NVIDIA Maxwell Speculation Thread

[McHuj]

I would argue that a modern SOC is much more complex than a GPU as it already contains a GPU plus a CPU that needs to hit clocks much higher than a GPU ever will. A mixture of a variety of components that run on various clock speeds is much harder to manufacture than a monotonous single function die.

When they say it's optimized for mobility, it's most likely that the transistors have been optimized for a good point of leakage and dynamic power at the expense of highest achievable frequency so no high end/high clock CPUs.

 

[entity279]

But is the "a monotonous single function die" an accurate description for a gpu? I would think not.

 

[Ailuros]

I knew that the 20nm process was not suitable for complex chips like gpus, just for the Soc like apple's ones

So why isn't any other SoC manufacturer - apart from Apple - (like QCOM for example) manufacturing any 20SoC SoCs in large volumes at TSMC within this year?

McHuj makes a couple of good points above; given that Apple's A7 is already at 101mm2@28nm and >1b transistors, a SoC like the A8 (which is going to pack a fairly higher amount of transistors) is anything but a "simple" chip.

NV and AMD haven't so far touched 20nm for whatever chip they might have developed because of the too high manufacturing costs, but it's not like Apple hasn't probably devoured the majority of early 20SoC capacities at TSMC with the usual volumes they're dealing with.

But is the "a monotonous single function die" an accurate description for a gpu? I would think not.

A GPU has it's own "monotonous" (one single tone) bandwidth it can use, while on a SoC you have a multitude of "polyphonic" units which are ripping each others garters into pieces for one and the same SoC bandwidth

 

[xDxD]

So why isn't any other SoC manufacturer - apart from Apple - (like QCOM for example) manufacturing any 20SoC SoCs in large volumes at TSMC within this year?

McHuj makes a couple of good points above; given that Apple's A7 is already at 101mm2@28nm and >1b transistors, a SoC like the A8 (which is going to pack a fairly higher amount of transistors) is anything but a "simple" chip.

NV and AMD haven't so far touched 20nm for whatever chip they might have developed because of the too high manufacturing costs, but it's not like Apple hasn't probably devoured the majority of early 20SoC capacities at TSMC with the usual volumes they're dealing with.

Cut:

Ok, the point is that there are fews (ot any) chances to see some nvidia gpus - mobile or discrete - made with 20nm this year and perhaps never (because they jump directly at 16nm process) ... Or not?

 

[Ailuros]

Ok, the point is that there are fews (ot any) chances to see some nvidia gpus - mobile or discrete - made with 20nm this year and perhaps never (because they jump directly at 16nm process) ... Or not?

I'm almost sure that NV wil use 20SoC at least for its Erista SoC in 2015. As for 16FinFET if there's any sort of (let's call it) "Apple effect" that affect TSMC's capacities or else availability for IHVs apart from Apple if the following has any merit we'll find out if there's anything to it sooner than later for 16FF also:

https://translate.google.gr/translate?hl=el&sl=zh-CN&tl=en&u=http%3A%2F%2Fudn.com%2FNEWS%2FFINANCE%2FFIN3%2F8892051.shtml

I know that there have been sitings claiming that NV intends to shrink upcoming GM20x chips to TSMCs 16FF. It could very well be true starting from Q2 15' the earliest IMHO; however it doesn't have to be the case either. It could be just another "misunderstanding" and it's just a testchip for 16FF to get GPxx0 out the door asap. I'm not saying it is or isn't, but those kind of "misunderstandings" happen all the time as much of a lot of (admittedly annoying ) "told you so" cases

 

[lanek]

Apple will not fill the TSMC capacity with 16nm Finfet ( or Finfet+ ).. Their lineup of product will be released next month and the next will not appear before 1 year .... So if they use 16nm for their next product, it will only be in 12 month minimum ..

But they will not use TSMC for their next products, its Samsung texas fundry who have allready the contract for the next one ... The problem, even today is not really Apple, fake informations given by financial media on USA... the real threat is Qualcomm who produce way more ARM based CPU in TSMC fundries that anyone.

But this is forget than Samsung Fundries, who alone produce way more chips than TSMC today.

