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

[dangerman1337]

Well if Maxwell is only 28nm sounds like NV is doing a repeat of GK104/110 split or some kind of a "hybrid" split. Perhaps we will see 28nm GM204 gaming flagship 880, 28nm GM200 as Titan 2 in 3-4 months for the HPC market, then in late 2015/early 2016 GM210 consumer gaming flagship #2 on 16/20nm.

I will be very surprised if a 2560 core, 256-bit Maxwell card will provide a 50% increase in performance over 780Ti and essentially be near the top for 2 years of Maxwell generation.

I see 28nm GM200 as Titan 2 (it was taped around late may/early June according to Eyrines at Beyond3D and seen in a manifest shortly after) but I am not sure we'll see GM21x either, they may as well just fast track Pascal if they can get a 16nm high end GPU in late 2015/early 2016 @ say around 780ti price (unless people feel like waiting until late Q4 2017/early 2018 for 10nm Pascal ). I think Nvidia is going to do a 'shorter' (in the sense that the higher end versions of the cards) life time for Maxwell than Kepler so they can push back die shrink transitions (not 'cheap' anymore) in order to keep costs lower.

Though I think for the GM204, the most I expect is 40% increase taking it to 7TF and a bit. Though i wonder how it will perform VS the GK110 @ 2560x1440 and 4k considering Nvidia is going to likely use a 256-bit bus (though isn't the GM107 on mobile performing very well against chips with larger buses?).

 

[Cloudfire777]

http://www.guru3d.com/news-story/nvidia-geforce-gtx-870-and-gtx-880-available-in-november.html

We have Sweclockers who make vague articles about GTX 880 and GTX 870 coming sometime after the summer in autumn, most likely in Q4 with no sources but just speculation.

WCCFTech shortly follows and add their own assumptions and speculation, that GTX 880 and GTX 870 suddenly comes out in November, in Winter ffs. Just like that, a month was pin pointed out of thin air.

Then we have the article you link to, which says the GM204 GPUs will be launched within 3 months, and that some Nvidia employee told them "see you soon". Guru3D too apparantly have been coloured by WCCFTech`s own BS assumptions, and says in November.

We know that Maxwell is coming soon and within 3 months. Thats all we know. A reseller told a guy from overclock that GTX 880 will begin selling in September, which is within 3 months. November isnt even inside the 3 month margin. July - October is.
Nvidia added driver support for upcoming mobile Maxwell cards just recently. Do 2+2...

Seriously, what a bunch of BS all the rumor sites are writing.
But atleast its good to see that Videocardz is staying away from posting these vague and incorrect articles.

Its so sad to see this BS from WCCFtech spreading like wildfire through many sites.

 

[dangerman1337]

We have Sweclockers who make vague articles about GTX 880 and GTX 870 coming sometime after the summer in autumn, most likely in Q4 with no sources but just speculation.

WCCFTech shortly follows and add their own assumptions and speculation, that GTX 880 and GTX 870 suddenly comes out in November, in Winter ffs. Just like that, a month was pin pointed out of thin air.

Then we have the article you link to, which says the GM204 GPUs will be launched within 3 months, and that some Nvidia employee told them "see you soon". Guru3D too apparantly have been coloured by WCCFTech`s own BS assumptions, and says in November.

We know that Maxwell is coming soon and within 3 months. Thats all we know. A reseller told a guy from overclock that GTX 880 will begin selling in September, which is within 3 months. November isnt even inside the 3 month margin. July - October is.
Nvidia added driver support for upcoming mobile Maxwell cards just recently. Do 2+2...

Seriously, what a bunch of BS all the rumor sites are writing.
But atleast its good to see that Videocardz is staying away from posting these vague and incorrect articles.

Its so sad to see this BS from WCCFtech spreading like wildfire through many sites.

Yeah, the only guy I will fully trust on this is Eyrines at Beyond3D since he has been accurate on Maxwell (first one to say that the GM20x will be on 28nm, taping out dates and the like) and has indicated September/October date but he isn't very active.

 

[Pinstripe]

Not sure if this is related, but a Maxwell announcement seems to be likely at gamescom in Germany.
https://twitter.com/NVIDIA_UK/status/489783801972023296

 

[mavere]

I still think GM200 may debut on 20nm.

...don't hold your breath.

 

[tviceman]

...don't hold your breath.

