[ PCGH ] Maxwell GTX 880 specifications leaked

[RussianSensation]

The bottom line is that the stats provided for the alleged GTX 880 make no sense. It would be ridiculous to pair that large a die, with that many shaders, with so few ROPs and such a narrow memory bus. It would be the opposite of future-proof; it would hurt performance in an area where the emphasis is clearly going to increase in the future.

The 880 specs state it's a 20nm chip with 7.9B transistors. That likely translates to a chip 340-375mm2, far from a large chip. The point you make about where the chip rests in the line, I don't necessarily agree with. Using your logic, had NV released GTX460 1GB before the 480, since it beat GTX285, you would have called it the flagship for the time being?

We suspected all along when 680 launched that it was NV's mid-range chip and we were proven right. 680 was only a flagship by virtue of 7970 being so much slower than GK110. On NV's product stack, GK104 was never really the flagship. They just got away with that strategy since Kepler was such a break-through compared to Fermi. 780Ti is ~ 2x faster than GTX580, the largest generational jump since 7900GTX --> 8800GTX:
http://www.computerbase.de/2013-12/grafikkarten-2013-vergleich/10/

The move from 8800GTX to 280/285 or from 285 to 480/580 was far smaller. That's why NV got away with 'marketing' 680 as the flagship.

If they repeat the same strategy, the code-name, the die size and the performance increase over 780Ti will solidify if the 880 is a mid-range Maxwell or flagship even if NV markets it as 880.

 

[toyota]

a low clocked mid range chip on 20nm with efficient Maxwell architecture rated for 230 watts is completely illogical.

 

[Ajay]

I've got a feeling that AMD will be using HBM for Pirate Islands similar to AMD using GDDR5 for R700.

While that would be amazing - no, I don't see it happening. Highly doubt AMD can get TSVs (stacked DRAM) even by 1H 2015. Intel's KC (14nm Xeon Phi) is using on-package HBM and it looks like they are at least two years ahead of everyone else.

 

[MrK6]

Kind of lackluster compared to Pirate Islands "leaked specs." Who knows if either is close, but considering the meager performance improvement I would have expected a better power profile considering Maxwell's successes on 28nm.

 

[moonbogg]

Could the argument be made that the 7970 was AMD's mid range chip? They did after all launch a larger die 28nm chip to compete with GK110 in the form of the R9 series. So both companies launched small die 28nm first, then larger dies later.

 

[Keysplayr]

a low clocked mid range chip on 20nm with efficient Maxwell architecture rated for 230 watts is completely illogical.

Kind of. Imagine a Titan Black at 250W on 28nm. Now shrink it to 20nm, and enhance the core with Maxwell architecture. Anyone here think it'll only consume 20W less than a Titan Black?

 

[Ajay]

Kind of lackluster compared to Pirate Islands "leaked specs." Who knows if either is close, but considering the meager performance improvement I would have expected a better power profile considering Maxwell's successes on 28nm.

Considering the efficiency of the GTX 750 Ti @ 28nm, there is no way the Power numbers are correct - or anything else for that matter. That's the thing with these sites; throw enough crap at the wall and eventually so will stick - and then they can say they were right :whiste:

 

[Silverforce11]

You guys keep mentioning the efficiency of the 750 ti as if you don't realize its an ultra low end compute-castrated chip made for that sole purpose. Any expectations that big Maxwell made for the HPC sector first and foremost to retain the same gaming efficiency is highly optimistic.

Let's face it, leak specs on an unknown architecture, you could see it however you wish, but we just don't know enough.

 

[tweakboy]

Will a 880 GTX be faster then 2x 790 SLI ?

Let me guess the 880 blows awawy current SLI configs.

 

[Keysplayr]

You guys keep mentioning the efficiency of the 750 ti as if you don't realize its an ultra low end compute-castrated chip made for that sole purpose. Any expectations that big Maxwell made for the HPC sector first and foremost to retain the same gaming efficiency is highly optimistic.

Let's face it, leak specs on an unknown architecture, you could see it however you wish, but we just don't know enough.

What do you mean we don't realize? And more to the point, why don't you realize that the same technique Nvidia used on GM107 could probably just as easily be used on bigger chips. What source do you have that lets you say 750Ti was a "one off" and the techniques will not be passed on to other Maxwell models? 28nm or 20nm alike?
That is why the efficiency of the 750Ti keeps being mentioned and WILL be mentioned a ton more while we wait for other Maxwell chips to emerge. Because people think that, I don't know, maybe, other Maxwell chips will be designed similarly.

