NVIDIA Maxwell Speculation Thread

[Blazkowicz]

Yes, GM206. It is a form of futureproofing, I'd like to keep it for 3-4 years.

In the recent news, Samsung is annoucing volume production of 8Gb gddr5 memory, which would natively give a 4GB capacity on a 128bit bus.
I suppose GTX 960 won't use that on the short term but later it is very much a possibility, memory is very fast too at 8 Gbps or in other terms 2000 MHz gddr5.

 

[A1xLLcqAgt0qc2RyMz0y]

GM200 die size 632mm^2 ± 2.5%

http://wccftech.com/nvidia-maxwell-gm200-die-shot-analysis

 

[Kaarlisk]

In the recent news, Samsung is annoucing volume production of 8Gb gddr5 memory, which would natively give a 4GB capacity on a 128bit bus.

If I'm to be honest, I had not thought of this issue.
4GB should still be possible with 4Gb chips in clamshell mode?

 

[dbz]

GM200 die size 632mm^2 ± 2.5%
http://wccftech.com/nvidia-maxwell-gm200-die-shot-analysis

Is TSMC's max die size (stated as 650mm^2 in the article) official?
The closest I've found would seem to put the maximum at 625

 

[Ailuros]

Alatar's estimation is based on the relative difference to GK110, so the border (chip vs package) issue doesn't apply here.

If the die is truly over 600mm2 it bears the question again WTF they used up all those transistors for and didn't add enough FP64 units. Depending on small area variations here in the estimates and variations in transistor density it can weigh between 8.2 and 8.9b transistors.

Note that all the so far infos I had received for different Maxwell cores were spot on; for GM200 I had for quite some time now ~560mm2/7.9b. And I beg your pardon it's a rare case that anyone - even here - estimates die area spot on from die shots.

 

[Grall]

GM200 die size 632mm^2 ± 2.5%

Gods. If that's true, I think I'll have to pre-emptively call an ambulance for my extreme priaptic condition...

If the die is truly over 600mm2 it bears the question again WTF they used up all those transistors for and didn't add enough FP64 units.

A proportionally higher amount of L2 perhaps. Just speculation, of course...

And I beg your pardon it's a rare case that anyone - even here - estimates die area spot on from die shots.

True, but let's not be a bunch of party poopers, right? 625sqmm would be glorious, wouldn't it? Let's just hope it doesn't cost $1000+ yet again, but yeah, if this GPU turns out to be anywhere near this size, and hooked up to 12 gigs of high-speed GDDR5, then yeah. It probably will.

 

[LordEC911]

Is TSMC's max die size (stated as 650mm^2 in the article) official?
The closest I've found would seem to put the maximum at 625

It is a reticle limit of the tools.
It depends on the exact dimensions of the ASIC but from what has been said from knowledgeable sources in the past, the hard limit was around 650mm2.

And I beg your pardon it's a rare case that anyone - even here - estimates die area spot on from die shots.

Too lazy to do a die estimate but I just want to mention I got 441mm2 for a Hawaii die size estimate back at the end of Sept '13.
http://semiaccurate.com/forums/showpost.php?p=195470&postcount=92

Though on the other hand, I got 416mm2 for GM204. At least it was within my 5% MOE.
http://hardforum.com/showpost.php?p=1041092935&postcount=92

 

[Jawed]

True, but let's not be a bunch of party poopers, right? 625sqmm would be glorious, wouldn't it? Let's just hope it doesn't cost $1000+ yet again, but yeah, if this GPU turns out to be anywhere near this size, and hooked up to 12 gigs of high-speed GDDR5, then yeah. It probably will.

We need a poll: how long until the fastest reference single-GPU gamer-orientated card is $1500? With people spending silly amounts on GTX980 for such a marginal gain in performance, it can't be long, can it?

 

[lanek]

Gods. If that's true, I think I'll have to pre-emptively call an ambulance for my extreme priaptic condition...


A proportionally higher amount of L2 perhaps. Just speculation, of course...


True, but let's not be a bunch of party poopers, right? 625sqmm would be glorious, wouldn't it? Let's just hope it doesn't cost $1000+ yet again, but yeah, if this GPU turns out to be anywhere near this size, and hooked up to 12 gigs of high-speed GDDR5, then yeah. It probably will.

I really doubt that the gaming part will got 12GB of memory.... the Quadro will, they need at least that against the Firepro who had 12 and 16GB of memory. but the gaming part.. 384bit bus + 6GB..... maybe they plan a Titan with that amount and if this is the case, yes it will cost a lot ( for marginal gain over the 700$ part )

 

[keldor314]

Hrmm. Judging by the relative size... 3072 cuda cores for the full chip?

 

[Ailuros]

A proportionally higher amount of L2 perhaps. Just speculation, of course...

Shall I ask what for?

True, but let's not be a bunch of party poopers, right? 625sqmm would be glorious, wouldn't it? Let's just hope it doesn't cost $1000+ yet again, but yeah, if this GPU turns out to be anywhere near this size, and hooked up to 12 gigs of high-speed GDDR5, then yeah. It probably will.

