[Videocardz] Maxwell 28 nm coming this year.

[Mondozei]

Earlier this year, I posted a thread about a scoop from Sweclockers. It basically said that we won't see no 20 nm Maxwell this year.

Well, according to shipment data that Videocardz - which has had a mixed history, but have gotten better over the past year - has looked at, we'll see a 28 nm Maxwell launch this year.
So still no 20 nm, which confirms the previous Sweclockers scoop.

More details:
http://videocardz.com/50301/nvidia-preparing-maxwell-gm204-kepler-gk210-gpus

Wccftech has also written a little about it:
http://wccftech.com/nvidia-maxwell-gm204-gpu-spotted-wild-kepler-gk210-works/

Remember, just rumors etc, so all caveats apply, but it would make sense. Not launching anything new aside from 750 Ti and Titan Z for the entire year(discounting minor tweaks to existing cards like Titan Black Edition etc) would look bad, especially as Tegra is one giant unholy kludge of fail for Nvidia right now.

 

[blackened23]

I just saw this on videocardz myself. If it's true, it'll be quite interesting to see how high NV will scale maxwell on 28nm.

 

[Stuka87]

This is not surprising. Pretty much what most of us expected.

But, given the way the 750 turned out, power consumption should be lower. Well, that or use the extra headroom for performance.

 

[x3sphere]

Interesting. Hoping it has HDMI 2.0 support.

 

[BrightCandle]

It is kind of surprising to have TSMC claiming mass production nearly 4 months ago and yet none of the GPU manufacturers using it. The costs of the new process must be enormous (low yield) for them to be skipping it until it matures. There isn't a whole lot they can do to improve performance dramatically on 28nm, they aren't going to get more transistors out of the process so performance is really stuck in a rut until that new process comes along.

 

[Cookie Monster]

It is kind of surprising to have TSMC claiming mass production nearly 4 months ago and yet none of the GPU manufacturers using it. The costs of the new process must be enormous (low yield) for them to be skipping it until it matures. There isn't a whole lot they can do to improve performance dramatically on 28nm, they aren't going to get more transistors out of the process so performance is really stuck in a rut until that new process comes along.

That is why Maxwell is interesting because of what the GTX750 Ti have showed us.

 

[Qwertilot]

Not such a shock if Apple and friends have eaten all the 20nm capacity for a good while

 

[paul878]

It is kind of surprising to have TSMC claiming mass production nearly 4 months ago and yet none of the GPU manufacturers using it. The costs of the new process must be enormous (low yield) for them to be skipping it until it matures. There isn't a whole lot they can do to improve performance dramatically on 28nm, they aren't going to get more transistors out of the process so performance is really stuck in a rut until that new process comes along.

That just might be the case. TSMC change per wafer, the buyer is paying for the working and none-working chips.

 

[BrightCandle]

In the past Nvidia has had quite a lot of issues getting its big chips to work at reasonable prices on new process technology, historically AMD has jumped to new nodes first. If AMD wont be touching it until next year that is highly indicative that Nvidia wont either. I knew the price was climbing but it seems kind of crazy to think its going to be 6-12 months with a new node being available before the GPU manufacturers use it, they are normally one of the first industries to jump on them.

 

[jpiniero]

Not such a shock if Apple and friends have eaten all the 20nm capacity for a good while

It does now look like the only high profile release using TSMC 20 this year is going to be the iProducts. Even Qualcomm's stuff isn't going to be released until 2015.

 

[Bateluer]

In the past Nvidia has had quite a lot of issues getting its big chips to work at reasonable prices on new process technology, historically AMD has jumped to new nodes first. If AMD wont be touching it until next year that is highly indicative that Nvidia wont either. I knew the price was climbing but it seems kind of crazy to think its going to be 6-12 months with a new node being available before the GPU manufacturers use it, they are normally one of the first industries to jump on them.

GloFlo just licensed Samsung's 14nm process. Not sure what AMD's contractual obligations are with GloFlo, but it would be interesting to see Radeons stamped on 14nm.

 

[tviceman]

GloFlo just licensed Samsung's 14nm process. Not sure what AMD's contractual obligations are with GloFlo, but it would be interesting to see Radeons stamped on 14nm.

Yes it would be interesting, 14nm is several years away. The license was to split the costs and R&D of developing the node.

 

[KaRLiToS]

I'm glad my R9 290x will last me longer. I have hard time keeping my stuff when something new releases.

 

[Wild Thing]

I'm surprised you don't have a 295X2 in your sig already Mr.Karl :wub:

 

[Bateluer]

Yes it would be interesting, 14nm is several years away. The license was to split the costs and R&D of developing the node.

Last I heard, Samsung was supposed to begin volume production of 14nm in 2015. Probably only a few months behind Intel's own 14nm, which was delayed to Q4 this year.

