NVIDIA Maxwell Speculation Thread

[LordEC911]

I'm not sure I believe that they'll skip 20 nm completely. For consumer level products, I think it's believable and at this point likely.

But for the high end, low volume where cost isn't a factor, I bet there will be a 20 nm GPU that will show up as a Tesla product and then maybe filter to a consumer GPU as a GTX Titan II or something similar with an outrageous price tag.

However, I don't think I buy the time table for the 16nm refresh. I think the earliest would be Q4 2015 and at that point they'll probably call it the 9xx series.

TSMC's 20nm isn't meant for high performance parts... 20G is no more. Adding in the increase in cost with minimal power/speed benefits and waiting for FinFet makes even more sense.
Finfet is right around the corner, there is absolutely no reason to manufacture at 20nm SOC, only to move to 16nm FinFet in 6months then to move to 16nm FinFet+ in another 6-8months.

 

[A1xLLcqAgt0qc2RyMz0y]

only to move to 16nm FinFet in 6 months then to move to 16nm FinFet in another 6-8 months.

???

You have two moves to the same 16nm FinFet process with 6-8 months time taken for the second move.

 

[McHuj]

TSMC's 20nm isn't meant for high performance parts... 20G is no more.

I'm not exactly sure what defines a process for high performance parts other than achievable clock speeds by the tech's transistors.

So it doesn't look like 20SOC can support clocks needed by high end CPU's (3+ GHz) but it should be suitable for GPU's (even upcoming mobile SOC's should be able to hit 2+ GHz).

And the reason I speculated about a 20nm part for the highest end part was mainly due to the die size of GK110, ~550mm. I don't believe that they can put out a bigger chip so they would need the density scaling of 20nm, despite the higher costs. But since we're talking about $999 and up price point per GPU board, it's not that big of an issue. But that's also with my assumption that NVIDIA needs a part like that in the next 12 months and 16FF is not viable for another year or more.

I'll be more than happy to be wrong and have new high performance consumer GPU's on 16FF next spring.

 

[Esrever]

videocardz.com/51009/nvidia-preparing-four-maxwell-gm204-skus

I like this part, however: Bear in mind that Charlie’s website is called semi accurate for a reason. It could all be wrong from the beginning.

Videocardz is more like neveraccurate.

 

[dbz]

I'm not exactly sure what defines a process for high performance parts other than achievable clock speeds by the tech's transistors.

TSMC's own roadmap?
TSMC went to some lengths to differentiate the CLN20G (General Purpose) process from the SoC up until CLN20G's cancellation a couple of years ago.

 

[LordEC911]

???

You have two moves to the same 16nm FinFet process with 6-8 months time taken for the second move.

Yes sorry. The second 16nm FinFet was missing the "+".
TSMC is working on a second offering of 16nm FinFet, known right now as 16nm FinFet+, that is supposed to offer additional benefits and is supposed to be "on-par" with Intel's offerings. It is a refinement of their process, I'm assuming due to FinFet, that directly affects their transistor structure.
20SOC, 16FinFet and 16FinFet+ all share very similar design rules and minimal tweaks will be needed, supposedly.

 

[swaaye]

I've been reading about notebooks lately and found out that the 128-bit DDR3-based GeForce 850M (640SP Maxwell) typically outperforms quite a few older desktop 256-bit GDDR5 cards like 560 Ti and 6870. How on earth does it do that with ~32GB/s bandwidth?! Amazing.

 

[A1xLLcqAgt0qc2RyMz0y]

I've been reading about notebooks lately and found out that the 128-bit DDR3-based GeForce 850M (640SP Maxwell) typically outperforms quite a few older desktop 256-bit GDDR5 cards like 560 Ti and 6870. How on earth does it do that with ~32GB/s bandwidth?! Amazing.

I believe the on-die 2 MB L2 Cache helps a great deal.

 

[constant]

I also conccur that the 2 MB L2 is helping Maxwell a great deal despite lower bandwidth.