I'm skipping next gen regardless so it's really of no consequence to me. I am merely speaking from a curiosity point of view, wondering how they can actually manufacture a noticeably better new ultra flagship on 28nm.

 

[Techhog]

...don't hold your breath.

ShintaiDK redeemed? 3+ more years of 28nm confirmed?

Warning issued for member callout.
-- stahlhart

 

[jpiniero]

I'm skipping next gen regardless so it's really of no consequence to me. I am merely speaking from a curiosity point of view, wondering how they can actually manufacture a noticeably better new ultra flagship on 28nm.

It won't be. It should however draw less power though.

 

[Loki726]

It won't be. It should however draw less power though.

I disagree. There is actually a lot of potential for big performance improvements just through better architecture, better circuits, and better software.

Think of it this way. Much of architecture, circuits, compiler, and system software research for the past 20-30 years has concentrated on sequential architectures (CPUs), and has delivered big performance improvements above and beyond advantages from technology scaling alone. Research into parallel architectures (GPUs) is much less mature, and there is still plenty of potential for big improvements at the same technology node.

 

[NostaSeronx]

GM204 and GM200 will be on 20-nm SOC.

GM204 = >GK110 @ GK104 original retail prices
GM200 = <Titan Z @ GK110 original retail prices

There will be two versions though one on 20-nm SOC and one on 16-nm FinFET. With Pascal being 16-nm FinFET plus coming 2016.

 

[boxleitnerb]

GM204 and GM200 will be on 20-nm SOC.

GM204 = >GK110 @ GK104 original retail prices
GM200 = <Titan Z @ GK110 original retail prices

There will be two versions though one on 20-nm SOC and one on 16-nm FinFET. With Pascal being 16-nm FinFET plus coming 2016.

And if you're wrong, will you stop posting personal speculations as facts? It would do this and all the other forums you are engaged in a favor.

 

[Cloudfire777]

GM204 and GM200 will be on 20-nm SOC.

GM204 = >GK110 @ GK104 original retail prices
GM200 = <Titan Z @ GK110 original retail prices

There will be two versions though one on 20-nm SOC and one on 16-nm FinFET. With Pascal being 16-nm FinFET plus coming 2016.

A 430mm2 GM204 in 20nm makes zero sense.
It will just be way too powerful due to all the Maxwell cores, even compared to GK110, and the size on a GM204 would ideally be around 300mm2 since they can shrink the transistors and keep the die smaller but still beating GK110 by a good margin and keeping the price down.

 

[NostaSeronx]

A 430mm2 GM204 in 20nm makes zero sense.

Oh, you mean that card with GK210 + Tegra K1(64-bit)? That was somehow leaked as GM204/GTX 880?

It is the successor to this machine; http://www.pcper.com/news/General-T...ent-ARM-GPU-Cluster-Homogeneous-GPU-Workloads
http://on-demand.gputechconf.com/gtc/2013/presentations/S3064-Pedraforca-ARM-GPU-Cluster-HPC.pdf
http://insidehpc.com/2013/06/12/ped...irst-to-combine-arm-cpus-gpus-and-infiniband/

System on PCIe is the future for HPC. Even, AMD is going that route with the open common slot architecture.

 

[boxleitnerb]

You are seriously confused.

 

[NTMBK]

Oh, you mean that card with GK210 + Tegra K1(64-bit)? That was somehow leaked as GM204/GTX 880?

Another classic bit of Seronx logic. "It's too big to be GM204 on 20nm! But since I have decided that GM204 is coming out on only 20nm and 16nm, that means it must be GM210."

 

[NostaSeronx]

Where did I say GM210... http://i.imgur.com/QMJ9fOn.png

I'm pretty sure I typed GK210 which is on the same process as GK20A.

The support was rather trivial on the Mesa/Gallium3D side since the GK20A is similar to the already-supported GK104 graphics processor while using the GK110 ISA.

GK20A and GK110 have the same ISA and so does the GK210.
=======
"New Kepler" on the xxx-node from TSMC.
"Maxwell" on the 20-nm node from TSMC.
"New Maxwell" on the 16-nm node from TSMC.
"Pascal" on the 16-nm node from TSMC.
"Volta" on the 16-nm or 10-nm node from TSMC. <== Not sure about this one as there is underground talks about Nvidia using 10-nm FinFETs from Intel.

