't stopped AMD from making a chip that is significantly larger than their average flagship with Tahiti, and rumors from AMD's side suggest they're going to go even bigger for their next round as well

thus far, all the evidence suggests that nVidia has been conservative to mitigate their risk in order to avoid another consumer Fermi fiasco, and nothing to suggest they're avoiding going big altogether

It wasn't long after the release of GF100 and GF104 that they released GF110 and then GF114 and all the woes of 40nm were pretty much gone. I don't see why we wouldn't see the same with 28nm, especially when nVidia avoided the issue in the first place and even has started to distance their HPC parts with their consumer ones, further mitigating process fab risk.

Whats more is that the rumors of the GPU to go into a GTX780 suggest it would be closer to 5B transistors than to the 7+B in GK110, so its not like they're going completely insane here with size.

I'm not talking about AMD having or not having difficulty. Look at how long it took for GK110. What happened to GK100? It was scrapped. We still don't know whether or not GK110 can be made in quantities to make it commercially viable for the consumer market. If so, it'll have taken so long as to have missed an entire series, GTX-600.

 

[boxleitnerb]

GK100 was not scrapped, it simply didn't exist. GK110 was always planned for last. Don't let names fool you - the code name GK110 doesn't necessarily imply that there ever was a GK100.

 

[3DVagabond]

GK100 was not scrapped, it simply didn't exist. GK110 was always planned for last. Don't let names fool you - the code name GK110 doesn't necessarily imply that there ever was a GK100.

You saying here that there was never a GK100 planned is the first time I've ever seen that claim. Either way it still took them a very long time to produce their big chip. There's still no evidence that would lead us to believe they can produce it in large enough quantities for consumer desktop use. They are a very limited availability sku. AFAIK, they haven't even made a graphics card out of it yet, just a Tesla card.

 

[Silverforce11]

GK100 was not scrapped, it simply didn't exist. GK110 was always planned for last. Don't let names fool you - the code name GK110 doesn't necessarily imply that there ever was a GK100.

Source?

 

[Rockinfella2k]

Well I hope this rumor is false, I just picked up a 690 and really don't want to upgrade in a month or two.

 

[boxleitnerb]

Source?

It was said by Ailuros over at beyond3d multiple times and he has good sources. After the Fermi debacle, it's only logical. Why would they make the same mistake twice?

Btw, where is the source that a GK100 existed? If it was scraped, it had to exist first. And if it existed, it has to be mentioned somewhere (and by that I mean some credible concrete source, not some speculative deduction like: GF100 existed, then GK100 must have existed as well).

You saying here that there was never a GK100 planned is the first time I've ever seen that claim. Either way it still took them a very long time to produce their big chip. There's still no evidence that would lead us to believe they can produce it in large enough quantities for consumer desktop use. They are a very limited availability sku. AFAIK, they haven't even made a graphics card out of it yet, just a Tesla card.

That claim is pretty old. Planned at some early point - maybe. Fully designed and taped out and cancelled due to bad characteristics - certainly not.

This is what I think happened:
Nvidia, seeing the problems they had with GF100, immediately changed their plans for Kepler and pushed the big chip to the last position in their release schedule. I suppose that was around 2009/2010. It's not like the hypothetical GK100 and the actual GK110 are completely different chips. Nvidia certainly didn't start at zero with GK110. Already accumulated knowledge and developments surely went into GK110. But the notion that they had something you could call "GK100", an actual complete design, not just bits and pieces in early stages of development, seems unlikely.

 

[Avalon]

I heard the next NVIDA GPU will be "a neural net processor; a learning computer."

God save us.

 

[3DVagabond]

That claim is pretty old. Planned at some early point - maybe. Fully designed and taped out and cancelled due to bad characteristics - certainly not.

This is what I think happened:
Nvidia, seeing the problems they had with GF100, immediately changed their plans for Kepler and pushed the big chip to the last position in their release schedule. I suppose that was around 2009/2010. It's not like the hypothetical GK100 and the actual GK110 are completely different chips. Nvidia certainly didn't start at zero with GK110. Already accumulated knowledge and developments surely went into GK110. But the notion that they had something you could call "GK100", an actual complete design, not just bits and pieces in early stages of development, seems unlikely.

To be clear, I think all that was scrapped was releasing it in the 600 series lineup. They needed more time to develop it and it slipped to the next series and was named GK110, so as for it not to be outdated almost upon release.

None of that matters. All I'm saying is it took them a lot longer to develop their big chip. Nothing you said contradicts that. It's also a very limited production sku, with only a Tesla card being produced so far.

 

[bunnyfubbles]

I'm not talking about AMD having or not having difficulty. Look at how long it took for GK110.

yeah, it was available in the HPC space sooner than GF110 was available relative to Cypress. If anything that shows they've actually learned from their Fermi experience and have thus far handled 28nm much better than 40nm.

What happened to GK100? It was scrapped.

1. you're assuming it even existed
2. even if it did and was scrapped, there could be a plethora of reasons as to why it was scrapped, many of which would have nothing to do with fabrication process but rather design flaw

We still don't know whether or not GK110 can be made in quantities to make it commercially viable for the consumer market. If so, it'll have taken so long as to have missed an entire series, GTX-600.

right, we don't know, but what we do know is that GK110 isn't actually intended for consumers, and likely never was, so it simply doesn't matter whether or not nVidia can viable produce 7.1B transistors on 28nm for consumers.

nVidia doesn't need to go from a 3.5B transistor chip to a 7.1B one to compete with AMD in the next round, they're doing that now against a 4.3B transistor Tahiti when, traditionally, nVidia has been the one that has needed a lot more transistors to compete.

