NVIDIA preparing four Maxwell GM204 SKUs (VideocardZ via S/A)

[n0x1ous]

Nope, you are the wrong. A quick 10 second Google search shows you that the 4890 was larger and had more transistors. And you can label AMD's actions as evolution or incremental or whatever you want. The fact remains unchanged, AMD released bigger and faster flagship replacements on each of the last 3 nodes. And NO ONE pitched a fit like people did over gk104 and now with the upcoming gk204. Hypocrisy at its finest.

it had that decap ring and a couple million more trannies. negligible and just there to help it achieve higher clocks - it was the same chip and you know it.

from anandtechs 4890 review...
"For the Radeon HD 4890 our hardware specs are pretty simple. Take a 4870 1GB and overclock it. Crank the core up 100 MHz to 850 MHz and the memory clock up 75 MHz to 975 MHz. That's the Radeon HD 4890 in a nutshell. However, to reach these clock levels, AMD revised the core by adding decoupling capacitors, new timing algorithms, and altered the ASIC power distribution for enhanced operation. These slight changes increased the transistor count from 956M to 959M. Otherwise, the core features/specifications (texture units, ROPs, z/stencil) remain the same as the HD4850/HD4870 series."

these AMD updates on a given node are usually like 20-25mm2 larger. nvidia "mid" to "big" die is typically +200mm2 - an exponential difference.

 

[tviceman]

@tviceman
re-chip-sales:GK104 destroyed GK106 on desktop.

I dont even have to look (again ) that Steam chart to know that GTX 670,760 & Co. are much more prevalent than GTX 660/650 Ti/Boost.

Compared to GF104/114 GK106 was a letdown imho

You are probably right, but gk106 didnt come in nearly as many variants as gk104, it was never rehashed (which helps revitalize sales), and I didn't think it was ever as attractively priced vs. the competiton as gk104's various cards.

 

[tviceman]

it had that decap ring and a couple million more trannies. negligible and just there to help it achieve higher clocks - it was the same chip and you know it.

30 mm2 bigger is not the same chip. It's not "exactly the same" like you said in your previous post, and it's not physically the same. It might be based off the same design, but it's not the same. It is noticeably larger and has more transistors.

This is seriously a point you cannot possibly win.

 

[n0x1ous]

30 mm2 bigger is not the same chip. It's not "exactly the same" like you said in your previous post, and it's not physically the same. It might be based off the same design, but it's not the same. It is noticeably larger and has more transistors.

This is seriously a point you cannot possibly win.

yet it has the exact same amount of ROPS/Stream Processors/ etc etc....so the larger die size and transistors are there to help it achieve higher clocks and thats about it.

 

[f1sherman]

yet it has the exact same amount of ROPS/Stream Processors/ etc etc....so the larger die size and transistors are there to help it achieve higher clocks and thats about it.

yeah but you responded regarding technicality (same chip), not necessarily disagreeing with him, but you are the one who's technically incorrect

just saying, sorry for bugging in

all in all hopefully we all learned something, so it's all good

 

[tviceman]

yet it has the exact same amount of ROPS/Stream Processors/ etc etc....so the larger die size and transistors are there to help it achieve higher clocks and thats about it.

So you're agreeing (without admitting to it) that it's a different chip. Ok. Keep backpedaling and saying whatever you need to say to make it work in your brain. And also keep buying / judging video cards based on die sizes and whats on the horizon at some unknown point in time. If GK104 was mid-range and if GM204 is going to be considered mid-range, then RV770 was mid-range, Cypress was mid-range, and Tahiti was mid-range. It's one big mid-range buffet and people are only complaining about one side having a longer line.

 

[blackened23]

Every chip ever made is mid range. You heard it here first. Nevermind yield issues or issues with putting 5000+ CUDA cores on 28nm (not possible), it's always mid range. ALWAYS. Tahiti was mid range. Cypress was mid range. GK104 was mid range...for a year. Now GK110 is mid range ...well it will be once Maxwell 2nd gen is released. Hawaii is the new mid range. Lets continue to argue semantics that are meaningless! We'll just ignore the real limitations of the transistors that are feasible on 28nm or yield issues. Nope. Always mid range. LOL.

 

[KaRLiToS]

Every chip ever made is mid range. You heard it here first. Nevermind yield issues or issues with putting 5000+ CUDA cores on 28nm (not possible), it's always mid range. ALWAYS. Tahiti was mid range. Cypress was mid range. GK104 was mid range...for a year. Now GK110 is mid range ...well it will be once Maxwell 2nd gen is released. Hawaii is the new mid range. Lets continue to argue semantics that are meaningless! We'll just ignore the real limitations of the transistors that are feasible on 28nm or yield issues. Nope. Always mid range. LOL.