 

[xDxD]

I'm almost sure that NV wil use 20SoC at least for its Erista SoC in 2015. As for 16FinFET if there's any sort of (let's call it) "Apple effect" that affect TSMC's capacities or else availability for IHVs apart from Apple if the following has any merit we'll find out if there's anything to it sooner than later for 16FF also:

https://translate.google.gr/translate?hl=el&sl=zh-CN&tl=en&u=http%3A%2F%2Fudn.com%2FNEWS%2FFINANCE%2FFIN3%2F8892051.shtml

I know that there have been sitings claiming that NV intends to shrink upcoming GM20x chips to TSMCs 16FF. It could very well be true starting from Q2 15' the earliest IMHO; however it doesn't have to be the case either. It could be just another "misunderstanding" and it's just a testchip for 16FF to get GPxx0 out the door asap. I'm not saying it is or isn't, but those kind of "misunderstandings" happen all the time as much of a lot of (admittedly annoying ) "told you so" cases

Ok i find it very interesting, thank you for the answer!

 

[Ailuros]

Apple will not fill the TSMC capacity with 16nm Finfet ( or Finfet+ ).. Their lineup of product will be released next month and the next will not appear before 1 year .... So if they use 16nm for their next product, it will only be in 12 month minimum ..

But they will not use TSMC for their next products, its Samsung texas fundry who have allready the contract for the next one ... The problem, even today is not really Apple, fake informations given by financial media on USA... the real threat is Qualcomm who produce way more ARM based CPU in TSMC fundries that anyone.

But this is forget than Samsung Fundries, who alone produce way more chips than TSMC today.

This is getting vastly OT but some of you act as if you actually know what Apple is planning. In that regard it's your gut feeling against following rumor: http://www.theverge.com/2014/8/26/6071193/larger-ipad-12-9-inches-apple-developing-for-early-2015

.....and now what?

 

[Arun]

There are a lot of *very* questionable assumptions being made here... So hey, I might as well join in the fun and make my own crazy bets!

GM20x will be 20nm.
GM107 and GM2xx are already tile-based (unlike Kepler).
GM204 will achieve GTX780-level performance with a 256-bit memory bus.

 

[elroy]

GM107 and GM2xx are already tile-based (unlike Kepler).

OMG they're using Gigapixel technology!!!!!!! Finally, 3DFX tech is being incorporated!!!!!

 

[Arun]

OMG they're using Gigapixel technology!!!!!!! Finally, 3DFX tech is being incorporated!!!!!

Hah!

GM107 doesn't show any of the performance characteristics of a traditional TBR with off-chip binning, so it's certainly not quite like Gigapixel, but it does show some very unusual characteristics that strongly imply they're doing on-chip binning for a relatively small number of triangles. I'm skeptical the 2MB L2 cache would make sense without that architecture (see http://forum.beyond3d.com/showthread.php?p=1856670#post1856670 and patents linked above my post).

BTW, the Qualcomm baseband on leaked iPhone 6 shots is 28nm so that's not where TSMC's capacity is going...

 

[Erinyes]

http://wccftech.com/nvidia-gm200-core-titan-ii-soon/

Hmm, is CP really Consumer Sample? Could it be Customer Sample and therefore be for the HPC market? 2 months sounds too early for a consumer GM200 doesn't it?

3-4 months should be more realistic. But the possibility is high that GM200 is pin-compatible to GK110/GK210, so the PCB designs are ready.

Maybe you can buy GM200 on some >$2k Titan at the end of '14.

GM200 is not planned for release before Q1'15 as per my info. Die size is ~560 mm2.

I'm almost sure that NV wil use 20SoC at least for its Erista SoC in 2015. As for 16FinFET if there's any sort of (let's call it) "Apple effect" that affect TSMC's capacities or else availability for IHVs apart from Apple if the following has any merit we'll find out if there's anything to it sooner than later for 16FF also:

https://translate.google.gr/transla...ttp://udn.com/NEWS/FINANCE/FIN3/8892051.shtml

Erista has already taped out on 20SoC a while back I have a few more details but I'll post in the relevant sub-forum.

There are a lot of *very* questionable assumptions being made here... So hey, I might as well join in the fun and make my own crazy bets!
GM20x will be 20nm.