I think the next round will have ~5B transistor chips from both sides, and nVidia will end up slightly on top instead of trading blows like this round, and it will be due to their architecture being more efficient for gaming because of how it seems that they're starting to separate their hardcore HPC designs from the consumer space.

Granted, that could be totally wrong, as we've only had wavering rumors to go off of, but I think its a pretty safe bet at this point in time.

 

[3DVagabond]

28nm was far better than 40nm for everyone. Look at when the 4770 was released and how long it was before we had mass availability of 40nm, even for AMD. Besides, being better than Fermi isn't exactly a big accomplishment.

Comparing transistor count isn't really a good comparison when we are talking stripped down gaming chips compared to fully capable GPU's. For example, no one thinks GK110 will be 2x as fast as GK104. Not in gaming, anyway. Compute's another story. That's where the bulk of those extra transistors are applied.

 

[monster88]

What 680 lacks is memory bandwidth.Switching to 384 bit will fix this and from the added features of GK110 it will definitely beat AMD.From last I heard the next gen from AMD was looking like ~20% faster than 7970.

remember when people were saying the same thing about the 680 before it came out?. people were actually believing the absurd claims of 40% faster than the 7970....see how that turned out.

lol....the things people believe.

 

[bunnyfubbles]

28nm was far better than 40nm for everyone. Look at when the 4770 was released and how long it was before we had mass availability of 40nm, even for AMD. Besides, being better than Fermi isn't exactly a big accomplishment.

so you agree with me, going bigger shouldn't be a problem...

Comparing transistor count isn't really a good comparison when we are talking stripped down gaming chips compared to fully capable GPU's. For example, no one thinks GK110 will be 2x as fast as GK104. Not in gaming, anyway. Compute's another story. That's where the bulk of those extra transistors are applied.

how is it not relevant? transistors are transistors, getting then functional for the task they are designed for is no less relevant

also, it doesn't matter how fast GK110 would be in the consumer space if its never going to be in the consumer space, the point is we can chop off a good ~2B transistors off of what GK110 established and end up with a ~5B transistor GPU that is still 50% more powerful than GK104

 

[NTMBK]

NVIDIA GeForce GTX 780 Specification

GPU: GK2xx
Base Clock: 1100 MHz
Boost Clock: 1150 MHz
Memory Clock: 1625 MHz (6500 MHz effective)
Memory type: 3GB GDDR5
Memory Interface: 384-bit

Given that we already know that Big Kepler is GK110, I'm inclined to ignore this rumour.

 

[3DVagabond]

so you agree with me, going bigger shouldn't be a problem...

Did you really get that from my post?

how is it not relevant? transistors are transistors, getting then functional for the task they are designed for is no less relevant

also, it doesn't matter how fast GK110 would be in the consumer space if its never going to be in the consumer space, the point is we can chop off a good ~2B transistors off of what GK110 established and end up with a ~5B transistor GPU that is still 50% more powerful than GK104

Another total fail understanding what I wrote.

I don't know what to say. I can't think of any way to make you understand me. :\

 

[boxleitnerb]

Yesterday the GTX680 Phantom 2GB dropped about $100 here in Europe in several shops. Maybe an indicator that the wait isn't too long now.

 

[tviceman]

Yesterday the GTX680 Phantom 2GB dropped about $100 here in Europe in several shops. Maybe an indicator that the wait isn't too long now.

There was an evga gtx 680 for $399.99 no rebate needed last week. Definitely the cheapest I have seen. In a semi-related news bit, Charlie says AMD is delaying the 8000 series roll out because of current 7000 series inventory levels.

 

[Fx1]

On the same 28nm with the same architecture you can guarantee the improvements will be rubbish for the increase in prices.

These releases are to Nvidia and AMD what the S is to Apples iphone 4S

 

[willomz]

Of course maybe the 680s are coming down in price in response to prices on the 7970?

 

[boxleitnerb]

On the same 28nm with the same architecture you can guarantee the improvements will be rubbish for the increase in prices.

These releases are to Nvidia and AMD what the S is to Apples iphone 4S

Nope, I don't think so. AMD still has some room if they lower voltage and Nvidia has quite some room for improvement if we look at K20X. I think 30% better performance are as good as guaranteed IF the 780 uses GK110. Anything above that...we'll have to see.

Of course maybe the 680s are coming down in price in response to prices on the 7970?

True, that could also be the reason.
But another thing is the GTX 660 SE. The 560 SE launched shortly before the 680 and the 460 SE shortly before the 580. I guess March 2013 is a good estimate. Roughly one year after GTX680.

 

[Fx1]

Nope, I don't think so. AMD still has some room if they lower voltage and Nvidia has quite some room for improvement if we look at K20X. I think 30% better performance are as good as guaranteed IF the 780 uses GK110. Anything above that...we'll have to see.



True, that could also be the reason.
But another thing is the GTX 660 SE. The 560 SE launched shortly before the 680 and the 460 SE shortly before the 580. I guess March 2013 is a good estimate. Roughly one year after GTX680.

Where do you get this from?

The 680 GTX has more shaders than the 670 and is only 3% faster. The 2GB b 3GB of Vram wont make much difference in anything but the biggest resolutions

No process shrink either so thats out of the question.

Basically all you get is a few more of what you got now which means about 14% better performance. No way you will get 30%