Kitguru : Nvidia to release three GeForce GTX 800 graphics cards this October

[jpiniero]

5.5B to 6B transistors

I had to look this up, and sure enough, GM107 is 1.87B transistors. *3 of that is 5.61B. Now whether that kind of simple scaling is really realistic I don't know. But it does look like they had to come up with something new when Maxwell 20 had to be canned, so doing something that doesn't stray much from a linear increase from GM107 would be the easiest way to get something out.

A 1920 core 880 would still be faster than the 780 Ti though.. at lower res at least.

 

[Fastx]

I concur. As a GK104 successor build on 28nm that is supposed to be affordable, don't expect a GTX 780 Ti slayer. That's just unrealisitc.

I was thinking and posted back on 7-6-14 here that Nvidia would release a Maxwell GTX 880 at around 15-20% faster than the GTX 780. So then some posters starting posting/predicting 20-35% faster than a 780Ti to my surprise (hard to believe imo) and at around $450-$549. So I am thinking if they are correct and this turns out to be right around their performance prediction, I will most likely sell my $356.00 Tri X 290 and go with a GTX 880. So now if KG turns out to be correct/close or +/- a few percent over a 780Ti, I will keeping be my Tri-X 290 oc. It will be interesting when released to see what the performance/price on the GTX 880 will be for sure.

 

[Pinstripe]

GTX 870 benchmarks leaked:
http://www.coolaler.com/showthread.php/317829

Via Google translate:

NVIDIA GTX 870, GTX 880 is the use of the second generation of Maxwell architecture, core GM204, mainly to replace GTX 770, GTX780 position, expected after the September debut.

Nice, exactly what I was expecting.

 

[Grooveriding]

GTX 870 benchmarks leaked:
http://www.coolaler.com/showthread.php/317829

Via Google translate:



Nice, exactly what I was expecting.

Hmmm. Performance is about 25% slower than a 780ti, not looking good if those are genuine.

 

[96Firebird]

Hmmm. Performance is about 25% slower than a 780ti, not looking good if those are genuine.

Have any links to stock 780 Ti 3DMark 11 results? They are hard to find, seems like most review sites skip these synthetics.

 

[blackened23]

Not sure what you're talking about. At non overclocked settings (looking at guru3d and other sites) that 3dmark11 score is higher than the GTX 780. So that would put the 870 as a slighter faster than GTX 780 part.

 

[blackened23]

Have any links to stock 780 Ti 3DMark 11 results? They are hard to find, seems like most review sites skip these synthetics.

As he did last time, he's probably basing this off of heavily overclocked 3dmark 11 suicide run benchmarks. Which, quite frankly, is silly. 3dmark11 extreme scores for the 780 at non overclocked stock settings for both the CPU and GPU are right around 4500. The 780ti is right around 5100.

Guru3d does 3dmark11 e and p. So does hardware.info among many others. If you look at the STOCK REFERENCE 780 and 780ti scores, non overclocked, stock, with a stock CPU (cpu speed affects the final score), 4500ish is the area for 780, 3500-3600 is the area for the 770.

So this essentially puts the 870 between the 780 and 780ti.

 

[ShintaiDK]

Hmmm. Performance is about 25% slower than a 780ti, not looking good if those are genuine.

It is a GTX870 tho. But again, there is only so much to do when stuck with 28nm. The high hope is still the power consumption. We still have to see that.

 

[blackened23]

It is a GTX870 tho. But again, there is only so much to do when stuck with 28nm. The high hope is still the power consumption. We still have to see that.

I don't think the 780ti is 25% faster than the 780. Again, the stock 780ti is around 5100-5200 in 3dmark11 extreme. The 780 is right around the 4400-4500 mark. Plus or minus a few points.

This is for STOCK CPUs, STOCK GPUs, nothing overclocked.

You're right of course about 28nm limitations, can only fit so many transistors on there. But these benches make the 870 right around 780 level of performance. Which, depending on the price, could be an okay fit.

 

[FatherMurphy]

If we are looking at an 870 that is slightly faster than a 780, based on Maxwell's power characteristics in the GM107, we should be looking at a part that consumes less avg. power than a GTX680, right? Perhaps that bodes well for overclocking potential.....

EDIT: Indeed, given that the GM204 is a larger chip (surface area) with potentially equal or lower power than a GTX680, an air cooler should be able to keep it very cool even under load.

 

[Grooveriding]

It is a GTX870 tho. But again, there is only so much to do when stuck with 28nm. The high hope is still the power consumption. We still have to see that.

I understand and agree with that. These cards are quite likely going to be marketed on their efficiency and price, but not penultimate performance in the face of beasts like the 780ti and Titan Black. Personally to me efficiency is a worthless marketing point in a gaming card, all that matters is performance, if efficiency helps you to get there - which it does - then so be it, but let's just see the performance numbers and how they compare to what else is available to measure the worth of a gaming card.

