[guru3d rumor] GTX 880- 8GB-3200 SP's

[dn7309]

Seeing that Nvidia was still to cheap to include 3 Gb of Ram when competing against the 7970 and 4 gb of ram when competing against the 290, I don't see how they suddenly make the jump up to 8 Gb of ram in one generation.

 

[96Firebird]

Maybe they got tired of people complaining?

 

[Mondozei]

This sounds like another GTX680 (GK104) midrange product. Just like the 6xx and 7xx series, GM104 will kick off the 8xx series and GM110 will arrive on 20nm somtime next year or by years end. GM104 will get promoted to GTX970 just as GK104 got promoted to GTX770. Anyways, thats what I think will go down.

The GTX 800-series will probably be a re-run of the GTX 600-series. Mediocre and overpriced.
The big news will come in the 900-series on 20 nm(either with or without FinFETs/Tri-gate).
GM210 is the one we're waiting for and that'll be due for the DX12 launch, or slightly before it, so next year by H2.

It's kind of ironic that we had big Kepler almost 1 and 1/2 years ago and we're still looking at 12-18 months until we get big Maxwell.

The bright side, if you can call it that, is that the average quality of people's GPUs should rise and become relatively evened out, i.e. a high base as people migrate to the 700 and 800 series who are now on HD5000 series or in the GTX 200-500 range.

Another good tidbit could be that both companies get their drivers into much better shape. AMD has done great work on Crossfire if you compare it today to just 18 months ago or so. And the reason is that if you want to go 4K with smoothness and high fps, you need SLI/Crossfire and you need good scaling since single GPUs are not advancing fast enough for 2160p60. So AMD/Nvidia are forced to do a much better job than they did historically. AMD especially. And this benefits people like me who are planning to get a second or a even a third Tri-X 290 and a 4K monitor the coming year.

But of course, I'd have preferred to get much stronger single GPUs instead.

 

[JoeRambo]

The specs make perfect sense for me, you guys are too fixated on ROP counts and memory bus width. This card is not targeting 4k, that is the job for big Maxwell.

So once that ridiculous idea of building a card for non yet existing market of 4k is gone, we are left with a card that has enough ROPs to pump pixels for 1080P and barely enough for 1440P.

It's even more clear when you think about it in terms of 580 -> 680 and 780 -> 880.
580 has 48 ROPs and mem with 384-bit, yet 680 had no trouble being faster than it even in syntetic pixel fill rate. I see no reason why that can't happen with new architecture (even if same mfg node) again.

If anything, the trend is for flops/pixel to rise, and with games like Witcher 3, Crysis/UE4 engine stuff etc incoming that trend will only speed up.

P.S. I am certain both AMD and Nvidia will have products for enthusiasts running high hz on 1440P or 4K, just that they probably can't be built on 28nm.

 

[Vesku]

Going to be a lackluster launch if the 880 performs worse at high resolutions than the 780/780 Ti.

 

[Keysplayr]

Nm

 

[Anarchist420]

The reason the 256 bit cards seem so much faster is because nv doesnt use a lossless z-buffer if SGSSAA isnt forced. 256 bit 7.4 GHz is nowhere close to 384 bit 6 Ghz and the IP system REALLY needs to be abolished if they are going to charge $700 for that. someone else could use their IP, make a better product with it, charge 1/3 the price for it, and that would force nvidia to compete. same thing with intel.

This is VC&G, not P&N. Stay on topic.
-- stahlhart

 

[OCGuy]

Maybe they got tired of people complaining?

They just announced that they beat Wall Street estimates for 2Q in a row due to GTX sales....that probably speaks louder than anonymous messages on the internet made mostly by people who don't have the disposable income to be in the market, anyway.

 

[Ajay]

This sounds like another GTX680 (GK104) midrange product. Just like the 6xx and 7xx series, GM104 will kick off the 8xx series and GM110 will arrive on 20nm somtime next year or by years end. GM104 will get promoted to GTX970 just as GK104 got promoted to GTX770. Anyways, thats what I think will go down.

Totally depends are which of the posted specs are correct. If it really has 3200 CCs then it's going to be a big die GPU and will need a 384b bus to provide sufficient bandwidth (maybe even 5x the GTX 750 Ti).

