Nvidia Corp. on Thursday unleashed its new graphics card based on Kepler architecture that the company positions as an ultimate solution for gamers. The novelty is based on the well-known GK104 chip with 1344 stream processors and thus provides excellent performance for modern video games. However, given positioning of the GeForce GTX 660 Ti product, the price tag of $299 seems to be too high.
Nvidia GeForce GTX 660 Ti graphics card is based on a cut-down version of the GK104 graphics processing unit (GPU) made using 28nm fabrication process, the same chip that is used on high-end models GTX 670 and 680. The model GTX 660 Ti has 1344 stream processing units, 112 texture units, 24 render back ends, 7 tessellation engines and 192-bit memory controller. Nvidia clocks the GK104 chip of GTX 660 Ti at 915MHz and recommends partners to install 2GB of GDDR5 memory clocked at 6008MHz onto the GeForce GTX 660 Ti graphics cards.
Given its Kepler origins, the GeForce GTX 660 Ti supports contemporary capabilities like DirectX 11.1, OpenGL 4.2, OpenCL 1.2, stereoscopic-3D, 4 multi-monitor output, PhysX, 4-way multi-GPU, PCI Express 3.0 and so on.

According to Nvidia itself, the GTX 660 Ti is 41% faster on average than the GeForce GTX 560 Ti from 2011 and can successfully compete against AMD Radeon HD 7950 that has recommended price-tag of $349. According to performance tests conducted by X-bit labs, in some cases the GeForce GTX 660 Ti is faster than the Radeon HD 7950, but in some cases the GTX 660 Ti fails to beat even its direct rival, the Radeon HD 7870 1GHz Edition. One advantage that the GeForce GTX 660 Ti clearly has over competing solutions from AMD is relatively low power consumption (no more than 150W in peak case scenario) and resulting short design of printed-circuit board, which means that the GTX 660 Ti can be installed into almost any computer case.

Nvidia releases its Kepler-series graphics card with $299 price-tag over five months after its arch-rival AMD first released its Radeon HD 7800-series "Pitcairn" GPUs. Therefore, it is not truly surprising that the GeForce GTX 660 Ti manages to beat the Radeon HD 7870 GHz Edition in loads of benchmarks. What is surprising is that Nvidia decided to address the market of graphics cards for gamers with the same chip it uses on enthusiast-class products. It is not a secret that TSMC charges AMD and Nvidia more for 28nm chips than it did for 40nm GPUs, which calls for development of specialized graphics processors for each market segment as despite of relatively small die size, the GK104 just cannot efficiently serve all markets from $299 to $999.
Another thing to observe is that due to increase of chips' price, both AMD and Nvidia now position graphics cards with $299 price-tags for gamers. Previously, the sweet-spot for gamers was something between $149 and $249. It remains to be seen whether gamers will actually start to buy more expensive solutions to play the latest titles or will wait for price-cuts or introduction of new products.
The Nvidia GeForce GTX 660 Ti GPU is available now from numerous add-in card suppliers, including ASL, Asustek Computer, Colorful, Elitegroup Computer Systems, EVGA, Gainward, Galaxy, Gigabyte, Innovision 3D, Jetway, Leadtek, MicroStar International, Palit, Point of View, PNY, Sparkle and Zotac.
For a limited time, gamers who purchase a GTX 660 Ti GPU from a participating retailer or e-tailer will receive a voucher for a free copy of Gearbox Software's Borderlands 2, which is launching next month.
Tags: Nvidia, Geforce, Kepler, GK104, 28nm, TSMC, AMD, Radeon, Pitcairn
Comments currently:
8
Discussion started: 08/16/12 06:14:42 PM
Latest comment: 08/19/12 03:57:58 PM
Expand all threads |
Collapse all threads
[1-4]
1.
I think what really happened is 2 separate teams were working on Kepler:
- GK104 and GK110
The team that worked on GK104 did such a good job that NV realized they could use that chip as a GTX670/680 against the 7970. That played well into NV's hand since GK110 is a massive chip, hot, power hungry and too expensive to manufacture on the new 28nm TSMC process and sell at $500. Instead, NV decided to post-pone that GK110 chip to Q4 2012 and sell it as K20 Tesla card to professionals for thousands of dollars. What this has allowed them to do is stay profitable on GK110 and still have the profitability and performance on GK104.
Since Tahiti XT sacrificed transistor space for GPGPU/double-precision, Kepler as a mid-range GF104 successor came out lean and energy efficient. After that, NV waited 6 months to build up failed GTX670 chips and we are finally seeing GTX660Ti. If NV didn't repackage the 1 failed memory controller/ROP GK104s, they would have had to throw them out. Surely it's better to sell them than to throw them out. Hence why it took so long to build up inventory of all the failed 670 chips.
Also, GTX680 was roughly just 35% faster than the 580, the lowest performance increase from 1 generation to the other in NV's history! Never in the history of NV did the new flagship have less memory bandwidth than the old one (GTX680 has the same as the 580). Again all the signs are there. GK104 is also lacking dynamic scheduler for compute, none of the double precision functionality of GF100/110. This is odd despite the fact that NV pushed GPGPU with G80, GT200B and GF100/110 Fermi. All of this further supports the view that GK104 was never meant to be high-end.
Finally, there are other hints from driver mistakes, packaging mistakes, and labels that hinted at GTX680 originally intended to be GTX670 Ti only.
http://www.techpowerup.co...X-680-as-GTX-670-Ti-.html
The fact that NV blocked voltage control further hints that GK104 is already being pushed to the limit. It's almost as if NV is using a GTX460 OCed (680) to compete against the 5870 (7970).
At the end, it doesn't really matter since GK110 (GTX780) hasn't launched this year. However, whether or not GK104 was always meant to be high-end will remain a mystery until 2013. If NV launches GK110 (aka GTX780) with full double precision performance and dynamic compute, then we'll know with 100% certainty that NV held back GK110 for manufacturing capacity / cost / profitability reasons and used GTX670/680 (GK104) as high-end because AMD under-delivered with Tahiti XT.
It makes sense strategically. GK110 is probably a massive 500mm^2 die chip. Why sell that for $500 when they can sell 294mm^2 GK104 for $500 and push those 500mm^2 GK110s as K20 Tesla cards for 5-7x more!
show the post
3
9
[
Posted by: BestJinjo