It would be more glorious if they could get the job done at a healthy portion less die area. Anything over 600sqm. is not necessarily a eulogy.

 

[silent_guy]

If the die is truly over 600mm2 it bears the question again WTF they used up all those transistors for and didn't add enough FP64 units. Depending on small area variations here in the estimates and variations in transistor density it can weigh between 8.2 and 8.9b transistors.

Note that all the so far infos I had received for different Maxwell cores were spot on; for GM200 I had for quite some time now ~560mm2/7.9b. And I beg your pardon it's a rare case that anyone - even here - estimates die area spot on from die shots.

My answer is the same a before: if a gm204 is 400mm2 and a gm200 has exactly 50% more of everything (ROP, MC, SM, TEX, L2 cache), then 600mm2 is your baseline, 560mm2 close to impossible, and 624mm2 some overhead (IO placement, routing inefficiencies, extra redundancy somewhere, who knows.)

The guy who wrote that starts out with a ton of reasons why the data is low quality but then still ends up with an error of just 2.5%. Right...

 

[Grall]

We need a poll: how long until the fastest reference single-GPU gamer-orientated card is $1500?

You might say we've been there already, with NV's dual Titan board having launched at $3k...

I really doubt that the gaming part will got 12GB of memory....

I would kind of expect a flagship part to get that much RAM, TBH. 6GB, there's software out right now that can blow that much video memory, the bleeding edge needs headroom to expand into.

Shall I ask what for?

Improved buffering of pixels, for example. It'd improve rasterization efficiency (and cut power useage) quite a bit I'd think, considering what we've seen from similar designs recently (including Intel's Crystalwell chip, and so on.)

 

[iMacmatician]

From Sweclockers: "Nvidia unveils GM200 on the GPU Technology Conference 2015" (original).

I'm guessing that we will see the Quadro M6000 at the GTC.

EDIT: By the way, can anyone get a GM206 die size estimate from the picture here? I get about 208-232 mm^2 depending on how much of the edge of the package is covered by the black enclosure (?).

 

[spworley]

It now seems likely GM200's die is 600+ mm^2. That very large size gives room for extra pins, possibly enough to allow a 512 bit memory bus. NVidia hasn't used such a wide bus since GT200's similar sized 572 mm^2 die. Fermi and Kepler largest chips were both smaller (529 and 561 respectively) and both had only had room for 384 bit busses. A 512 bit bus would give big Maxwell a rather dramatic and welcomed memory bandwidth boost.

A 512 bit wide bus would give big Maxwell memory capacities of 4, 8, or even 16 GB. 384 wide gives 3, 6, and 12.

GT200 was using GDDR3, not GDDR5, but I doubt that change limited bus widths.

 

[spworley]

Thinking more, my 512 bit bus width speculation is probably wrong. The leaked board shot shows 12 RAM chips on the front, not 8 or 16, so GM200 is likely the usual 384 wide. Unless the leaked board was 384 for testing and final boards are 512, but that's unlikely.

 

[Grall]

It's not room for I/O balls on the die that is the bottleneck; AMD has had 512-bit buses on GPUs significantly smaller than NV's monster dies.

 

[3dilettante]

AMD's most recent high speed interfaces were physically bulky. Hawaii significantly backed off on bus speed, so much so that it was a sliver of an improvement over Tahiti.
One of the more notable reversals in recent years has been Nvidia's trumping the co-developer of GDDR5 at implementing high speed interfaces, although even then Nvidia's interfaces were not as compact as the lower-speed bus variants.
Nvidia could be going with the set round of GDDR5 shrinks and another bump of the max data rate. Samsung is marketing 8 Gbps chips, which it probably wouldn't mention if it didn't have a likely buyer.

AMD has already floundered at high speeds with Tahiti, and its pawning-off of its high-speed IO engineering team may not bode well for its future there.
Another possible user of decent quantities of high-speed GDDR5 is Intel's MIC, but that's also not sticking with GDDR5 later in the year.
Nvidia's not promising a move beyond GDDR5 for another year, at least.

 

[Newguy]

http://www.tomshardware.com/reviews/nvidia-geforce-gtx-960,4038.html

http://www.anandtech.com/show/8923/nvidia-launches-geforce-gtx-960 - not actually a review, just basics

http://techreport.com/review/27702/nvidia-geforce-gtx-960-graphics-card-reviewed

 

[lanek]

Guru3D have his review up, i put this link only, because he have report every different gpu tested on each one ( Asus strix, Ka2, MSI, SLI etc etc )
http://www.guru3d.com/articles-pages/msi-geforce-gtx-960-gaming-oc-review,1.html
___________________

Anyway, looking at the Techreport review, is it me or the number they report are only with the Asus Strix ?
its not really clear as they allways write GTX960, but in listing of gpu tested it is the Asus Styx (( who run at a nicely 1317mhz boost min ).

Really strange test on their part, including the choice of some games.

Edit:
Edit: yes it seems if we look at the "testing methodology - system used" :
Asus Strix GTX 960 GeForce 347.25 1253 1317 1800 2048