[Pastebin/Forums] NVIDIA Maxwell GM1xx Specifications Leaked?

[parvadomus]

ARM cores? wtf? Biggest fake ever. /topic

 

[HurleyBird]

Sounds way too good to be true, but it would be awesome if it was.

 

[Keysplayr]

ARM cores? wtf? Biggest fake ever. /topic

Long ago rumors consisted of an on die or on card CPU on Nvidia's Maxwell.
I wouldn't discard this so quickly. That said, I am not quantifying any of the specs listed. I haven't a clue what Maxwell will be.

But here.

http://www.xbitlabs.com/news/graphi...well_Architecture_Will_Break_New_Grounds.html

 

[Cookie Monster]

That ARM chip is most likely be the creation from the denver project.

Sounds half-plausible.. but the idea that nVIDIA will incorporate an ARM chip into the GPU core i.e. resulting in a large SOC is probably 99% likely to happen. Its just when that is possible seeing as we won't see any 20nm products anytime soon. Thinking its a 2015 product, i.e. the initial Maxwell products will be the GPU core only and most likely based on the 28nm process.

I think maxwell's primary focus is efficiency? One thing for sure is that its going to be a new uarch so things will finally get interesting again along with AMD's new GPU.

 

[Keysplayr]

I guess they can "make" anything they want. As large a die as any process allows with integrated ARM CPU and Integrated Audio, LAN, Wifi. The question is, how well can software engineers access and harness the capabilities of this hardware?
It would be great to have an on die CPU, but what will it actually do? Triangle setup, thus removing that task from the actual system CPU? AI enhancements? Possibilities are endless. Just a matter of how well software takes advantage of it. I guess this is true with any bit of hardware out there. But this is a new idea. Could take years to perfect if at all.

 

[Cloudfire777]

ARM cores? wtf? Biggest fake ever. /topic

That Nvidia will use ARM along with the CUDA cores with Maxwell was already revealed last year.

http://uk.hardware.info/news/31510/nvidia-maxwell-gpu-integrates-64-bits-arm-processor-in-2014

 

[Erenhardt]

That Nvidia will use ARM along with the CUDA cores with Maxwell was already revealed last year.

http://uk.hardware.info/news/31510/nvidia-maxwell-gpu-integrates-64-bits-arm-processor-in-2014

Site doesn't work.
How is ARMs PCI interface? Cortex A9 had 1 line 32Bit (66MHz) which is not much. Custom design maybe? Or they use Jaguar with 4 PCI links? How is it going to access memory? Will it use GDDR from the card?

 

[ShintaiDK]

The ARM in Maxwell is most likely the mobile product called Parker.

Volta is where it gets interesting with stacked DRAM.

 

[Keysplayr]

Site doesn't work.
How is ARMs PCI interface? Cortex A9 had 1 line 32Bit (66MHz) which is not much. Custom design maybe? Or they use Jaguar with 4 PCI links? How is it going to access memory? Will it use GDDR from the card?

It would totally depend on what it will be designed to do. 66MHz might be waaaaay overkill. We don't know so these numbers being thrown around mean pretty much nil.

 

[Cloudfire777]

Site doesn't work.
How is ARMs PCI interface? Cortex A9 had 1 line 32Bit (66MHz) which is not much. Custom design maybe? Or they use Jaguar with 4 PCI links? How is it going to access memory? Will it use GDDR from the card?

The link works great here.

 

[blackened23]

With nonsense like this I wonder if it's a random nvidia fanboy who can't handle an AMD card launching and decides to make up a pile of nonsense and put it on a site like pastebin, or if it's nvidia's viral machine trying to get some attention in the midst of a big AMD launch.

Probably a more relevant discussion point than the random nonsense in the OP from pastebin. :awe:

The timing is hilarious indeed.

