NVIDIA Kepler speculation thread

[UniversalTruth]

AlexV - are you allowed to say anything about when people that actually know something will be allowed to say anything? Hitting refresh on this thread is tiring out my index finger.

Monday, March 12th? As per Charlie.

http://semiaccurate.com/2012/03/08/...o-nvidia-keplergk104gtx680-launch-activities/

 

[tviceman]

It's 5% behind 7970.

What interests me most though is the die size. Latest info suggests it's just a bit under 300mm^2.

I thought this post http://forum.beyond3d.com/showpost.php?p=1625114&postcount=2086 pegged it being very close to 320mm^2.

 

[Man from Atlantis]

That's the top metal layer, I think. Not much to see really - just the overall layout.

p.s.: By the I/O pad population, this chip seems to be a 256-bit device.

if my myopic eyes dont deceive me i see 12SMs or should i say 16?

 

[fellix]

More like 32 SMs. [strike]I'm too lazy now to draw an overlay map to illustrate it[/strike]. That would accommodate perfectly the rumored 1536 ALUs -- 48 per SM, like the GF114 configuration, and 128 TMU (4 per SM).

Green - SM partitions;
Red - setup pipes, command processor, ROPs & etc.;

 

[Picao84]

More like 32 SMs. [strike]I'm too lazy now to draw an overlay map to illustrate it[/strike]. That would accommodate perfectly the rumored 1536 ALUs -- 48 per SM, like the GF114 configuration, and 128 TMU (4 per SM).

Green - SM partitions;
Red - setup pipes, command processor, ROPs & etc.;

Thanks.
So, going by Charlie's indication of a 7 and 8 group chip, the salvage part should have 28 SMs or 1344 ALUs. Although there are some "sources" saying GTX670Ti has same number of ALUs than GTX680.

 

[iMacmatician]

I thought this post http://forum.beyond3d.com/showpost.php?p=1625114&postcount=2086 pegged it being very close to 320mm^2.

Well, it says Tahiti is 365 mm^2, but it's actually 352 mm^2. Using that number, I calculated (by measuring the large green and red blocks on the right-hand side) that GK104 is ~308 mm^2.

 

[DarthShader]

Would the "square" in the middle be the cache?

 

[DavidGraham]

So let me get this straight , as I didnt have the time to read through all the leaks . 680>7970 , 670>7950 ?

 

[tviceman]

Well, it says Tahiti is 365 mm^2, but it's actually 352 mm^2. Using that number, I calculated (by measuring the large green and red blocks on the right-hand side) that GK104 is ~308 mm^2.

It keeps shrinking! Well either way, assuming the performance rumors coming out of the chinese forums are true, Nvidia will have, for the first time in a LONG, LONG time, over taken AMD in performance per watt and performance per mm^2.

Here's to hoping they price it aggressively (or at least somewhat aggressively). GK104 noticeably smaller than GF114, but coupled with 1 gig of extra vram and assuming a ~20% cost increase with the 28nm process, R&D aside, Nvidia could sell a fully unlocked GK104 for $299 and make the same profit as they did with GF114's gtx560ti. Of course, I don't think Nvidia is going to go in that low, but even at $399, they would have a significantly higher gross profit margin while simultaneously undercutting AMD..... fingers crossed for decent pricing!

 

[fellix]

Would the "square" in the middle be the cache?

Nope, the L2 is distributed among the memory controllers. The middle structure is probably thread command and dispatch logic, similar to GF100.

 

[jimbo75]

Here's to hoping they price it aggressively (or at least somewhat aggressively). GK104 noticeably smaller than GF114, but coupled with 1 gig of extra vram and assuming a ~20% cost increase with the 28nm process, R&D aside, Nvidia could sell a fully unlocked GK104 for $299 and make the same profit as they did with GF114's gtx560ti.

Considering all the dual fan cards, hefty factory overclocks and quality VRM's I havedoubts that Nvidia was making any kind of profit on the 560 Ti. If they were their partners sure weren't.

Of course, I don't think Nvidia is going to go in that low, but even at $399, they would have a significantly higher gross profit margin while simultaneously undercutting AMD..... fingers crossed for decent pricing!

I'm pretty sure AMD could make a small profit on the 7870 at $150. That brings crossfire into the equation, which would be $299. Vs a $399 680 that's not a difficult choice.

AMD has left wide gaps for Nvidia to price the card no matter how it performs. If Nvidia decides to start another price war then they are plain stupid and if I was an Nvidia shareholder I'd be wanting to know why. They won't - if they had any intention of doing that the card would be called the 660 Ti.

This 680 will be $400 and will start a price war that neither of them can win, or $550 and probably won't. If the leaked performance is true It'll be closer to $550 or I'll eat my hat.

 

[DarthShader]

Nvidia will have, for the first time in a LONG, LONG time, over taken AMD in performance per watt and performance per mm^2.

Not pn performance per mm^2, unless GK104 is 50-60% faster than the 212mm^2 Pitcairn chip.