I really don't see it that way.

GK110 will never be a mid-range chip, and it never was.

 

[HurleyBird]

yeah but you responded regarding technicality (same chip), not necessarily disagreeing with him, but you are the one who's technically incorrect.

He is technically incorrect as I pointed out earlier, but tviceman is the one here who has the really dumb opinion. HD 4890 was hardly more significant than a stepping change -- functionally at least, it's identical. Compared to the current practice from both sides of releases multiple generations of chips on the same node it's apples and oranges.

 

[n0x1ous]

I really don't see it that way.

GK110 will never be a mid-range chip, and it never was.

exactly. cypress to cayman to tahiti were all new architectures and the flagship of their generation......the 550mm2 nvidia dies (that normally launch first, but came last this gen) are the same family/arch as their "mid-range" brethren.

tvice can keep pretending its the same but its not....this is my final post on this topic. good day.

 

[RussianSensation]

Nope, you are the wrong. A quick 10 second Google search shows you that the 4890 was larger and had more transistors. And you can label AMD's actions as evolution or incremental or whatever you want. The fact remains unchanged, AMD released bigger and faster flagship replacements on each of the last 3 nodes. And NO ONE pitched a fit like people did over gk104 and now with the upcoming gk204. Hypocrisy at its finest.

First, many people criticized AMD when they launched the 7970 for $550. Second, there is no hypocrisy. Even though AMD released midrange die chips, their price was low (4870 for $299, 5870 for $379, but really one could just buy a 5850 for $259+OC, and unlocked 6950 for $299). Then came 7970 at $550 but AMD again gave great value with $399 290. Excluding 7970, AMD provided unbeatable price/performance by delivering 80-90% of NV's flagship chip at good prices. The trend has been that it was AMD which continuously forces NV to do very large price drops and it is AMD that constantly establishes new price/performance benchmark SKUs. Whenever NV offered something cheaper with similar performance, it was by $30-50 only (680 and 760). GK104 vs. Tahiti is the exception, not the rule. AMD could very well start selling 25-35% faster "flagship" for $500-550 due to market changes as well, which is why people are saying that prices are going up / it's taking longer to get 2x the performance/$ increase if AMD and Nv will start bifurcating a generation.

Obviously it is logical now that both AMD and NV are reluctant to go back to the past since they saw people spend $500-550 on midrange chips. the strategy of splitting a generation into 2 halves could make more sense going forward since shrinks to future nodes are harder to come by and more expensive. This way AMD and NV could have new and exciting products every 12-18 months by bringing 25-40% 'halves' so to speak. All of this is speculation for Maxwell of course. 880Ti 25-30% faster at $550 is still very good, just not as good as in the past. And also using 780Ti's price is somewhat trivial since we know how overpriced that card is at the moment.

------

I'll ask this instead for all those people who say midrange Vs. High end doesn't matter. When was the last time any next generation NV flagship beat the previous generation flagship by only 25-30% on average? It has never happened (GTX680 is not a true successor to 580 so this doesn't count since we know GK110 is the real successor). Every new flagship from NV beat the last gen by about 45-100%.

http://www.computerbase.de/2013-12/grafikkarten-2013-vergleich/10/

And

http://www.computerbase.de/2011-10/bericht-grafikkarten-evolution/3/

7900GTX --> 8800 GTX = at least 100% faster, 300% faster when using DX9 games

8800GTX --> GTX280 = 61% faster

GTX 280 --> 480 = 49% faster (67% faster to 580 and almost all 480s could overclock to 580 speeds)

580 --> 780Ti = 94-109%

Is we go past that, the leap from GeForce 3 Ti500 to 4600 or 4600 to 5900U or 5900U to 6800U were all greater than 35%, more like 50-100%.

Now some people are saying that midrange vs. High end doesn't matter at all since whatever is the next gen chip automatically defines a next generation??? So it is a real generational flagship jump should 880Ti be only 25-35% faster than 780Ti?

Not to mention that a next gen flagship provides 80-90% of the performance of previous gen SLI cards too. I guess we should no longer expect 50-100% generational improvements???

What some of us are saying is that NV is still going to give us 50-100% generational increases but they could just split the generation into 2 halves. This way they have new faster products to sell every 12-15 months instead of waiting every 24-30 months for real flagships to come out. Is this bad? Maybe for AMD and NV this new way to launch generations makes sense but let's not kid ourselves: a next generation flagship that's only 25-35% faster is NOT a next gen flagship if we use 15 years of GPU history. Even Kepler vs. Fermi shows that this trend still holds.