If the bus bandwidth is correct, then there is no way it's a 3200 CC design because the SMMs are going to be starved for bandwidth, even with a better cache and more efficient controller.

The only reason that NV would build a new large die GPU would be for the lucrative professional/compute markets where Intel's Xeon Phi is giving them headaches. NV may need a 15%+ advantage to help it stay competitive until 20/16nm comes along.

If it's not a big die GPU, then it probably will be similar to GK104 (probably with a higher CC count) and then it will have a 256b bus with an ~7GHz clock. Oh, and cost $700, just to milk early adopters who want the latest thing.

Oh, the joys of the new PC GPU market.

 

[KaRLiToS]

They just announced that they beat Wall Street estimates for 2Q in a row due to GTX sales....that probably speaks louder than anonymous messages on the internet made mostly by people who don't have the disposable income to be in the market, anyway.

What does that have to do with the topic?

This year seems like a boring year without 20nm.

 

[Keysplayr]

What does that have to do with the topic?

This year seems like a boring year without 20nm.

Jesus folks, 20nm isn't the end all be all. I've said it before, I'd rather see innovations in architecture and super fine tuning than just continuously relying on die shrinks for performance increases and reduced power consumption.
Forget 20nm. It will get here when it gets here, and then we can all B & M about how long 16nm is taking.

Don't rush it. Let em work with what they have. Forces them to do better.

 

[KaRLiToS]

Jesus folks, 20nm isn't the end all be all. I've said it before, I'd rather see innovations in architecture and super fine tuning than just continuously relying on die shrinks for performance increases and reduced power consumption.
Forget 20nm. It will get here when it gets here, and then we can all B & M about how long 16nm is taking.

Don't rush it. Let em work with what they have. Forces them to do better.

How do you think is it possible to "super fine-tune" hawaii or GK110 ?

Do you think they can fine tune hawaii so it can run faster at a temperature of 145'C?

Even if you said it before, doesn't mean its possible. At this point, a die shrink is important for performance increase.

TSMC is slow. That's why AMD and Nvidia don't have 20nm yet.

 

[Silverforce11]

It's not Hawaii or Kepler, its going to be GCN 2.0 and Maxwell parts on 28nm.

At this point, rushing out a 20nm product that will collide with 16nm not long after is pointless.

 

[tviceman]

How do you think is it possible to "super fine-tune" hawaii or GK110 ?

Do you think they can fine tune hawaii so it can run faster at a temperature of 145'C?

Even if you said it before, doesn't mean its possible. At this point, a die shrink is important for performance increase.

TSMC is slow. That's why AMD and Nvidia don't have 20nm yet.

You are misinterpreting what he said / means. As nvidia has proven with GM107, there is plenty of room to improve both perf/mm^2 and perf/watt on the existing node. 20nm, while it will help both of those metrics in and of itself, is not absolutely needed to come out with significantly faster, more efficient GPU's working with the existing technology.

 

[KaRLiToS]

It's not Hawaii or Kepler, its going to be GCN 2.0 and Maxwell parts on 28nm.

At this point, rushing out a 20nm product that will collide with 16nm not long after is pointless.

Do you think they might skip 20nm?

You are misinterpreting what he said / means. As nvidia has proven with GM107, there is plenty of room to improve both perf/mm^2 and perf/watt on the existing node. 20nm, while it will help both of those metrics in and of itself, is not the absolutely needed to come out with significantly faster, more efficient GPU's working with the existing technology.

Got it :thumbsup:

 

[tviceman]

Totally depends are which of the posted specs are correct. If it really has 3200 CCs then it's going to be a big die GPU and will need a 384b bus to provide sufficient bandwidth (maybe even 5x the GTX 750 Ti).

If the bus bandwidth is correct, then there is no way it's a 3200 CC design because the SMMs are going to be starved for bandwidth, even with a better cache and more efficient controller.

The only reason that NV would build a new large die GPU would be for the lucrative professional/compute markets where Intel's Xeon Phi is giving them headaches. NV may need a 15%+ advantage to help it stay competitive until 20/16nm comes along.