However, when you get into high resolution settings on games and serious GPU compute applications you'll still be bound by the bandwidth to RAM,

My question is then why people find the rumours reasonable that the GM204 (GTX880) would have a 256-bit memory interface with 7 Ghz frequency (estimated) => ~224 GB/s bandwith.

Today the GTX770 has 224 GB/s bandwith while the GTX780 has 288-324.

Could they be looking to launch with 2.5 D memories? Or do they simply plan to slap on more cache and hope that that it'll be good enough for most uses cases?

 

[Blazkowicz]

See the Pascal thread (nvidia GPU, not the old language) if you're thinking about 2.5D memory ; )

 

[xDxD]

Videocardz is more like neveraccurate.

It has no journalistic aspirations, personally I find it convenient, as the site wccftech.com., Just because does "the dirty work" going to find and bring back the rumors around the web

 

[lanek]

This time it is pretty confirmed: Maxwell 28nm = November

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

Hilbert H. wrote:

I can write with pretty good indication that in the November 2014 we will see Nvidia launch their new GeForce GTX 870 and GTX 880 graphics cards. Last week at an Nvidia event a little birdie already whispered in my ears 'we'll see you again soon Hilbert'. More info from an external source however indicates that 28nm Maxwell based GPUs (indeed not 20nm) will be released within roughly three months time.

 

[Deleted member 13524]

Well.. at least we know the top-end 8xx cards aren't yet another Kepler respin.

 

[tviceman]

Well I hope GM204 on 28nm was designed to run at a higher TDP than GK104, otherwise it'll be a relatively unexciting new high-end card from an outright performance point-of-view. If GM204 is as efficient over GK104 as GM107 is over GK107 but uses the same real-world power envelope as GK104, then it'll only be 15-20% faster than gtx780 TI. While impressive nonetheless, it would be lackluster for being the new high end.

 

[McHuj]

Price matters too. The 780ti is still around $699. If they can deliver +15% of that performance for under $499 (or a 870 with 780ti performance for $399) then it's not that bad. I think there's a big market out there that's not will to pay more than $500 for a GPU that hasn't been addressed as those people are sitting on their 680/7970's.

I was hoping for an earlier launch like September instead of November.

 

[LeStoffer]

Price matters too. The 780ti is still around $699. If they can deliver +15% of that performance for under $499 (or a 870 with 780ti performance for $399) then it's not that bad. I think there's a big market out there that's not will to pay more than $500 for a GPU that hasn't been addressed as those people are sitting on their 680/7970's.

Exactly. And the 880 Ti will probably not be regarded as the high end from nVidia for very long. The GM200 / Titan II will fill that void once it is out in, say, the start of 2015.

I don't think, however, that we will see a 780/780 Ti-equivalent of the GM200 to sit between the 880 Ti and Titan II unless AMD releases something that forces nVidia to do that.

 

[Kaotik]

I don't think, however, that we will see a 780/780 Ti-equivalent of the GM200 to sit between the 880 Ti and Titan II unless AMD releases something that forces nVidia to do that.

Well, there is the >500mm^2 chip (which rumors call Fiji), which should beat the crap out of anything currently on the market by notable margin, so GM200 would probably be needed against that

 

[Alexko]

Well, there is the >500mm^2 chip (which rumors call Fiji), which should beat the crap out of anything currently on the market by notable margin, so GM200 would probably be needed against that

GM204 could be enough.

 

[Pinstripe]

This time it is pretty confirmed: Maxwell 28nm = November

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

"Within 3 months"

So not November then, but earlier. Gotcha.

 

[pjbliverpool]

Well, there is the >500mm^2 chip (which rumors call Fiji), which should beat the crap out of anything currently on the market by notable margin, so GM200 would probably be needed against that

500mm^2 is only 14% more than the 290x so I wouldn't expect miracles. Even 550mm^2 would only be 25% more. If that equated directly to a 25% performance increase then it would be the fastest GPU out there but not by a huge margin.