|
Date: 08/16/12 06:14:42 PM]
+ expand thread (1 answer)
- collapse thread
That's not the complete picture, although you do make a number of valid points. Not selling GK110 as a GeForce product, which I'm sure that they will eventually do (though at a much higher price than $500), would leave them without the ability to select the best GK110 chips for Tesla K20 cards (by yield culling). That would force them to either adjust K20 specifications and performance downward or take a massive hit to profitability as most cards would not be able to function with the specified number of CUDA units or at the specified clock frequencies.
However, with Intel MIC looking very promising and reportedly taking a number of large-scale supercomputer contracts from Nvidia, I'm not sure how profitable Tesla K20 will be whatever route they take.
0
0
[
Posted by: lol123

|
Date: 08/18/12 10:13:01 AM]
2.
Regarding the move to $299 price level from $199-229 for 460 768/1GB, I think it's a response for the cannibalized sub-$100 GPU space. Intel's and AMD's APUs have pretty much obsoleted budget gaming cards. NV will want to make up margins elsewhere which is why it makes a lot of sense to try and move the prices. Of course it depends on whether or not the consumer will respond to price increases. It is working well for them so far though.
With current generation of consoles looking very dated, consumers appreciating the convenience and prices of Steam/Origin, a lot of people are getting back to PC gaming. This is especially true in places like Russia, Brazil, India and Asia where PC gaming is expected to fuel hardware upgrades.
In its 33 country analysis of the gamer market, JPR has found very strong demand in the BRIC countries for systems, accessories, and upgrades approaching $4.7 billion in 2012 and growing to $7.7 billion by 2015.
http://www.techpowerup.co...-Jon-Peddie-Research.html
show the post
2
10
[
Posted by: BestJinjo

|
Date: 08/16/12 06:25:38 PM]
3.
Nice analysis.
The only question left out is will 660TI push Radeon prices down the road?
0
1
[
Posted by: Azazel

|
Date: 08/16/12 08:28:55 PM]
4.
300$ for this? You serious???
show the post
1
4
[
Posted by: TAViX

|
Date: 08/17/12 12:58:06 AM]
+ expand thread (1 answer)
- collapse thread
The best case scenario is always going to be that the entire 192bit bus is in use by interleaving a memory operation across all 3 controllers, giving the card 144GB/sec of memory bandwidth (192bit * 6GHz / 8). But that can only be
done at up to 1.5GB of memory; the final 512MB of memory is attached to a single memory controller. This invokes the worst case scenario, where only 1 64-bit memory controller is in use and thereby reducing memory bandwidth to a much more modest 48GB/sec.
http://www.anandtech.com/...force-gtx-660-ti-review/2
Not a real 2GB card? Have to spend $30 more a 3GB version....

show the post
1
10
[
Posted by: BestJinjo

|
Date: 08/19/12 03:57:58 PM]
[1-4]