As already stated, if NV decides to use 16nm/20nm for GM210 (or w/e it's called) then releasing a 28nm GM204 makes sense. NV may also think that AMD will not launch anything 50-60% faster than 290X because they are also power consumption constrained on the same 28nm node. From a business perspective, it would be more profitable selling a 430mm2 chip at $500-550 on a very mature node if NV feels that in the next 6 months AMD can't bring anything that can seriously beat the 880Ti.

It is not just about AMD vs. NV, or the definition if a midrange chip, but the hypothesis that AMD and NV will bifurcate a generation and in turn sell us midrange next gen performance (if people don't like the idea of calling it midrange chip) for $500-550. I think that's the main point that's being missed here. Alternatively, we have to accept that next generation flagships should only be 25-35% faster and throw the historical 50-100% out the window.

P.S. And the comments before that 7970 was only 40% faster than 6970 tell half the story. First, an after market 7970 hits 1125-1175mhz with ease. Second, if a gamer looks at 7970Ghz/280X benches vs. 6970, in modern games the increase is 65-90%. The only reason this doesn't seem impressive is because how much better NV executed with Kepler (580->780ti) and because of how weak 6970 is now vs. 580 due to heavy use of tessellation in modern games.

 

[PrincessFrosty]

So basically we have to wait yet another generation of products for a meaningful jump in technology, I've been putting off a card upgrade since my 580s waiting for a real jump not just rebranding and relatively minor increases in speed, looks like I'll probably end up waiting another round, sigh.

 

[RussianSensation]

So basically we have to wait yet another generation of products for a meaningful jump in technology, I've been putting off a card upgrade since my 580s waiting for a real jump not just rebranding and relatively minor increases in speed, looks like I'll probably end up waiting another round, sigh.

Well in your case the increases are cumulative since you skipped 680/780Ti. Since aftermarket 780Ti is 100% faster than a 580, 30% over 780Ti is 130% faster than your card. That's a solid upgrade for you, and more so if your 580s are the 1.5GB variants. If GM210 is 35% faster than 880Ti, it will be 175.5% faster than a 580. I think for you it's probably better to get a card 130% faster than wait yet another year for a 175% card. In other words, how impressive 880Ti looks will depend on what you feel is a good enough performance increase (35%, 50%, 75%, 100%, 200%, etc.)

 

[HurleyBird]

Well in your case the increases are cumulative since you skipped 680/780Ti. Since aftermarket 780Ti is 100% faster than a 580, 30% over 780Ti is 130% faster than your card.

160%

 

[RussianSensation]

160%

You are right. Thanks for the correction (2x faster than 580 x 1.30
= 260% of 580s performance or 160% faster). Brain lapse

 

[SirPauly]

When was the last time any next generation NV flagship beat the previous generation flagship by only 25-30% on average?

It has happened with the 8800 ultra and the 9800 GTX -- a smaller die with a high end naming convention. The G-92 strategy was very similar to the Gk-104!

The constructive nit-pick with the introduction of 28 sku's from nVidia and AMD was this to me:

Even though the node and arches were more-so substantial and significant the price performance felt more-so incremental and evolutionary!

 

[SirPauly]

And this isn't new nVidia offering smaller dies for flag ship naming -- with the 7900 family and the 7900 GTX, the die was remarkably small compared to the 1900 XTX from ATI.

 

[Wild Thing]

First, many people criticized AMD when they launched the 7970 for $550. Second, there is no hypocrisy. Even though AMD released midrange die chips, their price was low (4870 for $299, 5870 for $379, but really one could just buy a 5850 for $259+OC, and unlocked 6950 for $299). Then came 7970 at $550 but AMD again gave great value with $399 290. Excluding 7970, AMD provided unbeatable price/performance by delivering 80-90% of NV's flagship chip at good prices. The trend has been that it was AMD which continuously forces NV to do very large price drops and it is AMD that constantly establishes new price/performance benchmark SKUs. Whenever NV offered something cheaper with similar performance, it was by $30-50 only (680 and 760). GK104 vs. Tahiti is the exception, not the rule. AMD could very well start selling 25-35% faster "flagship" for $500-550 due to market changes as well, which is why people are saying that prices are going up / it's taking longer to get 2x the performance/$ increase if AMD and Nv will start bifurcating a generation.