 

[witeken]

Last I heard, Samsung was supposed to begin volume production of 14nm in 2015. Probably only a few months behind Intel's own 14nm, which was delayed to Q4 this year.

Err, Intel already started volume production. If you want 14nm GPUs, you'll have to wait until 2016, not 2014. And those 14nm Samsung transistors aren't as good as Intel's, because they're 1st gen FinFETs vs. 2nd gen Tri-Gate, and the density is the same or just slightly better than 20nm, so there's no place for more shaders.

 

[Mondozei]

Err, Intel already started volume production. If you want 14nm GPUs, you'll have to wait until 2016, not 2014. And those 14nm Samsung transistors aren't as good as Intel's, because they're 1st gen FinFETs vs. 2nd gen Tri-Gate, and the density is the same or just slightly better than 20nm, so there's no place for more shaders.

Yeah, but Samsung is good as a fast adopter. They have been closing the node distance over the past 5 years.

Still, Intel is licensing out their fabs, too, would be hilarious to see Nvidia ride those out on 14 nm next year on Maxwell, two node jumps in one.
Hilarious, and unlikely, but still.

 

[Mondozei]

Given that a more efficient 28 nm node process gave us the GTX 800M series - basically the same kepler cards but with higher performance - plus a new architecture (Maxwell), I'm not so sure we will see paltry results with desktop GTX 800 series on 28 nm.

Especially given the fact with what we saw with 750 Ti. Big Maxwell will not come until 20 nm is ready, so 2015, but hopefully GTX 870 will perform at least like a 780 or maybe even a 780 Ti. That's be awesome for 300 dollars.


Edit:

This should also give impetus to AMD to relase their new GCN 2.0 series cards on 28 nm. Because what are they going to do otherwise, wait for 8-10 months as even middle-range cards go toe to toe (if we see serious improvements on Maxwell + more efficient 28 nm process) with R9-290?

 

[blackened23]

Nvidia took a 60W TDP part and doubled the performance per watt. If they scale that up to 150 or even 200W, it will easily be faster than the 780ti.

Although I think the halo part would not be possible on 28nm due to transistor count, 28nm doesn't have the density to do it. But 28nm should be fine for a meaningful increase over the 780ti, easily, because NV doubled the performance per watt. If they doubled it at 60W, that means a 120 or 200W part will also have double the PPW.

 

[BrightCandle]

Maxwell does well on power consumption compared to its predecessor but its also quite a bit larger in terms of die area, 1.87B compared to 1.3B of its predecessor. You can't just take a 780 ti and convert the cores to Maxwell and get a nice reduction in power consumption and a performance boost because it wont fit, its too big to be fabricated.

If we take BF4 as an example the 750 ti achieves 48 fps at high quality (1) and the 650 ti gets 37.9. So while it gains in power efficiency it looses out in transistor efficiency. That is actually one of the better cases, in a lot of games its lead is smaller. So while its 27% faster its also 44% larger. This poses a problem when we start talking about large die chips like 780's because the spare transistor budget for a 44% increase just isn't there, it would be capped at a similar size to the 780, ie 7.1B or so. Given that the performance of Maxwell would probably be lower than that of Keplar as its less efficient with its transistor in raw performance.

What it looks like to me is that Maxwell is designed for a new silicon process. They reduced power consumption with the expectation of near double the transistor count on a new GPU on a new process and the end result would be similar power consumption today and a top GPU around 14B transistors and a corresponding increase in performance somewhere around 50-100%. But its problematic using this architecture on 28nm, because the transistor budget means the design is limited, it will be very power efficient but it wont be faster unless something else changes.

 

[Saylick]

Maxwell does well on power consumption compared to its predecessor but its also quite a bit larger in terms of die area, 1.87B compared to 1.3B of its predecessor. You can't just take a 780 ti and convert the cores to Maxwell and get a nice reduction in power consumption and a performance boost because it wont fit, its too big to be fabricated.

If we take BF4 as an example the 750 ti achieves 48 fps at high quality (1) and the 650 ti gets 37.9. So while it gains in power efficiency it looses out in transistor efficiency. That is actually one of the better cases, in a lot of games its lead is smaller. So while its 27% faster its also 44% larger. This poses a problem when we start talking about large die chips like 780's because the spare transistor budget for a 44% increase just isn't there, it would be capped at a similar size to the 780, ie 7.1B or so. Given that the performance of Maxwell would probably be lower than that of Keplar as its less efficient with its transistor in raw performance.

As I stated before, Maxwell will do well because of IPC improvements AND area improvements. Nvidia managed to squeeze out 90% of an SMX (192 cores) using their SMM layout (128 cores) while also bringing a 35% improvement in instructions per core/clock through scheduler improvements.

