Alleged NVIDIA GK104 Specifications

[ocre]

If that is true, and with what AMD's rumored pricing is going to be, AMD is going to have a very, very short ride. Nvidia will stomp all over them in the high end and performance segment. The only real traction AMD will have is with whatever they do with the barts shrink.

i just heard pretty much the same thing posted in the 7970 thread:

A friend of mine mentioned to me over the weekend that they had run into some roadblocks and expected significant delay (month or so), then figured out a workaround was significantly easier than anticipated to put them back on track.

Cannot name source, so I never posted here, especially because he couldn't mention any firm dates and nobody would believe me anyway, but that's what I heard.

Even with that, Dec 22 is sooner than expected.

I was told a couple weeks ago AMD would not be able to get cards out soon because of a delay. But now i am hearing its all good. I suspect the launch will be in very short supply. I also think AMD is under pressure and want to be as far ahead of nvidia as possible.

Kepler will be a beast, but nvidia will not have the real kepler flagship for sometime. considering their situation, I think they have a pretty decent plan of action. They are really ambitious. I think they will execute much better this time around.

Its just gonna be great to see the 28nm cards row out. I heard mixed things on what AMDs core next will be like. I really cant wait to see how it turns. Cant wait

 

[Silverforce11]

Benchmarks on a bunch of games can always be biased one way or another, some games just favor different architectures, or even game settings. ie. BF3 with MSAA, NV wins up to 1080p easily, but at 1600p, they tank hard to unplayable. BF3 with NO MSAA on high or ultra, cheaper radeons match or outperform the NV counterpart. Crysis 2 with tessellation enabled kill AMD cards, until you set the tessellation factor in catalyst (preventing rabid over-tessellation) and there's no huge perf drop.

Casting a wider net with more games, to sites such as TPU or ABT:

Damn even this round, with the outlier gtx580, which is fine since its also an outlier in terms of price & TDP. What's AMDs advantage is offering similar perf at a less to much less power use. NV has CUDA. Its the consumers call.

 

[RampantAndroid]

Huh?

Also, where are you getting your prices from?

He said a 6990 costs the same as a GTX580. I guess $500 == $700 in his little messed up world.

 

[nsavop]

Leaked kepler performance slide? link

 

[Silverforce11]

Leaked kepler performance slide? link

To have legit benchmarks one would assume they have engineering samples. If they do indeed have samples that advanced, they wouldn't be launching THAT late in 2012. Conflicting "leaks" from the same article = a lot of bullshit for page hits.

Logic 101.

1024 CUDA cores and everything else upped is just an extrapolation away from gtx580 performance. Multiply things by 1.8 - 2.0 and you get roughly the "appropriate, not too bogus to be completely unbelievable" result.

 

[tviceman]

The more I think about it, the more a 768 core GK104 sounds reasonable. GT200b was 240 cores and just under 500 mm^2, while GF104/GF110 was 384 cores and around 350mm^2 (a little more or little less depending on who you believe). GF104 increased it's core count by more than 50% over the previous flagship chip, so a 768 core 28nm second tier chip sounds more and more reasonable. Assuming Nvidia releases a fully unlocked GK104 from the start, I am guessing real world performance will be ~40% above (or higher if they get really good yields and higher clock speeds) a gtx580.

The gtx660 or whatever this chip will be branded could give the high end tahiti a serious run for it's money when it comes to gaming. I am really excited for this next gen round!

 

[aphelion02]

Leaked kepler performance slide? link

So they are skipping the 600 series?

 

[tviceman]

So they are skipping the 600 series?

Hahaha I just noticed that. I am guessing the slide is fake. If high end kepler isn't coming until Q2 Nvidia more than likely is not making marketing slides yet.

 

[Genx87]

So they are skipping the 600 series?

The person who made up that fake chart probably confused Nvidias next gen series numbers with AMDs

 

[nsavop]

So they are skipping the 600 series?

Maybe the 600 series will be used for the rumored fermi refresh (GK104) and 700 series for high end kepler. But yeah most likely fake.

 

[SirPauly]

When nVidia has execution problems they make a lot of noise about how great their architecture will be. nVidia has really been quiet and things are very confusing.

 

[AtenRa]

I believe Gk100 will be released in Q1 (Feb, March??), Gk104 will be released a few months later.

Personal estimation.

 

[SHAQ]

I think they are less behind AMD than most people think. It is not a new architecture only a simple die shrink with more stuff added. Fermi took so long because it had to be redesigned a couple of times. Not so this time. It depends on when TSMC has reached process maturity and how much capacity they have to sell.

That slide seems more believeable to me since it says 780. A forgery wouldn't miss something so obvious. The 580 was a refresh of the 480 and the numbering scheme was changed instead of calling it a 490. That slide could be an estimate of the refresh of the next round or maybe the high end will have a 100 higher series number going forward to distinguish it more. But remember Fermi was always supposed to have 512 cores and a lot of the blame for missing that target at launch belongs to TSMC so 580 is really a misnomer as it is a same generation product.