With every leak it's really more and more looking like what we have in the 870 and 880 will be gtx 780 and 770 replacements with better price points, but we're not getting a significant new flagship. 4 coming on 5 years of 28nm and the inability for nvidia to afford the newer TSMC process is starting to cripple performance leaps.

 

[Grooveriding]

If we are looking at an 870 that is slightly faster than a 780, based on Maxwell's power characteristics in the GM107, we should be looking at a part that consumes less avg. power than a GTX680, right? Perhaps that bodes well for overclocking potential.....

I would of thought that as well, but the reality is overclocking gains are not consistent on the same architecture and are generally more chip iteration specific. An example is GK104 in gtx 680 & 770 does not see the same sort amazing scaling out of an overclock that GK110 in 780 and 780ti does. Partly this is because most of the gains seen on GK104 are from overclocking the memory because it has the 256 bit bus giving it less available memory bandwidth than GK110 has. Whereas on GK110 you get big gains from core overclocks because there is more available memory bandwidth.

I have my doubts that there will be much in the way of gains out of overclocking the core on 870 and 880 because it will suffer the same weakness that 680 and 770 have with limited memory bandwidth. It will probably be the same situation where it is how well your memory overclocks that dictates what you can do overclocking wise.

 

[FatherMurphy]

That's true, Groove. And there can't be much more headroom in GDDR5 over 7Ghz, right?

EDIT: Now I'm thinking, where was it confirmed that GM204 has a 256-bit memory interface. Was it the memory layout on the engineering sample board that was pictured a while back? Nvidia had to create a much larger die to accommodate the Maxwell shaders, it would be relatively much easier to put a 320 or 384-bit memory interface on such a large die, right? I could just be showing my ignorance of the technical details of such an implementation, though.

DOUBLE EDIT: Also, even if the GM204 is 256-bit, Maxwell's architecture includes a significantly larger L2 cache that might alleviate memory bottlenecks that were observed in the 680/770. Speculation, of course, but it's something to think about.

 

[Grooveriding]

That's true, Groove. And there can't be much more headroom in GDDR5 over 7Ghz, right?

Not that I know of. I haven't read anything saying that there has been faster GDDR5 released as yet, so the assumption is these cards will use the same 7ghz GDDR5 found on GTX 770 and 780ti. So the 870 and 880 will have the same memory bandwidth as found on the gtx 770 given all that. Also from experience I can say the 7ghz ddr5 does not have much overclocking headroom. This is why on cards like the Kingpin 780ti they use custom Samsung memory instead of the stock stuff you find on 780ti as standard, because the overclocks on the stock 7ghz stuff are not that great.

As far as 3DMark11 scores, guru is worthless and you need to be more discerning to get accurate data. They just give you the flat score for the whole system. This is part of why 3Dmark is largely worthless because CPU and memory play a huge role in the results. You need to pay attention to graphics score and the fps results for graphics tests 1 through 4. There are plenty of resources out there to find results showing you all that information, feel free. Once you've managed that you'll see this possible leak is pretty weak in the performance department for a new card.

780ti is quite a bit faster than 780, it has 7ghz gddr5 vs the 6ghz on 780, and unlike in the past with X80 X70 part comparisons, where there was just a single enabled SMX difference, the 780ti has 15 SMX vs the 12 SMX on 780. It's a significant leap to go from 780 performance to 780ti.

Are these cards due next month ? We should start to see more and more leaks like this as it gets much closer to get a real idea of this is what the performance is going to be or not.

 

[ShintaiDK]

Personally to me efficiency is a worthless marketing point in a gaming card, all that matters is performance, if efficiency helps you to get there - which it does - then so be it, but let's just see the performance numbers and how they compare to what else is available to measure the worth of a gaming card.

For me its a good thing and high on the want list. I would love GTX780 performance at 150W for example.

 

[96Firebird]

So no 3DMark 11 links?

The GPU-Z screenshot shows 7GHz Samsung memory.

 

[Blue_Max]

For me its a good thing and high on the want list. I would love GTX780 performance at 150W for example.

Me too! It also means COOLER and SMALLER cards that'll fit in more machines like your typical mATX prebuilt HP, Acer or Dell.

Make it efficient enough, those same people can buy your card without having to worry about a new power supply as well! (I'd like to see them start including SATA-to-6pin power adapters though, most OEMs no longer use 4-pin Molex connectors at all.) From a marketing standpoint, the GTX 750TI is a solid-gold winner!

 

[RussianSensation]

A 1920 core 880 would still be faster than the 780 Ti though.. at lower res at least.

Assuming those CUDA cores give 1.35X increase in IPC and 1178mhz clock speeds. However, it would depend largely on the game. If a game is shader limited, it could beat 780Ti by 10-15% but most games are not entirely shader limited. 780Ti will have higher ROP and texture performance, which can also be just as limiting factors in games.