Obviously it is logical now that both AMD and NV are reluctant to go back to the past since they saw people spend $500-550 on midrange chips. the strategy of splitting a generation into 2 halves could make more sense going forward since shrinks to future nodes are harder to come by and more expensive. This way AMD and NV could have new and exciting products every 12-18 months by bringing 25-40% 'halves' so to speak. All of this is speculation for Maxwell of course. 880Ti 25-30% faster at $550 is still very good, just not as good as in the past. And also using 780Ti's price is somewhat trivial since we know how overpriced that card is at the moment.

------

I'll ask this instead for all those people who say midrange Vs. High end doesn't matter. When was the last time any next generation NV flagship beat the previous generation flagship by only 25-30% on average? It has never happened (GTX680 is not a true successor to 580 so this doesn't count since we know GK110 is the real successor). Every new flagship from NV beat the last gen by about 45-100%.

http://www.computerbase.de/2013-12/grafikkarten-2013-vergleich/10/

And

http://www.computerbase.de/2011-10/bericht-grafikkarten-evolution/3/

7900GTX --> 8800 GTX = at least 100% faster, 300% faster when using DX9 games

8800GTX --> GTX280 = 61% faster

GTX 280 --> 480 = 49% faster (67% faster to 580 and almost all 480s could overclock to 580 speeds)

580 --> 780Ti = 94-109%

Is we go past that, the leap from GeForce 3 Ti500 to 4600 or 4600 to 5900U or 5900U to 6800U were all greater than 35%, more like 50-100%.

Now some people are saying that midrange vs. High end doesn't matter at all since whatever is the next gen chip automatically defines a next generation??? So it is a real generational flagship jump should 880Ti be only 25-35% faster than 780Ti?

Not to mention that a next gen flagship provides 80-90% of the performance of previous gen SLI cards too. I guess we should no longer expect 50-100% generational improvements???

What some of us are saying is that NV is still going to give us 50-100% generational increases but they could just split the generation into 2 halves. This way they have new faster products to sell every 12-15 months instead of waiting every 24-30 months for real flagships to come out. Is this bad? Maybe for AMD and NV this new way to launch generations makes sense but let's not kid ourselves: a next generation flagship that's only 25-35% faster is NOT a next gen flagship if we use 15 years of GPU history. Even Kepler vs. Fermi shows that this trend still holds.

As already stated, if NV decides to use 16nm/20nm for GM210 (or w/e it's called) then releasing a 28nm GM204 makes sense. NV may also think that AMD will not launch anything 50-60% faster than 290X because they are also power consumption constrained on the same 28nm node. From a business perspective, it would be more profitable selling a 430mm2 chip at $500-550 on a very mature node if NV feels that in the next 6 months AMD can't bring anything that can seriously beat the 880Ti.

It is not just about AMD vs. NV, or the definition if a midrange chip, but the hypothesis that AMD and NV will bifurcate a generation and in turn sell us midrange next gen performance (if people don't like the idea of calling it midrange chip) for $500-550. I think that's the main point that's bring missed here. Alternatively, we have to accept that next generation flagships should only be 25-35% faster and throw the historical 50-100% out the window.

P.S. And the comments before that 7970 was only 40% faster than 6970 tell half the story. First, an after market 7970 hits 1125-1175mhz with ease. Second, if a gamer looks at 7970Ghz/280X benches vs. 6970, in modern games the increase is 65-90%. The only reason this doesn't seem impressive is because how much better NV executed with Kepler (580->780ti) and because of how weak 6970 is now vs. 580 due to heavy use of tessellation in modern games.

Very nice summation and analysis Mister Russian.
You da man!:thumbsup:

 

[RussianSensation]

It has happened with the 8800 ultra and the 9800 GTX -- a smaller die with a high end naming convention. The G-92 strategy was very similar to the Gk-104!

9800GTX is not a real new generation vs. 8800 Ultra so this example doesn't apply at all. Everyone knows that the successor to 8800 Ultra was GTX280, not 9800GTX/9800GTX+. The 9800GTX/GTX+ chips were refreshes of G80 architecture only as a result of a die shrink (G80->G92). GM204 is a completely new architecture (Kepler vs. Maxwell) which doesn't even apply to the 8800 U vs. 9800GTX comparison. NV should have never even called 9800GTX/+ a GeForce 9. This was more like 8850GTX.

And this isn't new nVidia offering smaller dies for flag ship naming -- with the 7900 family and the 7900 GTX, the die was remarkably small compared to the 1900 XTX from ATI.