 

[tviceman]

FAAAAAAAAAAAAAAAAAAAAAAAAAAAAKE


But interestingly with these fake specs and rumors, I've talked about on beyond3d that Nvidia may not move to GDDR6 vram with first gen Maxwell parts. They've been slow to migrate to new rams in the past. IF GM104 leads 6+ months ahead of GM100/GM110, cards based on that chip may in fact use a 320-bit memory interface with 7.0ghz vram (giving it gtx780 bandwidth), and that GM100/GM110 would be Nvidia's first chip to feature GDDR6 vram if it's in full production by then.

 

[RussianSensation]

Doubling of the cores moving to a new production process is normal.

It is?

AMD - HD6970 to HD7970: 1536 SP --> 2048 (33% increase)
NV - GTX580 to GTX780: 512 CUDA --> 2688 (5.25x increase)

Without knowing how Maxwell architecture differs from Fermi/Kepler, trying to estimate the # of CUDA cores for GM100/110 is a pure guessing game.

Also, the move from 28nm to 20nm won't be as large as the move from 40nm to 28nm. This makes it a lot more difficult to predict specs. Then there is the uncertainty of rumours pointing to 28nm vs. 20nm Maxwell cards. If Maxwell is 20nm, it's not launching for a while I bet. If NV releases 28nm Maxwell cards, we may not see the true successor to 780 until 2015 on 20nm.

 

[Ibra]

Is this NV's attempt to mitigate mantle? Looks like they're concerned, although pastebin as a source implies it's just a fan trying to spread fud.

Might as well read tea leaves. Anyways the rumor mill appears to be gearing up to try discredit the amd launch.

Fantasy GPUs on a fantasy node using fantasy finfet technology. Sounds legit.

Perhaps start the wild rumours closer to 20nm release next time, m'kay?

With nonsense like this I wonder if it's a random nvidia fanboy who can't handle an AMD card launching and decides to make up a pile of nonsense and put it on a site like pastebin, or if it's nvidia's viral machine trying to get some attention in the midst of a big AMD launch.

Probably a more relevant discussion point than the random nonsense in the OP from pastebin. :awe:

Saying dudes which believed AMD's XRD2, 20 nm, 4096 stream pr. till last second. :sneaky:

 

[wand3r3r]

Saying dudes which believed AMD's XRD2, 20 nm, 4096 stream pr. till last second. :sneaky:

Sorry but I can't even understand what you're saying.

 

[96Firebird]

We'll be waiting a while for these...

 

[Erenhardt]

We'll be waiting a while for these...

Yep! That is what we (nv marketing department) wanted to hear! Bear with us!

 

[seitur]

Next GPUs from Nvidia rleased in 2014 will be 28nm. Get real.

 

[blackened23]

Next GPUs from Nvidia rleased in 2014 will be 28nm. Get real.

That's the entire point as to why this rumor is ridiculous. 20nm won't be possible until 2H 2014. This article, hilariously, mentions TSMC's FinFETs. Anyone with knowledge of TSMC's track record and schedule would realize their 20nm with FinFET (oh wait, TSMC's marketing calls it "16nm" :awe won't be ready until 2H 2015 or 2016.

TSMC is pretty hilarious. Their marketing has no shame to call their 20nm with FinFETs process "16nm". Intel, for their part, are honest about their nodes. The same cannot be said about TSMC - TSMC also claimed to be shipping 28nm since 1H 2010. Yeah okay. TSMC marketing at its' finest.

This tangent aside, any mention of Maxwell and FinFETs is absolute stupidity. Not unless Maxwell is a 2015-2016 product. There *is* a very real possibility of Maxwell being 28nm - 20nm will not be ready until 2H 2014 for volume shipments and the wafer costs will be astronomical to the point that the break even points in terms of profit will be 750k units (last I heard). That means that, initially, only giant corporations like Apple or Qualcomm will be able to buy into 20nm from the offset. Maybe nvidia can? I don't know. Certainly they're in a better position than AMD, but, it won't be easy for nvidia either. And that is still 2H 2014 for 20nm regardless of whether they're willing to pay for 20nm wafer costs.

 

[jpiniero]

I was thinking that Maxwell was backported to 28 nm when they realized that 20 wasn't going to be available until 2015 due to Apple.