What it looks like to me is that Maxwell is designed for a new silicon process. They reduced power consumption with the expectation of near double the transistor count on a new GPU on a new process and the end result would be similar power consumption today and a top GPU around 14B transistors and a corresponding increase in performance somewhere around 50-100%. But its problematic using this architecture on 28nm, because the transistor budget means the design is limited, it will be very power efficient but it wont be faster unless something else changes.

The 650 Ti is based on the GK106 die and not the GK107 die, which has a transistor count of 2.54B. Going by your numbers, this means that the 750Ti beats out the 650Ti in terms of power AND in transistor count. If you want to compare the 750Ti to the 650, according to Anand's bench, the 750Ti leads the 650 by about 2x in most if not all games tested.

 

[blackened23]

Maxwell does well on power consumption compared to its predecessor but its also quite a bit larger in terms of die area, 1.87B compared to 1.3B of its predecessor. You can't just take a 780 ti and convert the cores to Maxwell and get a nice reduction in power consumption and a performance boost because it wont fit, its too big to be fabricated.

If we take BF4 as an example the 750 ti achieves 48 fps at high quality (1) and the 650 ti gets 37.9. So while it gains in power efficiency it looses out in transistor efficiency. That is actually one of the better cases, in a lot of games its lead is smaller. So while its 27% faster its also 44% larger. This poses a problem when we start talking about large die chips like 780's because the spare transistor budget for a 44% increase just isn't there, it would be capped at a similar size to the 780, ie 7.1B or so. Given that the performance of Maxwell would probably be lower than that of Keplar as its less efficient with its transistor in raw performance.

What it looks like to me is that Maxwell is designed for a new silicon process. They reduced power consumption with the expectation of near double the transistor count on a new GPU on a new process and the end result would be similar power consumption today and a top GPU around 14B transistors and a corresponding increase in performance somewhere around 50-100%. But its problematic using this architecture on 28nm, because the transistor budget means the design is limited, it will be very power efficient but it wont be faster unless something else changes.

Correct. Die size is not inconsequential. At a certain point in scaling, 20nm will be required for full transistor density. This is why I stated that the true high end halo Maxwell part will require 20nm; there is really no avoiding this, once you're at that range it's just not going to be possible on 28nm. But that doens't mean that an increase over the 700 series isn't possible - they can still scale Maxwell on 28nm with the given PPW characteristics and likely get meaningful performance increases over the 700 generation. Given that the GK107 doubles performance per watt at 60W, it's very conceivable that Maxwell will do the same at 150W or 200W. How would a 175W TDP part perform with double the performance per watt? I think a Maxwell part in the 150-200W TDP range with similar performance per watt, scaled, will pass the 780ti. By napkin math, if the 750ti doubled the performance at a 60W TDP, then a Maxwell part with a 150-175W TDP could theoretically easily pass the 780ti.

The rumors indicate either a GTX 870 or 880. If so, that means it performs applicable to the intended SKU assigned to it, in other words, if any of these upcoming parts are going to be assigned an 870 or 880 monkier, they will perform as an 870 or 880. Meaning they will be meaningfully faster than the 780ti. Assuming the rumors are true, that is, and that 870 and 880 parts are indeed in the works. If they're releasing anything with an 880 SKU name i'd definitely expect it to perform accordingly.

I don't know though. That's what i'm theorizing, certainly these are rumors and absolutely nothing is set in stone. I do think that NV will release something this year because they've released a new generation yearly for some time now. Or pretty close to yearly. So I guess we'll see. It'll be interesting in any case.

 

[blastingcap]

This is an old article, but it makes you wonder. http://www.extremetech.com/computin...y-with-tsmc-claims-22nm-essentially-worthless

"20 or 14nm cost barely goes below the previous one, no saving!"

 

[Leadbox]

The 650 Ti is based on the GK106 die and not the GK107 die, which has a transistor count of 2.54B. Going by your numbers, this means that the 750Ti beats out the 650Ti in terms of power AND in transistor count. If you want to compare the 750Ti to the 650, according to Anand's bench, the 750Ti leads the 650 by about 2x in most if not all games tested.

Ain't it the 650Ti Boost thats based on GK106(nVidia's response to the 7790)? The 650Ti and plain old 650 are based on GK107.

 

[Saylick]

Ain't it the 650Ti Boost thats based on GK106(nVidia's response to the 7790)? The 650Ti and plain old 650 are based on GK107.

That's what I had initially thought until I looked up AT's article on the 650 Ti.

Coming from the GTX 660 and its fully enabled GK106 GPU, NVIDIA has cut several features and functional units in order to bring the GTX 650 Ti down to their desired TDP and price. As is customary for lower tier parts, GTX 650 Ti ships with a binned GK106 GPU with some functional units disabled, where it unfortunately takes a big hit. For the GTX 650 Ti NVIDIA has opted to disable both SMXes and ROP/L2/memory clusters, with a greater emphasis on the latter.