If it's not a big die GPU, then it probably will be similar to GK104 (probably with a higher CC count) and then it will have a 256b bus with an ~7GHz clock. Oh, and cost $700, just to milk early adopters who want the latest thing.

Oh, the joys of the new PC GPU market.

It isn't the bandwidth that is suspect with these specs. Look at GK107 vs. GK104. GK107's bandwidth was 80 gb/s, while GTX 770's is 224 gb/s. The GTX 770 is about 3.33x faster than GTX 650, yet does so with only 2.8x the memory bandwidth.

Extrapolate those figures and compare them with GM107 and these rumored 3200 SP maxwell chip specs. GTX 750 TI has 86.4 gb/s bandwidth. The rumored specs of this new maxwell part point to 7.4ghz on a 256-bit bus, which would add up to 237 gb/s. So if Maxwell's various chips scale with performance in respect to Nvidia's factory-set memory speeds, this new chip has 2.75x more memory bandwidth, almost exactly the same gulf as GTX 650 and GTX 770. So, as you see, bandwidth isn't "OMG what do we do" low.

I think the biggest issue with the specs listed is the ROP's. If it's the same ROP's as GM107, then that will be where the bottleneck is. If the ROP's have been beefed up, which is plausible given the timeframe of this chip's tape out relative to GM107's tapeout, then this chip should be about as balanced as GK104 was. That is, there will still be a small bandwidth and ROP bottleneck, but nothing detrimental. But if the ROP's are relatively unchanged, then we could be looking at a weirdly unbalanced part.

Oh, and while we're at it, if this chip scales similarly in performance over GM107 like GK104 did over GK107, then it should be significantly faster than a gtx 780 TI. I don't think it will be significantly faster, not with 32 ROP's and ~240gb/s bandwidth, but I do think with the given specs it will be somewhat faster.

 

[Keysplayr]

How do you think is it possible to "super fine-tune" hawaii or GK110 ?

Do you think they can fine tune hawaii so it can run faster at a temperature of 145'C?

Even if you said it before, doesn't mean its possible. At this point, a die shrink is important for performance increase.

TSMC is slow. That's why AMD and Nvidia don't have 20nm yet.

N/M. Tviceman nailed it.

 

[Silverforce11]

Do you think they might skip 20nm?

Yup. We've only just got the big R295X2 and TitanZ is still to come. Then GM204 later this year on *28nm* its clear that 20nm will not be used by AMD/NV, since 16nm is just around the corner.

@tviceman
Given the 3200 shaders and 32 ROPs and clocks on Maxwell, it should be significantly faster than 780ti at the common resolutions of 1080p to 1600p. This is its goal since afterall, it is a mid-range chip. 4K is for the big Maxwell, like its for big GK110.

Edit: Typo, GM104/204. aka "Mid-range" Maxwell for outrageous prices.

 

[RussianSensation]

@tviceman
Given the 3200 shaders and 32 ROPs and clocks on Maxwell, it should be significantly faster than 780ti at the common resolutions of 1080p to 1600p. This is its goal since afterall, it is a mid-range chip. 4K is for the big Maxwell, like its for big GK110.

Yup. On paper, it has 3200 CUDA cores, 200 TMUs and 32 ROPs, yet the efficiency of the architecture has gone up in aggregate 35% (NV claims that 128 CUDA Maxwell cores provide 90% of the performance of 192 Maxwell cores). Apply that 35% and you are looking at Kepler "equivalent" 4320 CUDA cores.

I said it before - looking at the specs of 660 vs. 750Ti, on paper, 660 should destroy the 750Ti but it's not much faster given the dramatic difference in their CUDA cores, TMUs, ROPs and memory bandwidth. On paper, 660 is beating 750Ti by close to 50% at minimum in each of those metrics and yet the performance is only 20-21% faster!

Let's not forget that in the past we have seen many cards that couldn't fully utilize all of their available memory bandwidth - GTX285 vs. GTX460.

GTX285 = 159 GB/sec (38% greater)
GTX460 = 115.2 GB/sec
http://www.gpureview.com/show_cards.php?card1=605&card2=632

Specs on paper from different generations show time and time again that you can't just compare CUDA cores, TMU, ROP and memory bandwidth efficiency without knowing the increased IPC and peculiarities of the new architecture.

