NVIDIA Kepler speculation thread

NVIDIA GeForce GTX 750 Ti pictured and detailed

Source

The GTX 750 TI would be equipped with 960 CUDA cores, 80 TMUs and 32 ROPs. So technically this is a faster GTX 660 (has more ROPs — 32 vs 24), not necessarily a rebrand.
Additionally the clocks are not the same as on GTX 660, we are looking at 1033 MHz base clock and 1098 boost clock. Although this might just be the sample that was leaked (it doesn’t look like a reference board).
GeForce GTX 750 Ti has 2GB GDDR5 memory across 256-bit interface (another major change in comparison to GTX 660). With a memory clock at 1502 MHz, we have a bandwidth of 192.3 GB/s.
So is this a modified GK106 or a chopped GK104? even more so than the GTX 760. :???:
Also i doubt this will be put against the $199 R9-270X, so potentially a $149 product?
 
I don't remember any x4 or xx4 desktop GeForce part being cut down this much in terms of CCs, but still I'm expecting this (if the specs are real) to be a GK104, since I don't see any reason why a 192-bit bus and 24 ROPs are not part of the full configuration for GK106.
 
Also i doubt this will be put against the $199 R9-270X, so potentially a $149 product?
Why? Even if it's really GK106, the chip is a bit larger and therfore slightly more expensive to make than Pitcairn, so why sell it for much less?
The card will be roughly on par with 270X in terms of performance, so anything less than 189$ would surprise me, it would cut their margin more than necessary.

I don't remember any x4 or xx4 desktop GeForce part being cut down this much in terms of CCs, but still I'm expecting this (if the specs are real) to be a GK104, since I don't see any reason why a 192-bit bus and 24 ROPs are not part of the full configuration for GK106.
Nvidia launched the 192-bit GF106 as a 128-bit product at first, too (GTS 450). Maybe NVidia just likes to keep an ace up their sleeve.
 
I don't remember any x4 or xx4 desktop GeForce part being cut down this much in terms of CCs, but still I'm expecting this (if the specs are real) to be a GK104, since I don't see any reason why a 192-bit bus and 24 ROPs are not part of the full configuration for GK106.

660TI with full memory bus. ( 256bit ).. ( as the 660TI was a GK104. )

I bet for a 750TI, GTX 760 with 2 SM lost, should end between the 660 and 660TI performance wise depending the turbo clock speed. ( lol, not seen the graph with % perf, well look right ).

( at this rythm they will do us a 768SP GK104 ( half SM disabled ).
 
Last edited by a moderator:
So, a GK104 again? It looks like nvidia just churns out many of them and the dies can simply be sold on many products. It's their new G92.
GK106 is the red-headed child, just like G94 was : it's not incredibly smaller (a quick check tells GK104 is 294 sq. mm, GK106 is 221 sq. mm). It's more successful though, moves a lot of units on the midrange - 650 ti and 660 are cheap and fast. It is useful in laptops too.


Nvidia launched the 192-bit GF106 as a 128-bit product at first, too (GTS 450). Maybe NVidia just likes to keep an ace up their sleeve.

They launched the GK106 in 192bit parts first, then 128bit later ;)
 
The 270X is not pitcairn, it is Tahiti(7950) on a 256bit bus.

Also it is worth pointing out that the recently released firestrike score of a 290X being used with a 3960X scored ~9250 points, where as the AMD slide showed it around 8000, the discrepancy is most likely due to AMD using their FX9590 CPU.
That makes all their scores actually much lower than what they should be, including the 5500 score of the 270X.
Also a 7870 @ 1150Mhz core/1400Mhz memory using a 3770k @ 4.2Ghz scores 5550.


The 270X will be competition for the GTX 760, the 760 will receive a price drop to $199.

So where does that leave the 750 Ti? most likely at the $149 range.

Edit; Ignore this.
 
Last edited by a moderator:
Additionally, Gibbo made this comment about the 270X The comment has since been edited/removed from the thread on the overclockers website.
I suggested early on 270/x was Curacao but it would make sense to use the Tahiti LE.
 
Last edited by a moderator:
The 270X is not pitcairn, it is Tahiti(7950) on a 256bit bus.
BIOS says Curacao XT: http://www.techpowerup.com/vgabios/146203/ati-r9270x-2048-130416.html


GTX 780i: 1920SPs@1033/1150MHz + 384-bit@6Gbps?

R9 290 counter or 560Ti 448 style card, to get rid off <12SM GK110 chips?
Sweclockers report about a new GK110 SKU launching 16/17th October: http://www.sweclockers.com/nyhet/17706-nvidia-forbereder-nytt-geforce-med-gk110
... also a 3GiB Titan could be launched.
 
(the last couple of posts and this needs to be moved to the volcanic island thread)

1400mhz 256 bit memory doesn't really smell like a chopped tahiti either - then it would likely be higher. But it's on the other hand quite high for a "tweaked pitcain".
 
(the last couple of posts and this needs to be moved to the volcanic island thread)

1400mhz 256 bit memory doesn't really smell like a chopped tahiti either - then it would likely be higher. But it's on the other hand quite high for a "tweaked pitcain".
My post was about why i think the 750Ti will be priced $149.

Edit: Nevermind; The 750 Ti will be $199 due to the 270X being pitcairn.

 
Last edited by a moderator:
Concerning the ROP count on the rumoured 750 Ti; I was unaware that Nvidia tie their memory controllers with their ROP Units.

This is from the 660Ti review, i don't believe it is mentioned in the 660 review, but seems likely to be the same.

This easily explains the ROP count in the rumoured specs; Though this is probably old news for most people, I'm a little late to the party. :oops:
 
Concerning the ROP count on the rumoured 750 Ti; I was unaware that Nvidia tie their memory controllers with their ROP Units.

This is from the 660Ti review, i don't believe it is mentioned in the 660 review, but seems likely to be the same.

This easily explains the ROP count in the rumoured specs; Though this is probably old news for most people, I'm a little late to the party. :oops:
Indeed this is REALLY old news, nvidia has done it this way since basically forever (the same basic arrangement since at least g80, but I'm quite sure just about all chips before also worked like that, the only difference is that some chips have one quad-rop tied to each 64bit channel whereas others have one octo-rop tied to each 64bit channel).
 
http://www.geforce.com/whats-new/ar...day-bundle-featuring-heroes-pirates-and-spies

vRRa7ac.png
 
In nvidia's recent driver annoucement ( https://devtalk.nvidia.com/default/...inux-solaris-and-freebsd-driver-331-17-beta-/ ) there's support for two new denominations :

* GeForce 705A
* GeForce GT 730A
So the first one is old gen (like GeForce 405 was), it's yet another GF119. Techpowerup database says it has a 15W TDP, pretty cool (in a thermal sense) if verified

GT730A should be a GK208 but the details I found are for GT730M instead. I wonder what the exact characteristics are and if the price will get low ; and especially if it does have a displayport or minidisplayport as standard.
What is the "A" for, though? :p. I hope it doesn't mean it's a Fermi and that they're pushing GF108 again.
 
A SKUs are for All-in-One PCs, so its soldered or MXM.
You can buy GK208 in GT 640 GDDR5 and GT 630 (newest revision DDR3 64-Bit).
 
It is not sold in my country or continent though. Launching it as a desktop GT730 card would fix that, and there'd be lower price, more cooling options (like a full height passive card, or 80mm fan at least)

GK107 with displayport might be a nice board too (compatible with G-Sync, lol)
 
Back
Top