Nope, the L2 is distributed among the memory controllers. The middle structure is probably thread command and dispatch logic, similar to GF100.

Right, my bad, although I don't quite see where those caches would "stuck in". Here's a Fermi shot for comaprision:

The shaders certainly take up more relative area on chip now.

 

[lanek]

Interesting… SemiAccurate claimed "8 group" vs. "7 group" GK104s and now here's a claim of 670 Ti CCs = 680 CCs.

Besides the possibility that one source or the other is incorrect, maybe that gives the possibility of 3 models: a 670 with fewer CCs and a lower clock, a 670 Ti with all the CCs and a lower clock, and a 680 with all the CCs (I'm assuming it's all the CCs…) and a higher clock? Two variants of an NVIDIA chip of ~300 mm^2 or larger with full CC counts are rare, the 8800 GTX/Ultra and the GTX 275/285 are the only ones I can find since the 8000 series and in either case, one chip launched some time before the other. If the link in the info is true, then I'm guessing that the 670 Ti is more of a "standard" clocked GK104 and the 680 is to try to beat Tahiti at least until GK110 comes along.

If the discrepancies in supposed clock speeds (705/950) weren't explained by the dynamic clocking, the above could explain that too (although the difference is quite large…).

Well a solution is just some have got not retail version where the CC's was not disabled. ( I remember the case on previous launch, but i think it was with AMD this round ).. This could well explain the difference.

 

[tviceman]

Considering all the dual fan cards, hefty factory overclocks and quality VRM's I havedoubts that Nvidia was making any kind of profit on the 560 Ti. If they were their partners sure weren't.

Nvidia didn't enjoy 50% gross margins on the entire year of 2011 on their single best selling chip (GF114) breaking even. The same could be said for all of Nvidia's partners. Once Fermi was released, no Nvidia-exclusive AIB's folded. Common sense.

Not pn performance per mm^2, unless GK104 is 50-60% faster than the 212mm^2 Pitcairn chip.

Pitcairn is an absolute beast in that regard. I was specifically referring to the "high end" although we all know that GK104 was meant to be and will eventually end up as Kepler's' mid-range. We shall see how well Kepler's performance per watt and mm^2 trickles down. Pitcarin will be very, very hard for GK106 to beat in both respects.

 

[jimbo75]

Nvidia didn't enjoy 50% gross margins on the entire year of 2011 on their single best selling chip (GF114) breaking even. The same could be said for all of Nvidia's partners. Once Fermi was released, no Nvidia-exclusive AIB's folded. Common sense.

GF114 wasn't their best chip that's why, and neither was it close to being their best selling chip.

Remove their professional market and how much profit did they actually make? It's pretty obvious that tesla and quadro have been subsidising geforce for the past couple of years.

 

[silent_guy]

GF114 wasn't their best chip that's why, and neither was it close to being their best selling chip.

Remove their professional market and how much profit did they actually make? It's pretty obvious that tesla and quadro have been subsidising geforce for the past couple of years.

Profit is irrelevant if you want to judge whether or not you should sell a chip. (Or at least start by defining 'profit' in this context.) What matters is gross margin.

Do you think GMs on GF114 were close to 0?

 

[jimbo75]

Profit is irrelevant if you want to judge whether or not you should sell a chip. (Or at least start by defining 'profit' in this context.) What matters is gross margin.

Do you think GMs on GF114 were close to 0?

I think they pushed the chip to it's limits in order to make it look a lot better than it was, but that doesn't come cheap.

On the flipside they got to name 6 different cards and 3 different chips "560" and made more money out of the confusion. For me the actual 560 Ti was probably a loss leader.

 

[tviceman]

GF114 wasn't their best chip that's why, and neither was it close to being their best selling chip.

Remove their professional market and how much profit did they actually make? It's pretty obvious that tesla and quadro have been subsidising geforce for the past couple of years.

No, you are considerably off base. GF104/GF114 was by far their best selling chip on Fermi. That isn't evn debatable. Nvidia makes an incredibly high percentage of profit in the professional space, but they do not sell enough GPU's in that market to completely offset their entire consumer-gpu category, nor do they sell enough mid-range chips in the pro space to sell the same mid-range at a loss or zero profit in the consumer space.

 

[jimbo75]

No, you are considerably off base. GF104/GF114 was by far their best selling chip on Fermi. That isn't evn debatable.

Sub $150 cards sell far more than the rest combined. This really shouldn't be a surprise to anyone as it's the same as with CPU's.

As far as I can tell the 560 Ti dropped to $180 lowest and was generally >$200.

If you meant the 560 was the best selling Fermi "card" however then sure...but that's only because there were 6 of them spanning 3 different chips.

 

[kalelovil]

I doubt that. GF108/118 most likely was by far, considering its low cost and all the laptop and desktop OEM designs its various incarnations featured in.
GT 415m, 420m, 425m, 435m, 525m, 540m, 550m, 555m, 630m, 635m, 420, 430, 440.
Nvidia's profit margin on it wouldn't have been great though.