I've heard there will be no GPU's on 20nm.

GM107 and GM2xx are already tile-based (unlike Kepler).

I've heard the same thing, Maxwell is already tile based in certain ways and Kepler also had a few minor things. Pascal is another (bigger) step in that direction.

GM204 will achieve GTX780-level performance with a 256-bit memory bus.

It will definitely come close IMHO..and I hope they price it well.

 

[xDxD]

Seems confirmed that gtx 980 will not beats gtx 780ti

videocardz.com/52122/overclockersuk-reference-geforce-gtx-980970-4gb-8gb-models-also-planned

It seems strange to me a new top vga slower than previous one (gtx 780ti), even at equal production process, not a good new for - entusiasths- gamers

 

[A1xLLcqAgt0qc2RyMz0y]

Seems confirmed that gtx 980 will not beats gtx 780ti

videocardz.com/52122/overclockersuk-reference-geforce-gtx-980970-4gb-8gb-models-also-planned

It seems strange to me a new top vga slower than previous one (gtx 780ti), even at equal production process, not a good new for - entusiasths- gamers

The 780ti is based on the GK110 and priced at $699.
Rumors have the GTX 980 priced at $499.

For a $200 discount to the 780ti it doesn't have to beat it.

The GTX 770 is based on the GK104 and new GTX 980/970 is GM204 based so that is where the comparisons should be.

EDIT: Quote from your link:

They replace 780, not 780Ti.

Performance wise, my lips are sealed on how it compares to 780Ti, but of course 980 is quicker than 780 for sure!

http://videocardz.com/52122/overclockersuk-reference-geforce-gtx-980970-4gb-8gb-models-also-planned

Note that the GTX 980 (GM204) will beat a GTX 780 (GK110).

 

[pjbliverpool]

Seems confirmed that gtx 980 will not beats gtx 780ti

That definitely doesn't confirm it, he just says he can't comment. If it comnfortably beats the 780 as stated then it at least has a shot at the 780Ti.

What I want to know is why have they skipped the 880 and 880Ti? It's sounds like the performance isn't going to justify the naming leap.

 

[xDxD]

The 780ti is based on the GK110 and priced at $699.
Rumors have the GTX 980 priced at $499.

For a $200 discount to the 780ti it doesn't have to beat it.

The GTX 770 is based on the GK104 and new GTX 980/970 is GM204 based so that is where the comparisons should be.

I know that, but those are vga for entusiasths, I think Nvidia better wait to have gm200 ready for market, in this situation you have gtx580<gtx680<gtx780<gtx780ti>gtx980...

That definitely doesn't confirm it, he just says he can't comment. If it comnfortably beats the 780 as stated then it at least has a shot at the 780Ti.

What I want to know is why have they skipped the 880 and 880Ti? It's sounds like the performance isn't going to justify the naming leap.

Let's wait and see , for the commercial name seems because there are Kepler mobile gpus named 8xx

 

[trinibwoy]

GM107 doesn't show any of the performance characteristics of a traditional TBR with off-chip binning, so it's certainly not quite like Gigapixel, but it does show some very unusual characteristics that strongly imply they're doing on-chip binning for a relatively small number of triangles. I'm skeptical the 2MB L2 cache would make sense without that architecture (see http://forum.beyond3d.com/showthread.php?p=1856670#post1856670 and patents linked above my post).

Isn't this a curious time to be investing in tiling with as a bandwidth saving measure with on-package/stacked DRAMs on the horizon?

 

[silent_guy]

I don't think we'll see stacked DRAMs on lower GPUs anytime soon. And on-chip RAM still has a lower latency, which is always nice. And lower power as well.

 

[RecessionCone]

Isn't this a curious time to be investing in tiling with as a bandwidth saving measure with on-package/stacked DRAMs on the horizon?

On-package memory is a nice one-time boost. But the laws of physics unfortunately make transferring data expensive relative to computing on it. Even with on-package memory, we need all the bandwidth savings we can get.

 

[Ailuros]

Isn't this a curious time to be investing in tiling with as a bandwidth saving measure with on-package/stacked DRAMs on the horizon?

Tiling doesn't suck only when NV uses it