If TSMC is behind again we may be looking at a similar approach and the 780 will be the same gen but delayed a few months.

 

[Vesku]

I'm starting to think NVIDIA is waiting for solid 7970 numbers so they can tweak their response. If they start showering tech sites in PR on 7970 launch than things are probably pretty dicey for them.

 

[dangerman1337]

So they are skipping the 600 series?

They pretty much did that for the performance and Enthusiast segment for the 300 series IIRC.

 

[tviceman]

They pretty much did that for the performance and Enthusiast segment for the 300 series IIRC.

The only gpu's that carried a 300 designation were in laptops. Since nvidia rebranded a few fermi's to be 600 series laptop gpu's, it does lend credence to the idea that nvidia will opt to name kepler gpu's with 700 series names.

 

[Vesku]

It would bring leading digit parity....

 

[RussianSensation]

Leaked kepler performance slide? [/IMG]

So we have HD7970 rumored to be up to 50% faster than the 6970 and now GTX780 being > 100% faster than GTX580? So what's that supposed to make the 7970, a 6850 of the next generation?

HD6970 (Base = 100%) --> GTX580 (115-120%) --> HD7970 (150%) --> GTX780 (100% faster than GTX580 --> 230-240%)

So that would make a GTX780 effectively 53%-60% faster than HD7970?



For real? You are telling me in 1 generation NV will surpass AMD by 1 entire generation? :whiste:

 

[tviceman]

So we have HD7970 rumored to be up to 50% faster than the 6970 and now GTX780 being > 100% faster than GTX580? So what's that supposed to make the 7970, a 6850 of the next generation?

HD6970 (Base = 100%) --> GTX580 (115-120%) --> HD7970 (150%) --> GTX780 (100% faster than GTX580 --> 230-240%)

So that would make a GTX780 effectively 53%-60% faster than HD7970?



For real? You are telling me in 1 generation NV will surpass AMD by 1 entire generation? :whiste:

Hahaha yeah the slide is a fake. Marketing slides with performance numbers don't show up several months in advance of a GPU's release. The specs and performance of Kepler are the biggest unknown factors in all of this. The real question is, will the hd7970 really only beat the gtx580 by 30% on average? If so, then I think AMD might not be able to touch Kepler's high end GPU. Time will tell though.

 

[Silverforce11]

7970 doesn't NEED to compete with high-end kepler, it just needs to compete with gk104 (aka beefed up Fermi), not high-end gk100 (is it even coming?) or even gk110 (rumored 2013).

The point is due to the different time to market, 7970 has no competition. If gk104 is in late Q2 2012, its not a direct competitor, it may compete with a refreshed 7990. Within 1-2 months, fine, i can understand direct competition.

 

[tviceman]

7970 doesn't NEED to compete with high-end kepler, it just needs to compete with gk104 (aka beefed up Fermi), not high-end gk100 (is it even coming?) or even gk110 (rumored 2013).

The point is due to the different time to market, 7970 has no competition. If gk104 is in late Q2 2012, its not a direct competitor, it may compete with a refreshed 7990. Within 1-2 months, fine, i can understand direct competition.

This is all very true. The timing of Nvidia's release will dictate how and what GPU's compete with each other.

 

[rgallant]

"That slide seems more believeable to me since it says 780. A forgery wouldn't miss something so obvious.

obvious like miss spelling Light
Metro: Last Light he has Metro Last Night
-pointed out in another forum

 

[Lonbjerg]

When nVidia has execution problems they make a lot of noise about how great their architecture will be. nVidia has really been quiet and things are very confusing.

The only thing lacking from the G80 launch is misdirection (aka "we might not go unified")...

 

[boxleitnerb]

If you look up Metro, half the internet thinks it is Last Light, the other half goes for Last Night. There was quite a bit of confusion about this. Could be just a typo.

 

[SHAQ]

So we have HD7970 rumored to be up to 50% faster than the 6970 and now GTX780 being > 100% faster than GTX580? So what's that supposed to make the 7970, a 6850 of the next generation?

HD6970 (Base = 100%) --> GTX580 (115-120%) --> HD7970 (150%) --> GTX780 (100% faster than GTX580 --> 230-240%)

So that would make a GTX780 effectively 53%-60% faster than HD7970?



For real? You are telling me in 1 generation NV will surpass AMD by 1 entire generation? :whiste:

I would expect it to be close to twice as fast as the 580. Every new gen matches or exceeds the old dual GPU card. The 285 has 240 shaders and the 580 has 512. Since 28nm is more of a shrink the 680 should have 1024 shaders so at least part of the card will be twice as fast. Also since 28nm is rumored to clock higher it could be over twice as fast in at least shader bound games. Anything less than 80% faster than a 580 would be a disappointment for the flagship card. Now Nvidia is moving to a 2 year cycle with a midlife refresh so that 1024 shader card may not come out until the refresh and we will have to settle for the rumored 768 shader card at launch which might only be 50%+ faster than a 580.