Maxwell Tidbits

wasabiman123

Member
May 28, 2013
132
1
81
Is this for 20nm or 28nm H1 2014 Maxwell parts O: I can see this definitely for 20nm, but that seems to good to be true for the current 28nm process.
 

gun5l1ng3r

Member
May 20, 2013
29
0
0
That sure is a high core count...makes me consider not going SLI and just waiting for one of these beasts...
 

BallaTheFeared

Diamond Member
Nov 15, 2010
8,115
0
71
I thought I read AMD was tapping out 20nm products in the first half of 2014.

I don't think Nvidia will bring Maxwell to 28nm, this has been one of their best nodes in awhile and 290X isn't changing the landscape it's just getting AMD into the fight for the holiday season.


Maxwell + GSync will be a potent combination for me (for AMD to get my money again) in my next upgrade cycle, either next year or the following for Intel 14nm.
 

Cookie Monster

Diamond Member
May 7, 2005
5,161
32
86
Is this for 20nm or 28nm H1 2014 Maxwell parts O: I can see this definitely for 20nm, but that seems to good to be true for the current 28nm process.

So from the general gist of things..

The initial Maxwell offerings will be using the 28nm process. Im not sure how they will fit all of that seeing as the GK110 is already at the maximum physical limits..

20nm will be used for the Maxwell refresh which would also have the denver ARM CPU incorporated into the die (or package).

Have no idea when these will be available though, but most likely the Maxwell refresh is a 2015 product while the initial Maxwell is perhaps aiming for a 1H 2014 release.

On a side note, SKYMTL over at XS (owner of hardwarecanucks) hinted that Maxwell 28nm samples were made in summer... although not sure what samples they were.
 

Majcric

Golden Member
May 3, 2011
1,369
37
91
The GPU roadmap sure looks different now, the last one I remember showed Maxwell as being a huge leap over Kepler. Oh well, grain of salt chart.
 

Keysplayr

Elite Member
Jan 16, 2003
21,209
50
91
Speculation: 6144 is double what GK110 was "supposed" to be. A 16SMX 3072 cuda core chip.
Maxwell, if it ends up with 6144, will be only 2x the shaders. And AMD will come out with double 290x with a huge amount of shaders as well, and lo and behold will yet again be trading blows with NV. Miraculous.
 
Last edited:

wasabiman123

Member
May 28, 2013
132
1
81
So from the general gist of things..

The initial Maxwell offerings will be using the 28nm process. Im not sure how they will fit all of that seeing as the GK110 is already at the maximum physical limits..

20nm will be used for the Maxwell refresh which would also have the denver ARM CPU incorporated into the die (or package).

Have no idea when these will be available though, but most likely the Maxwell refresh is a 2015 product while the initial Maxwell is perhaps aiming for a 1H 2014 release.

On a side note, SKYMTL over at XS (owner of hardwarecanucks) hinted that Maxwell 28nm samples were made in summer... although not sure what samples they were.

Thanks for the response, if the 20nm parts with Denver Armv8 on die is going to be an early 2015 ordeal...Ugh, did not want to wait that long to upgrade from the 670, don't know if going to 28nm Maxwell would be worth it....Especially without the onboard ARM cores.
 

Cookie Monster

Diamond Member
May 7, 2005
5,161
32
86
core count means nothing without knowing about the architecture.

Quite true as Kepler's CC were quite weaker than Fermi's CC. Maxwell may have (probably 99.9% true) somewhat further weakened cores but smaller in terms of size.
 

tviceman

Diamond Member
Mar 25, 2008
6,734
514
126
www.facebook.com
Is this for 20nm or 28nm H1 2014 Maxwell parts O: I can see this definitely for 20nm, but that seems to good to be true for the current 28nm process.

I don't see Maxwell coming on 28nm. Not when they're coming out with new 700 series products this and next month, and not when 20nm will be available sometime in the second half of next year. When was the last time Nvidia shrank an existing architecture? I believe it was GT200, and that was a looooong time ago. They might do it again when it makes sense logistically, but bringing a new product out in the 1st half of next year on the cusp of a new node coming doesn't make that logistical sense.

Anyways, I imagine these specs are entirely made up. Talking about core clocks now is a dead give away.
 
Last edited:

HurleyBird

Platinum Member
Apr 22, 2003
2,670
1,250
136
Quite true as Kepler's CC were quite weaker than Fermi's CC. Maxwell may have (probably 99.9% true) somewhat further weakened cores but smaller in terms of size.