Do not confuse my comment here to mean that I accept any of the leaks we have been hearing lately. Especially from the sources. I do not.

 

[Techhog]

Could the argument be made that the 7970 was AMD's mid range chip? They did after all launch a larger die 28nm chip to compete with GK110 in the form of the R9 series. So both companies launched small die 28nm first, then larger dies later.

Tahiti clearly wasn't intended to be midrange like GK104 was though.

 

[Silverforce11]

What do you mean we don't realize? And more to the point, why don't you realize that the same technique Nvidia used on GM107 could probably just as easily be used on bigger chips. What source do you have that lets you say 750Ti was a "one off" and the techniques will not be passed on to other Maxwell models? 28nm or 20nm alike?
That is why the efficiency of the 750Ti keeps being mentioned and WILL be mentioned a ton more while we wait for other Maxwell chips to emerge. Because people think that, I don't know, maybe, other Maxwell chips will be designed similarly.

Do not confuse my comment here to mean that I accept any of the leaks we have been hearing lately. Especially from the sources. I do not.

I just told you, none of us knows anything to speculate further. It's just a guessing game.

 

[Keysplayr]

I just told you, none of us knows anything to speculate further. It's just a guessing game.

How hard is it to speculate that other Maxwell skus will share similar designs with GM107?



Not very hard.

 

[Keysplayr]

Tahiti clearly wasn't intended to be midrange like GK104 was though.

Yes, GK104 and GK110 are both Kepler architecture.

 

[blackened23]

Considering the efficiency of the GTX 750 Ti @ 28nm, there is no way the Power numbers are correct - or anything else for that matter. That's the thing with these sites; throw enough crap at the wall and eventually so will stick - and then they can say they were right :whiste:

While you obviously can't discern everything with specifications alone, this is quite true. Maxwell as an architecture has twice the performance per watt than Kepler. ON 28NM. Yet these specifications would indicate a 230W TDP card with more transistors and an incremental spec increase.

Fake. It's already been confirmed as fake anyway. There's simply no way this could be a 230W TDP variant of maxwell, given the performance per watt of the architecture itself.

 

[Ajay]

Let's face it, leak specs on an unknown architecture, you could see it however you wish, but we just don't know enough.

Unknown??? We already know the fundamental architectural changes from Kepler to Maxwell. For gaming GPUs, there won't be much changed from the GM107 (just more CCs, ROP, cache, etc. - which are really implementation variations). There will be more differences in NV's 'big die' HPC GPU - I think some details can be inferred docs at Nvidia's CUDA. Dev site, but I'm not at my computer right now.

Anyway, we are not totally in the dark and the GTX 750 serves well as a pointer to things to come.

 

[Cloudfire777]

Direct actual translation from the leaker. The clock speeds and TDP in the article this thread links to is BS

GTX 880, which is GM104, will be the first flagship from Nvidia's Maxwell architecture. Single 6-pin power supply and it is rated at 150W TDP. The card will use the same building materials as GTX 660, but the PCB will be a little bit shorter. Performance will be about 15-20% higher than GTX Titan.
GTX 880 will be built with TSMC's 20nm process but due to slow progress with 20nm, the first quarter have managed to produce 1000pcs of GM104 chips. Expect this production to increase during this quarter and the upcoming quarters.

GTX 880 specifications: 3200 CUDA cores, GM104 256bit 4GB GDDR5 VRAM, and it integrates an ARM CPU

GTX 880 will launch in June 2014

 

[Silverforce11]

Unknown??? We already know the fundamental architectural changes from Kepler to Maxwell. For gaming GPUs, there won't be much changed from the GM107 (just more CCs, ROP, cache, etc. - which are really implementation variations). There will be more differences in NV's 'big die' HPC GPU - I think some details can be inferred docs at Nvidia's CUDA. Dev site, but I'm not at my computer right now.

Anyway, we are not totally in the dark and the GTX 750 serves well as a pointer to things to come.

This is my point several times now which you guys choose to ignore and go on your merry way assuming a ultra low end chip designed purely for efficiency is somehow going to be representative at the high end. We just don't know.

It comes down to how long a bowstring do you want to pull? Because its a definite that "There will be more differences in NV's 'big die' HPC GPU"..

