NVIDIA Maxwell Speculation Thread

[McHuj]

This original quote:

makes no mention that the "20nm production capacity" is only for ARM SOCs. Large die sizes such as GPUs and FPGAs will take up a lot of that 20nm production capacity and AMD is not mentioned as one of those who procured early 20nm production capacity from TSMC.

About a week or so ago, Xilinix announced that they're Ultrascale 20nm FPGA is available for shipping. That's a big chip, but a good sign that commercially viable parts are available (albeit probably very, very low volume). If I recall, I want to say that the first 28nm FPGA were out late summer of 2011 (about 4-6 months before Tahiti shipped).

So I don't think it is completely out of the realm of possibility that we'll see a 20nm GPU in Q2 of next year.

 

[silent_guy]

About a week or so ago, Xilinix announced that they're Ultrascale 20nm FPGA is available for shipping. That's a big chip, but a good sign that commercially viable parts are available (albeit probably very, very low volume).

Of course, $20000 per chip makes it quite a bit easier to have a commercially viable part.

 

[Esrever]

This original quote:

makes no mention that the "20nm production capacity" is only for ARM SOCs. Large die sizes such as GPUs and FPGAs will take up a lot of that 20nm production capacity and AMD is not mentioned as one of those who procured early 20nm production capacity from TSMC.

Its not a coincidence that nearly all the others mentioned makes ARM SoCs in the low power space?

Chang said mobile devices will remain mainstream products for the next two years, and TSMC will continue to play a critical role in foundry business.

I would imagine most of the talk about be focused around mobile anyways from this quote from your link.

 

[xDxD]

Here some rumors: http://videocardz.com/48610/nvidia-maxwell-details-revealed-ces-2014

 

[lanek]

A controller ? Maybe more something like on the AMD gpu's ( understand the scalar units )

 

[Ailuros]

Here some rumors: http://videocardz.com/48610/nvidia-maxwell-details-revealed-ces-2014

Well apparently Denver is (almost) ready after all, meaning I severely doubt there's anything behind that controller rumor. I don't think anyone expects the Maxwell top dog within this half of the year and since Denver cores will make it into K1-reloaded in H2 I don't see why the biggest Maxwell wouldn't want to have a few of them.

 

[iMacmatician]

From Fudzilla: "Nvidia Maxwell coming in Q1 is GM117."

The [GM117] chip is going after mobile designs as Maxwell is designed to be power efficient but it will play games just fine. With GM117 there might be more mobile design wins as well, as the chip has great thermals. On the desktop side the GM117 replaces the GK208, a Kepler based entry level part, and OEM 640 and 630 cards, but we expect that a 650 or faster replacement is on the way to follow in the months to come.

Some rumors not long ago claimed a midrange-level GM107 so I'm guessing at least one rumor has some important detail(s) wrong or there are in fact two Maxwell chips coming out in the next 3-4 months.

EDIT: They made a typo apparently, it's actually "GM107."

 

[Kaotik]

From Fudzilla: "Nvidia Maxwell coming in Q1 is GM117."

Some rumors not long ago claimed a midrange-level GM107 so I'm guessing at least one rumor has some important detail(s) wrong or there are in fact two Maxwell chips coming out in the next 3-4 months.

GM117 would be, based on NV's codenames in the past, somehow redesigned GM107, not completely different chip

 

[boxleitnerb]

Or codenames don't mean anything, at least for us who don't know their detailed internal meaning.

 

[iMacmatician]

Just so nobody (including me) gets mislead:

Fudzilla made a typo which they corrected, the "GM117" is actually "GM107." So I guess all the rumors are indeed talking about one "7" chip.

 

[Ailuros]

Just so nobody (including me) gets mislead:

Fudzilla made a typo which they corrected, the "GM117" is actually "GM107." So I guess all the rumors are indeed talking about one "7" chip.

A 1 or a 0 in the middle shouldn't make any major difference; however a "7" as a last digit based on track record doesn't suggest anything more than a low end chip.

 

[UniversalTruth]

A 1 or a 0 in the middle shouldn't make any major difference; however a "7" as a last digit based on track record doesn't suggest anything more than a low end chip.

No very big difference but if they go to GM117, most probably it would mean they had some type of problems with the original implementation forcing them or deciding to scrap it

 

[DSC]

http://videocardz.com/48726/nvidia-geforce-gtx-750-ti-maxwell-gpu-february

 

[Ailuros]

No very big difference but if they go to GM117, most probably it would mean they had some type of problems with the original implementation forcing them or deciding to scrap it

Not again with that "1" in the middle nonsense. Since we've unfortunately inhereted that fairy tale from the supposed "GK100" that never existed I think we can finally get back to reality, exactly because NV has these days the luxury to re-design entire product families and trash the old ones I guess.

http://videocardz.com/48726/nvidia-geforce-gtx-750-ti-maxwell-gpu-february

Sweoverclockers mention in their link though that there's one Denver core in the GPU; since online translators are usually a disaster I wonder if they mean 1 CPU core for all GPU chips irrelevant of size (doesn't make much sense) or if the specific chip (most likely GM107?) will carry 1 Denver core. If it truly is to replace the 650 Ti boost the latter has 768SPs@980MHz. Now 2,3 or 4 clusters?

 

[DSC]

I guess 2 GPCs or 2.5 GPCs like GK106 above if it's replacing GTX 650 Ti and GTX 650 Ti BOOST.

 

[silent_guy]

There's a whole industry of shady dealers in chip scrap.

 

[Picao84]

Sweoverclockers mention in their link though that there's one Denver core in the GPU; since online translators are usually a disaster I wonder if they mean 1 CPU core for all GPU chips irrelevant of size (doesn't make much sense) or if the specific chip (most likely GM107?) will carry 1 Denver core. If it truly is to replace the 650 Ti boost the latter has 768SPs@980MHz. Now 2,3 or 4 clusters?

I wonder another thing, though. G(N)1x7 has always been primarily for notebooks (yes, you can find on desktops, but I am not sure there really is a market for it outside OEMs) with low power/heat. I would be very skeptical about this part having similar power/heat characteristics to GK107 and performance around GTX650 Ti Boost.

What do you think? Would it possible? Maybe with the same 28nm process they would be using for TK1? Does it make sense?

 

[Picao84]

There's a whole industry of shady dealers in chip scrap.

I hope I am not the only one who did not get that

 

[silent_guy]

I hope I am not the only one who did not get that

I was thinking of tech journalists. And those who think that you scrap a chip because you can't make the previous one work, but then can make it work by giving it a different name.

Not saying that projects don't get cancelled, it happens all the time, but I can't remember a single case where it happens because someone can't make it work (whatever that means.) Projects get cancelled because market realities change: a competitor comes out with something drastically new, or because the design time is too long during the pre-tape out stage.

 

[Ailuros]

I wonder another thing, though. G(N)1x7 has always been primarily for notebooks (yes, you can find on desktops, but I am not sure there really is a market for it outside OEMs) with low power/heat. I would be very skeptical about this part having similar power/heat characteristics to GK107 and performance around GTX650 Ti Boost.

What do you think? Would it possible? Maybe with the same 28nm process they would be using for TK1? Does it make sense?

Of course is it possible and yes theoretically they could have used 28HPm for it.