As I said, put aside the mid-range vs. high-end die and focus on performance instead. Let's talk about the GeForce 7 family.

GeForce 7800GTX 256MB => 54% faster than 6800U
GeForce 7800GTX 512MB => 2x faster than 6800U
GeForce 7900GTX 512MB => 128% faster than 6800U
Source

You see, whenever NV replaced a previous generation flagship with a real new generation flagship (whether they simply used more CUDA cores/units as in G80 8800 U --> GT200 GTX280 OR whether they replaced an architecture GT200/b GTX280/285 --> GTX480/580), the performance increase is usually 50-100%, never 25-35%. By loose definition of GPU industry in the last 15 years, every new real generation resulted in a performance increase of 50-100% until AMD/NV started making up generations such as GeForce 9 (completely made up generation) and HD6970 (this isn't a real next generation vs. 5870). Starting with 7970/R9 290X and 680/780Ti, both AMD and NV split up a generation into 2 halves, essentially milking the mid-range for much longer by sustaining much higher ASPs. $450 HD7950 anyone?

I mean if GM204 launches with 25-35% more performance and then 1 year later GM200/210 brings another 30-40%, it will be obvious that GM204 is the spiritual successor to GK104 and GM200/210 is the real successor to GK110. That's the point: Splitting the gen into 2 halves could very well be the new trend in GPUs as a result of more expensive die shrinks and longer time to market for lower nodes.

 

[BrightCandle]

I think splitting the generations to some extent makes sense. Both the 290 and the 780 are really big cores, the 780 being very close to the absolute maximum possible due to lithography limits and the 290 is smaller but still very big historically. Nvidia has been beaten to market multiple times by AMD because its wanted to lead with big cores and found significant problems producing them on a new process. A new architecture and a new process all at the same time is prone to delays if things don't go to plan.

So leading with the mid range means you can still get saleable yields from the smaller cores when the process isn't great and then as the process matures you can put out the bigger design that needs a lower error rate to be profitable in the consumer market. With prices per wafer going up the alternative plan of leading with the big core doesn't make any sense, they would be ludicrously expensive.

The only caution I have about GM200/210 performance is that its not on a new process, so they have the same limits on total transistors as they did before. A focus on reducing the power consumption and improving efficiency will only get them so far and in the past when they did this (the 580 over the 480, the 6970 over the 5870 etc) the performance gains have been more moderate. There is a reasonable chance this architecture will on this process only really bring moderate gains, like that of a refresh on the existing process rather than the full on gains because they don't have a density increase. It would be impressive and a little embarrassing if they did manage to get 2x performance from the same process node within the same limits.

 

[jpiniero]

25-35% more performance than what? The 780 Ti? That's going to be tough.

 

[96Firebird]

The only caution I have about GM200/210 performance is that its not on a new process, so they have the same limits on total transistors as they did before. A focus on reducing the power consumption and improving efficiency will only get them so far and in the past when they did this (the 580 over the 480, the 6970 over the 5870 etc) the performance gains have been more moderate. There is a reasonable chance this architecture will on this process only really bring moderate gains, like that of a refresh on the existing process rather than the full on gains because they don't have a density increase. It would be impressive and a little embarrassing if they did manage to get 2x performance from the same process node within the same limits.

Just my opinion, but I don't think they will release big Maxwell on 28nm. I don't think they can get a big enough boost in performance for workstation cards to sell (where they make the real money). They designed the chip to be on 20nm, and scaling it back to the 28nm process seems like it would cut too much from the original design.

 

[Rezist]

Obviously cards aren't getting much faster, and the rate which there releasing is getting slower. Atleast no one can complain that PC's are costing more then consoles since there slowing down so much, you wont have anything to upgrade to.

 

[tviceman]

First, many people criticized AMD when they launched the 7970 for $550. Second, there is no hypocrisy. Even though AMD released midrange die chips, their price was low (4870 for $299, 5870 for $379, but really one could just buy a 5850 for $259+OC, and unlocked 6950 for $299). Then came 7970 at $550 but AMD again gave great value with $399 290.

I replied to you but the forum apparently didn't accept it. Doesn't matter. I agreed with some of what you said and disagreed with other things you said. Still doesn't matter. What we are arguing about - the product placement of a GPU - doesn't matter one iota. For anyone here who's criteria to buying a GPU is greater than a particular die size, company code-name correlation, or thinks it'll be displaced relatively quickly, sit on the side lines if your criteria isn't met. It's really that simple.

 

[3DVagabond]

Just don't want people to be fooled by the name (gtx-880) into believing this is the GK110 replacement.