 

[f1sherman]

This is my point several times now which you guys choose to ignore and go on your merry way assuming a ultra low end chip designed purely for efficiency is somehow going to be representative at the high end. We just don't know.

It comes down to how long a bowstring do you want to pull? Because its a definite that "There will be more differences in NV's 'big die' HPC GPU"..

Within same architecture - basic metrics are always quite similar in-between particular chip implementations, and even superposition principle works quite well.
So yeah I'd say we know much about Maxwell's efficiency

If efficiency was everything they cared about, 750/Ti would come with lower clocks (like in M versions) and combined with binning would proly doubled on its efficiency (~30W).

Even so, particular chip implementation can take you only so far - it's architecture itself that IS focused on efficiency.

Notice there is no difference between ultra low end and HIGH END Kepler (with NV anyway). If anything Titan is almost at top of Kepler's perf/W.

 

[Techhog]

Direct actual translation from the leaker. The clock speeds and TDP in the article this thread links to is BS

GTX 880, which is GM104, will be the first flagship from Nvidia's Maxwell architecture. Single 6-pin power supply and it is rated at 150W TDP. The card will use the same building materials as GTX 660, but the PCB will be a little bit shorter. Performance will be about 15-20% higher than GTX Titan.
GTX 880 will be built with TSMC's 20nm process but due to slow progress with 20nm, the first quarter have managed to produce 1000pcs of GM104 chips. Expect this production to increase during this quarter and the upcoming quarters.

GTX 880 specifications: 3200 CUDA cores, GM104 256bit 4GB GDDR5 VRAM, and it integrates an ARM CPU

GTX 880 will launch in June 2014

June? AHAHAHAHAHAHA

 

[Ajay]

This is my point several times now which you guys choose to ignore and go on your merry way assuming a ultra low end chip designed purely for efficiency is somehow going to be representative at the high end. We just don't know.

It comes down to how long a bowstring do you want to pull? Because its a definite that "There will be more differences in NV's 'big die' HPC GPU"..

God give me strength! First off, I'm an not "you guys". Secondly, I make the assumption the the 880 is a mid range card (like the 680) and hence will scale in a fairly predictable way compared to the GM107 (albeit, on 20nm instead of 28 - so even more efficient unless TSMC or NV blows it).

Lastly, I singled our NV's HPC GPU as being a somewhat different case which will depend on the actual changes in CUDA Compute 5.0 and their implementation. I haven't read any of the docs @ https://developer.nvidia.com/maxwell-compute-architecture, because I'm no where near playing with that hardware - but I'm sure they provide _some_ illumination.

If you choose to presume that you have absolutely no clue as to what improvements might be seen - then fine, just stop bashing us because we don't conform to your world view. Clearly we have a different opinion.

 

[blastingcap]

What do you mean we don't realize? And more to the point, why don't you realize that the same technique Nvidia used on GM107 could probably just as easily be used on bigger chips. What source do you have that lets you say 750Ti was a "one off" and the techniques will not be passed on to other Maxwell models? 28nm or 20nm alike?
That is why the efficiency of the 750Ti keeps being mentioned and WILL be mentioned a ton more while we wait for other Maxwell chips to emerge. Because people think that, I don't know, maybe, other Maxwell chips will be designed similarly.

Do not confuse my comment here to mean that I accept any of the leaks we have been hearing lately. Especially from the sources. I do not.

I agree with Keys. NV's designs for the last several years have been quite modular (the smallest complete unit being the SMX for Kepler) and I expect there to be pretty good scaling. Not perfect scaling, because nothing is ever 100.00% perfect and there may be bottlenecks here or there, but pretty good scaling. And people who don't want to speculate should know that other people enjoy speculating so let them have their fun in this thread and go to another one you like more instead.

 

[moonbogg]

June? AHAHAHAHAHAHA

I called it first as an April fool's joke. So if its true, I get all the props, ok?

 

[Bubbleawsome]

Wow guys, you lasted through 3 pages of non-caps, non-bolded text. Good job. :|
Might as well let the masses in now.

 

[Tristor]

If these specs are legit, then this is a repeat of Kepler and I'll be annoyed. This is going to hurt their sales with the people who actually want to build at the top end. What motivation do I have to upgrade if this is all they're offering for Maxwell out the gate? I've been planning a Haswell-E/DDR4/Maxwell build for 2015 for doing 4k Project CARS, but what's the point if this is it?