Even if we throw all of these variables and assume that GTX880 is only 15-20% faster than 780Ti, if NV prices it at $549, that will be more than enough to hold them over for another 1-1.5 years until GM210 is launched, unless of course AMD exceeds expectations with R9 390X. Secondly, GTX880 may not be aimed at 780Ti users. For 680/7970Ghz users, 20% faster than 780Ti would translate into a card that's 73-84% faster. At $550 that would sell like hot cakes.

 

[NTMBK]

Hopefully both NVidia and AMD can squeeze a little more out of 28nm. Both have shown some promising low end products- NVidia's 750ti and AMD's Mullins both show pretty decent perf/W improvements.

 

[DigDog]

i just bought a 770, next time i will upgrade GPUs will have 1024bit buses and 16Gb VRAM. (i.e. circa 5 years)

edit: actually, i think this would make a great "totally speculatory 5-year GPU thread"

 

[dangerman1337]

One problem with this rumor; it assumes that they'll be a GM210 yet in the drivers there's a GM200 which has been rumored to use the 28nm process. Also I dobut Nvidia is going to simply use a 20nm shrink of the GM204 as stepping just means re-spins and not shrinks. The GT200 when it got shrinked to 55nm was called the GT200B yet according to this rumor a B stepping means 20nm which does not make any sense at all.

I'm going with others that there won't be any desktop GPUs with 20nmSoC, it is just pointless and woud be far too expensive for most for relatively minor gains. If the full fledged GM204 uses the same power as the GTX770 yet is 5.7 TF in SP which is approximately 1.8x as powerful on 28nm as the GTX770 then GM200 on 28nm could see similar gains (maybe 1.7x gains compared to the GK110?) but maybe lower as the die size is less flexibe as the GM107 had a die size increase from the GK107 around 25%.

Yup. We've only just got the big R295X2 and TitanZ is still to come. Then GM104 later this year on *28nm* its clear that 20nm will not be used by AMD/NV, since 16nm is just around the corner.

There isn't a GM104, only GM204. I'm very curious why Nvidia is calling the GM204, GM206 and the GM200 when all three (except maybe the GM200) will be using TSMC's 28nmHP process.

 

[Tristor]

You are misinterpreting what he said / means. As nvidia has proven with GM107, there is plenty of room to improve both perf/mm^2 and perf/watt on the existing node. 20nm, while it will help both of those metrics in and of itself, is not absolutely needed to come out with significantly faster, more efficient GPU's working with the existing technology.

I'm overjoyed at the improvements they were able to make in GM107, but I'd like to see those same improvements extrapolated out to a true big die GPU and encompassed in a die shrink. The additional perf/watt a die shrink will net us should make such a GPU a worthwhile upgrade over the current generation and might be usable as a starting point for 4k gaming. With games like Project Cars and Witcher 3 coming out, I want 4k ASAP, but it's just not really feasible right now within my parameters.

 

[tviceman]

I'm overjoyed at the improvements they were able to make in GM107, but I'd like to see those same improvements extrapolated out to a true big die GPU and encompassed in a die shrink. The additional perf/watt a die shrink will net us should make such a GPU a worthwhile upgrade over the current generation and might be usable as a starting point for 4k gaming. With games like Project Cars and Witcher 3 coming out, I want 4k ASAP, but it's just not really feasible right now within my parameters.

It still won't be entirely feasible with next gen high end single-card setups, either. Some games will be playable, but almost all will need many graphical features turned down to maintain consistently good frame rates.

 

[Stuka87]

Yup. We've only just got the big R295X2 and TitanZ is still to come. Then GM204 later this year on *28nm* its clear that 20nm will not be used by AMD/NV, since 16nm is just around the corner.

@tviceman
Given the 3200 shaders and 32 ROPs and clocks on Maxwell, it should be significantly faster than 780ti at the common resolutions of 1080p to 1600p. This is its goal since afterall, it is a mid-range chip. 4K is for the big Maxwell, like its for big GK110.

Edit: Typo, GM104/204. aka "Mid-range" Maxwell for outrageous prices.

20nm has been "Just around the corner" for two years, and is just now becoming available for SoC's. You seriously think that an HP 16nm process will be available in 2016?