 

[NTMBK]

Where did I say GM210... http://i.imgur.com/QMJ9fOn.png

I'm pretty sure I typed GK210 which is on the same process as GK20A.GK20A and GK110 have the same ISA and so does the GK210.

Oh, I'm terribly sorry, let me rephrase.

"It's too big to be GM204 on 20nm! But since I have decided that GM204 is coming out on only 20nm and 16nm, that means it must be GK210."

 

[NostaSeronx]

So, NTMBK, Nvidia can't have two architectures coming out at once?

GM200
GK280/GK180
GM204
GK210
GM117/GM107
GK208

^== I know it is confusing a bit but that is the 800 series. GM2xx = 20-nm, and GM1xx and GK2xx = 28-nm.

====
If Charlie says this then he must be right! Especially, when he has a super negative bias about Nvidia. *wink wink*

 

[Cloudfire777]

Oh, you mean that card with GK210 + Tegra K1(64-bit)? That was somehow leaked as GM204/GTX 880?

It is the successor to this machine; http://www.pcper.com/news/General-T...ent-ARM-GPU-Cluster-Homogeneous-GPU-Workloads
http://on-demand.gputechconf.com/gtc/2013/presentations/S3064-Pedraforca-ARM-GPU-Cluster-HPC.pdf
http://insidehpc.com/2013/06/12/ped...irst-to-combine-arm-cpus-gpus-and-infiniband/

System on PCIe is the future for HPC. Even, AMD is going that route with the open common slot architecture.

No, you clearly havent paid attention to the leak. The die was branded with a number that match the GM204 shipping documents from Zauba.

It is 100% a GM204 and its on 28nm unless Nvidia put some ARM cores on the same die which makes it bigger than it normally would be if everything was Maxwell cores

 

[NostaSeronx]

Wafer: T6A996.0A0
=/=
Wafer: T6A995.005 / T6A995.007

 

[SinOfLiberty]

Wafer: T6A996.0A0
=/=
Wafer: T6A995.005 / T6A995.007

Good point.

But, keep in mind that these lot numbers are for technical parts that gpu is made of. Especially with Maxwell, when NV designs them from bottom to top, there will be little to no difference in lot number.

 

[Cloudfire777]

Its probably just batch numbers etc that change from time to time, or design revisions.

The die was also marked with manufacturing date, week 21 in late May. Why on earth would Nvidia manufacture a GK210, stick with Kepler, reduce GPGPU performance GREATLY from GK110 (because of the die size) and add ARM processors on the same die?

But its not like that anyway, because the computer above with Tegra have the CPU on the side, not on the same die as the GPU....

 

[RussianSensation]

GM204 and GM200 will be on 20-nm SOC.

GM204 = >GK110 @ GK104 original retail prices
GM200 = <Titan Z @ GK110 original retail prices

There will be two versions though one on 20-nm SOC and one on 16-nm FinFET. With Pascal being 16-nm FinFET plus coming 2016.

There is no evidence which supports that next generation GPUs from AMD /Nv will be on anything other than 28nm. Regarding Pascal, NV already slipped with Fermi by 6 months and their roadmap initially had Kepler for 2011. Kepler in turn slipped to 2012. In 2012 when Kepler came out and NV projected Maxwell for 2014, none of us thought that would mean GM204 would launch first and at that only in Q4 2014. Even if you look at Kepler, the flagship never made it in 2012, with 780Ti launching in late 2013. That's a huge slip from 2012.

This actually could imply flagship Maxwell may not be out until mid- to late-2015. At this point, Nv might launch some low end (750Ti sku) Pascal in 2016 but I doubt we will see anything that resembles flagship Pascal before 2017.

Look at AMD as well, it took them 2 years to squeeze 35% more on 28nm. I don't see AMD lagging NV by 5-6 months with 780 vs 290 but all of a sudden have 20nm parts ready Q1 2015. Also, if NV made a 430mm2 GM204 on 20nm, this would imply a lot more than 2560 CUDA cores. But then the card could have too much shader power and not enough memory bandwidth and pixel full rate performance.

 

[Mand]

There is no evidence which supports that next generation GPUs from AMD /Nv will be on anything other than 28nm.

There is also no evidence which supports that next generation GPUs from AMD/Nv will be on 28nm.

 

[SinOfLiberty]

There is also no evidence which supports that next generation GPUs from AMD/Nv will be on 28nm.

Be prepared for links to rumor sites(treated as official sources) and some other baseless info.

:biggrin: