NVIDIA Kepler speculation thread

http://www.geforce.com/hardware/desktop-gpus/geforce-gtx-660-oem/specifications

nVidia announces GTX 660 (OEM), again GK104.
With all the delays etc, starts to look like GK106 has been canned.

Looks like you may be right and that Nvidia will handle the entire 600 line with just two parts. The GK104 and the GK107.

Also here is a link to an GT645 (OEM) part which seems to be a FERMI GPU with DDR5 memory:

http://www.geforce.com/hardware/desktop-gpus/geforce-gt-645-oem/specifications

If Nvidia decided to not make the GK106 then that would by itself be the reason for the GTX660 delay as the 28nm supply would really have to expand to allow for the much higher sales these mid-range cards would entail.

A couple questions:

#1 How much does Nvidia save in NRE costs in not producing the GK106

#2 Will the (GK104 or the GK107) handle the GTS650

GK106 isn't canned. Last two conference calls have stated that they are supply constrained, and last conference call hinted at Kepler products hitting all the important price points below $400. But when there isn't enough supply, it's best to allocate to whatever makes the most money. In this case, GK104 makes more money per wafer. GK107 is necessary to finalize all the design wins and supply all the OEM's / Vendors for the new laptop refreshes (new GPU models are almost always a must with laptop refreshes, hence the reason there are so many rebadges).

The gtx660 OEM exists because Nvidia is using as many dies as possible as a functional chip from their GK104 wafers. Nvidia binned so many different chips from both GF110 and GF104 that it got pretty ridiculous (4 desktop GF110 skus, 4 desktop GF104 skus). With GK104, now they have 4 skus coming from 1 chip, so it's safe to assume they are using almost the entirety of each GK104 wafer, with half (or more) of each chip in a $400+ video card (i.e. high dollar, high margin). GK106 will be, what, $220 at most? Almost half of what a gtx670 goes for. With the same PCB and vram costs as a gtx670, the margins are much lower and they won't pull in as much money per wafer even though the chip is smaller would yield more video cards per wafer than GK104.

When the constrain eases (which is probably going to happen very soon thanks to AMD's price cuts), Nvidia will get GK106 out. The rumors at the end of July were for a September launch. I still think that is going to happen.
 
Not only compared to the competition but also to NV's own Kepler family variants. GK107 is way too humble IMHO to cover let's call it the middle part of the 660 SKUs.

NV re-using GK104 even for 660 OEM SKUs could either mean that GK106 is facing unexpected delays or they've simply an unhealthy amount of GK104 chips (binning yields).

I think the current configuration of 660 Ti is a strong indicator against yield issues with GK104. Otherwise they'd have launched two SKUs - one with less shaders, one with less frequency.

I think it's rather because of limited wafer supply, that Nvidia wants to concentrate on only two chips for the time being - and from a company point of view this makes sense to me, especially given the rather high price level for 28nm chips in general.
 
I think the current configuration of 660 Ti is a strong indicator against yield issues with GK104. Otherwise they'd have launched two SKUs - one with less shaders, one with less frequency.

I think it's rather because of limited wafer supply, that Nvidia wants to concentrate on only two chips for the time being - and from a company point of view this makes sense to me, especially given the rather high price level for 28nm chips in general.

Why castrate N amount of 294mm2 chips if there are more units operational, instead of going into production with N% smaller cores where wafer yields will be X% higher than in the first case?

Yes 28nm manufacturing should still be quite expensive, but that's exactly the point for the above.

Note that in my book (which might be worded wrong = open to correction) "binning yields" don't stand for wafer yields, but for yields per core ie A% of chips with 15 SMXs, B% of chips with 14 SMXs, C% of chips with 13 SMXs etc.
 
Why castrate N amount of 294mm2 chips if there are more units operational, instead of going into production with N% smaller cores where wafer yields will be X% higher than in the first case?

Yes 28nm manufacturing should still be quite expensive, but that's exactly the point for the above.

Note that in my book (which might be worded wrong = open to correction) "binning yields" don't stand for wafer yields, but for yields per core ie A% of chips with 15 SMXs, B% of chips with 14 SMXs, C% of chips with 13 SMXs etc.

Just cause they are not viable in term of marketing... If Nvidia was coming with a 660TI, a 660 who is not competitive against AMD performance wise, there image will suffer a lot. They need something 6months later who is price wise competitive and performance wise competitive.

Hence why they up the chips to the superior sku...

GK106 660 become GK104 660TI - 660 .. GK106 castrated 650TI-650 become the full SP GK106 ( initially 660 ).

They dont have the choice, and i believe they have do the right one. This is the first time a company make 5 cards on the same sku from top range to middle range.
 
Last edited by a moderator:
I tend to agree that GK106 is most likely canned in preference of going directly to GK116. Mindshare is important to Nvidia and I doubt GK106 was anywhere near good enough to take on Pitcairn.

One look at newegg is enough to see that they are still sitting on a lot of unsold 560's so they probably weren't in a huge rush to get it out anyway.
 
nvidia didn't even launch the GK107 card with enough bandwith. it's the GTX 650, which shows liberal usage of the letter "X" again.
I wonder if it beats the "lameduck" gf106/gf116
 
I think the current configuration of 660 Ti is a strong indicator against yield issues with GK104. Otherwise they'd have launched two SKUs - one with less shaders, one with less frequency.
The GTX660 OEM version has it all, less shaders (only 6 SMx/1152 SPs) and less frequency (823/880 MHz base/turbo) than the GTX660Ti. That's where the bad dies end up. ;)
And probably in some Quadros, where half of the chip is deactivated.
 
GTX660 (OEM) - $200

Alienware has added the recently launched GTX 660 (OEM-only) graphics card as a choice in its X51 Mini Gaming Desktop

The GTX 660 OEM-only graphics card features a total of 1152 CUDA cores and works at 823MHz base and 888MHz Boost clock. It features either 1.5GB or 3GB of GDDR5 memory (Alienware decided to go for the 1.5GB model) clocked at 5.8GHz and paired up with a 192-bit memory interface.

The GTX 660 adds US $200 to the total price and according to Alienware, it should be enough to give 50+ frames per second in Battlefield 3 at 1080p resolution.
http://www.fudzilla.com/home/item/28481-alienware-adds-gtx-660-to-its-x51-mini-gaming-desktop
 
From VR-Zone: "No NVIDIA GeForce GTX 780 before March 2013, Maxwell Only in 2014?"

The Maxwell delay is well known, but the other rumors in the report seems interesting:
  • GK106 has been canceled or otherwise won't be released, neither will there be any sort of "GK116"
  • GK104/107 refreshes could have 25-30% higher performance over GK104/107
  • "Maxwell will be the first top-to-bottom GPU architecture, powering everything from Tegra to Tesla"

————————

So if there will be no GK106/116, I wonder how far GK104 will move down. Maybe to (a) GeForce model(s) with 4 or 5 SMXes?
 
If Nvidia is going to service their entire lower-mid to top end from one chip, they really have pulled a rabbit out of a hat. I suspect that we're about to see a complete cratering of their discrete desktop gpu share however.
 
If Nvidia is going to service their entire lower-mid to top end from one chip, they really have pulled a rabbit out of a hat. I suspect that we're about to see a complete cratering of their discrete desktop gpu share however.

Not one chip, one architecture, just what everyone has always on desktop gfx cards etc, but extended to use same architecture on mobile Tegras too
 
Not one chip, one architecture, just what everyone has always on desktop gfx cards etc, but extended to use same architecture on mobile Tegras too

I think he meant GK104, if GK106 really is cancelled then GK104 is pretty much it for the whole range on top of the low end.
 
If Nvidia is going to service their entire lower-mid to top end from one chip, they really have pulled a rabbit out of a hat. I suspect that we're about to see a complete cratering of their discrete desktop gpu share however.

Cratering seems quite excessive, but they do appear to be losing some market share: http://techreport.com/discussions.x/23482

GK104 can't go quite as high as Tahiti XT2, and can't be very profitable when pitted against Pitcairn Pro, so it's not an ideal situation, but it seems manageable; at least for now.

If and when the HD 7850 gets to $140 or below, I'm not sure how NVIDIA could respond.
 
Back
Top