 

[blackened23]

I was thinking that Maxwell was backported to 28 nm when they realized that 20 wasn't going to be available until 2015 due to Apple.

Maxwell could certainly be 20nm but if so, it will be 2H 2014 and it will take nvidia a ton of sales to meet a break even point. Nvidia could, presumably, ship a 28nm Maxwell sooner.

I dunno. Gains are still possible on 28nm but there are definitely diminishing returns; you just can't fit many more transistors than GK110 already has. That said, Apple has done tremendous work on 28nm (see their A7) and good things are possible. Gains are still possible on 28nm, make no mistake.

20nm is a possibility with Maxwell, but no one knows when other than 2H 2014 if nvidia wants to take the hit on wafer prices, or later if not. 28nm is also a VERY real possibility. FinFETs? That is pretty lol-worthy. Maxwell and FinFETs will not happen unless nvidia wants to ship it in 2016 - and that is what this rumor alludes to (Maxwell using FinFETs). Give me a break.

 

[tviceman]

It is?

AMD - HD6970 to HD7970: 1536 SP --> 2048 (33% increase)
NV - GTX580 to GTX780: 512 CUDA --> 2688 (5.25x increase)

As we know, though, Kepler cores are approximately equal to 1/2 of a Fermi core. So Keys got it mostly right, GK104 doubled GF114, GK106 doubled GF116, and GK107 doubled GF118.

 

[tviceman]

This tangent aside, any mention of Maxwell and FinFETs is absolute stupidity. Not unless Maxwell is a 2015-2016 product. There *is* a very real possibility of Maxwell being 28nm - 20nm will not be ready until 2H 2014 for volume shipments and the wafer costs will be astronomical to the point that the break even points in terms of profit will be 750k units (last I heard). That means that, initially, only giant corporations like Apple or Qualcomm will be able to buy into 20nm from the offset. Maybe nvidia can? I don't know. Certainly they're in a better position than AMD, but, it won't be easy for nvidia either. And that is still 2H 2014 for 20nm regardless of whether they're willing to pay for 20nm wafer costs.

Well crap, with the margins for 28nm have been record breaking for Nvidia, I don't see why the trend will stop at 20nm. If Nvidia beats AMD to 20nm GPU's by 4-6 months (which I think is a real possibility at this point, unless AMD is going to shrink Hawaii), they will be able to jack prices through the roof, especially if Maxwell brings the same leap in performance over Kepler as Kepler did over Fermi. If Hawaii cards are still going for $499 when Maxwell launches, Nvidia can just come over the top with a ~320mm^2 GM104 die (just throwing out an arbitrary die size a little bit larger but in the same ballpark as GK104) for $599-649 and live off that until AMD gets there 20nm cards out in October (or so).

 

[BoFox]

Looks like Maxwell will be by far the most bandwidth-bottlenecked of all architectures to date, before Volta with its stacked DRAM (or Einstein).

That's probably why they have to do all kinds of crazy stuff with L1/L2/L3 cache optimizations.

With the rumored specs in the OP, GM104 would be a massive let-down if it came out first, with GM100 coming out at least 6 months later on. The specs look too much like GF104 vs GT200, proportionally - imagine if GTX 460 was the only thing for 6 months, barely giving you anything over what GTX 285 gave you (ignoring DX11, of course).

 

[iMacmatician]

This was posted earlier at B3D forum.

And 3DCenter forum before that.

When it mentions TSMC's finFets, you know it's fake. Unless Maxwell is a 2015-2016 product.

That's the entire point as to why this rumor is ridiculous. 20nm won't be possible until 2H 2014. This article, hilariously, mentions TSMC's FinFETs. Anyone with knowledge of TSMC's track record and schedule would realize their 20nm with FinFET (oh wait, TSMC's marketing calls it "16nm" :awe won't be ready until 2H 2015 or 2016.

According to the "leak," the FinFETs appear to be for the refresh Maxwells, and if the first line of 20 nm Maxwells comes in late 2014 or so then it would not be surprising for the refresh to come in late 2015 or even 2016.