Fermi to Kepler was a transition from scalar to superscalar, which was pretty radical. Undoubtedly there will be changes to all of the functional units but unless Nvidia goes back to scalar shaders or adopts VLIW you are unlikely to to see anything as dramatic as the change from Fermi to Kepler
 

tviceman

Diamond Member
Mar 25, 2008
6,734
514
126
www.facebook.com
Fermi to Kepler was a transition from scalar to superscalar, which was pretty radical. Undoubtedly there will be changes to all of the functional units but unless Nvidia goes back to scalar shaders or adopts VLIW you are unlikely to to see anything as dramatic as the change from Fermi to Kepler

GF104/GF114 was already superscalar, and fermi wasn't much (if any) improvement in perf/watt over GT200b.
 

HurleyBird

Platinum Member
Apr 22, 2003
2,670
1,250
136
GF104/GF114 was already superscalar, and fermi wasn't much (if any) improvement in perf/watt over GT200b.

I just knew someone would bring this up. GF104 was just barely superscalar, and certainly not aimed at extracting ILP to anywhere near the same extent as GCN or Kepler.
 

ShintaiDK

Lifer
Apr 22, 2012
20,378
145
106
20nm at best case is H2 2014. More likely Q1 2015.

Also these specs is the same fake specs posted some time ago.
 

boxleitnerb

Platinum Member
Nov 1, 2011
2,601
2
81
First Maxwell will come in Q1 2014, successor to GK104, 28nm, 1500-2000 SP.

If the card lives about 8-10 months, that's absolutely okay. After that, 20nm.
 

Genx87

Lifer
Apr 8, 2002
41,095
513
126
If Maxwell comes on 28nm it will have to be a slimmed down version. Or it will be massive, hot, and comsume a lot of power.

I think they wait until 20nm is ready near the end of 1H 2014. They have a decent enough product line to last until then.
 

ShintaiDK

Lifer
Apr 22, 2012
20,378
145
106
Maxwell on 28nm makes no sense. The benefits would be close to non exisistant.

GPUs dont really make progress without dieshrinks or larger dies.

The only thing you could really do on 28nm is take a GK104 and ~double it since its a more gaming oriented chip.
 

blackened23

Diamond Member
Jul 26, 2011
8,548
2
0
I think gains are still possible on 28nm, but the major problem is transistor density can't get much better - Maxwell won't have room to grow in terms of transistor count on 28nm. If it is initially released as a 28nm part, I wouldn't expect much more than a GTX 480 > 580 type of increase - 15% perhaps (just estimating). The real exciting prospect is 20nm with 1.5 to 2 times the transistors, that is where the real performance increase would come into play. Unfortunately, it looks like 2H 2014 is the soonest 20nm will happen with TSMC based on their history and the fact that 20nm is still in risk production. Risk production basically means, it isn't anywhere near ready for prime time.
 

boxleitnerb

Platinum Member
Nov 1, 2011
2,601
2
81
In the future with new nodes getting more and more expensive, the focus shifts to architecture.

A 28nm Maxwell GM104 (or similarly named) could take the performance position of the GTX 780, but with a much smaller die size and a little lower power consumption. It would compete better with AMDs new GPUs that are bound to come out some time in spring (Maui, Iceland). Or do you think they will stick to Tahiti and Pitcairn until 20nm?

Anyway, right now the signs clearly point to an early Maxwell release:
Editors of hardwarecanucks and videocardz.com, informed users on XS forums are all in agreement that samples were done in summer this year. This makes a spring launch very much possible. GK104 will be 2 years old by then and ripe for succession.

And last but not least, don't forget the 28HPM process. It's quite new and as far as I have read, it offered very significant gains to some Qualcomm SoCs vs 28HP. 28HPM is TMSC's best 28nm process with improved perf/W characteristics. Additionally, you can select a broader range of threshold voltages, in essence fine tuning all the transistors in your GPU (you have slow transistors which use little power and fast transistors which use lots of power). If you're power limited, 28HPMs gain over 28HP is almost as much as the gain from 20nm TSMC, or so I was told. This mainly applies to SoCs like in smart phones, but there was a rumor going around that the XB1 SoC uses the 28HPM process.
 
Last edited:

Keysplayr

Elite Member
Jan 16, 2003
21,209
50
91
I think gains are still possible on 28nm, but the major problem is transistor density can't get much better - Maxwell won't have room to grow in terms of transistor count on 28nm. If it is initially released as a 28nm part, I wouldn't expect much more than a GTX 480 > 580 type of increase - 15% perhaps (just estimating). The real exciting prospect is 20nm with 1.5 to 2 times the transistors, that is where the real performance increase would come into play. Unfortunately, it looks like 2H 2014 is the soonest 20nm will happen with TSMC based on their history and the fact that 20nm is still in risk production. Risk production basically means, it isn't anywhere near ready for prime time.

What if they're at the end of risk production?