Groove seems to be spot on that 870/880 will have the same memory bandwidth as the 770. Leaks are starting to solidify 32 ROPs, 256-bit bus, 224 GB/sec memory bandwidth, 4GB of VRAM, for at least the lower end GM204 chips:
http://videocardz.com/51172/nvidia-geforce-gtx-870-specifications-leaked

With these specs this card could easily be made on 770's PCB. All the specs, including up to 1920 CUDA cores and rumored $400-450 price now squarely point to this card being a next gen mid-range Maxwell product, a direct successor to 680/780.

Sounds like 2560-3200 CUDA cores were just early wishful thinking rumors. Well there goes my prediction that this card will beat 780Ti by 30-35% on avg. The question now is what is NV going to use then to fill in the performance between a 780Ti and GM200/210? If 880 is a 680 successor then how expensive will 780/780Ti successors be? $650 for 780's and $800 for 780Ti's?

 

[ShintaiDK]

Me too! It also means COOLER and SMALLER cards that'll fit in more machines like your typical mATX prebuilt HP, Acer or Dell.

Make it efficient enough, those same people can buy your card without having to worry about a new power supply as well! (I'd like to see them start including SATA-to-6pin power adapters though, most OEMs no longer use 4-pin Molex connectors at all.) From a marketing standpoint, the GTX 750TI is a solid-gold winner!

We use 2 Silverstone SG08B cases here at home for gaming. One Haswell and one IB based. GTX680 and GTX670.

If these can give a good power reduction and say 30% speedboost. We will buy and give the old cards away.

 

[alcoholbob]

If the leaks are true about stock base clockspeeds at 1051mhz i wonder if these cards will have much overclocking headroom or if they are already running close to their limits. We might be at the situation of outside of custom bios and water cooling, 700 cards may actually be faster with overclocking headroom than potential performance of 800 cards on air.

 

[n0x1ous]

If the leaks are true about stock base clockspeeds at 1051mhz i wonder if these cards will have much overclocking headroom or if they are already running close to their limits. We might be at the situation of outside of custom bios and water cooling, 700 cards may actually be faster with overclocking headroom than potential performance of 800 cards on air.

well going off the 750ti....im thinking they will still have headroom.

plenty of 1.35 even 1.4 ghz OC's on little maxwell

 

[r3pshow]

VC: NVIDIA GeForce GTX 870 specifications leaked :

Here

what do you guys think about this? seems legit ?

 

[RussianSensation]

We use 2 Silverstone SG08B cases here at home for gaming. One Haswell and one IB based. GTX680 and GTX670.

If these can give a good power reduction and say 30% speedboost. We will buy and give the old cards away.

Considering it will be almost 3 years since 7970 launched and more than 2.5 years since 680 launched, beating them by 30% even at 150W TDP for a next gen product is very underwhelming to me.

In a similar time frame, I went from HD4890 (August 2009) to 470 to 6950 to 7970 (June 2012), which is 3X the performance increase and 3X the VRAM increase! Now in 3 years 30-40% is good? That would be horrible if we even compare NV's own leap from GF104 to GK104. A next gen card that barely beats 780 for $400-450 would be hardly worth getting excited about. I was hoping at least 15-25% faster than 780Ti for $499 but now it's looking underwhelming to say the least unless pricing is <$399. I mean 780 costs $450-500 but it's only as fast as a $380 290. So really for progress to be made, a slightly faster 780 1.5 years later should be $380-400 tops

Still hoping 880 beats 780Ti or otherwise even the naming makes little sense.

 

[ShintaiDK]

Considering it will be almost 3 years since 7970 launched and more than 2.5 years since 680 launched, beating them by 30% even at 150W TDP for a next gen product is very underwhelming to me.

In a similar time frame, I went from HD4890 (August 2009) to 470 to 6950 to 7970 (June 2012), which is 3X the performance increase and 3X the VRAM increase! Now in 3 years 30-40% is good? That would be horrible if we even compare NV's own leap from GF104 to GK104. A next gen card that barely beats 780 for $400-450 would be hardly worth getting excited about. I was hoping at least 15-25% faster than 780Ti for $499 but now it's looking underwhelming to say the least unless pricing is <$399. I mean 780 costs $450-500 but it's only as fast as a $380 290. So really for progress to be made, a slightly faster 780 1.5 years later should be $380-400 tops

Still hoping 880 beats 780Ti or otherwise even the naming makes little sense.

It is underwhelming. But its a reality we have to accept. Nomatter what else we experienced in the "good old days".

Transistor cost for anyone but Intel(And possible Apple and Qualcomm) is rising below 28nm. Specially 14/16nm is very expensive. Around 60% more than 28nm at the time.

While it doesnt exclude designs below 28nm. It sets a severe limit to their transistor count and/or price/performance. And it would still be cheaper gate wise to simply make an even bigger 28nm design.

EUV may help whenever it may come. But thats far away.

 

[Mand]

Still hoping 880 beats 780Ti or otherwise even the naming makes little sense.

I can't see how anyone could possibly think 880 won't beat 780Ti. Nvidia is not stupid or incompetent enough for that.