New - Nvidia 560 Ti 448 Cores, ETA 11/29

[toyota]

Actually I think the opposite is happening. We often just look at shaders because it's the easiest # to remember/compare.

GTX460 vs. GTX285
Pixel fill-rate (# ROPs x clocks) = 21600 MPixels/sec vs. 20736 MPixels/sec
Texture fill-rate (# TMUs x clocks) = 37800 MTexels/sec vs. 51840 MTexels/sec
Memory bandwidth = 115.2 GB/sec vs. 158.976 GB/sec

On paper, GTX285 should crush the GTX460 without question. And yet, GTX460 is pretty much as fast!

Clearly, Fermi is far more efficient than GT200b was.

yeah but as for as actual SP, it looks awful when just 240 on the gtx285 can match the 336 of the gtx460 at nearly the same clocks too.

 

[happy medium]

Clearly, Fermi is far more efficient than GT200b was.

And a gtx680 Keplar will be 15/20% more efficient than Fermi.

"The 28LP process boasts a 20 percent speed improved over the 40LP process at the same leakage/gate. "

And the gtx780 (680 refresh) will be even faster.

"The 28HP process supports a 45 percent speed improvement over the 40G process at the same leakage/gate."

http://www.tsmc.com/english/dedicatedFoundry/technology/28nm.htm

 

[tviceman]

do gpus ever get faster clock for clock on the same amount of cores?

Yeah AMD did it with the hd6850/6870 when compared to the hd5850/5870. Nvidia has been using their unified core structure since G200. And again, Huang has been beating the performance per watt drum with Kepler pretty heavily. I think there is more going on than the standard node-shrink-power-draw-improvements with Kepler. Of course I don't have any numbers to back it up, but the fact that Kepler already has more design wins now than Fermi ever had at any time is pretty telling.

 

[ocre]

Yeah AMD did it with the hd6850/6870 when compared to the hd5850/5870. Nvidia has been using their unified core structure since G200. And again, Huang has been beating the performance per watt drum with Kepler pretty heavily. I think there is more going on than the standard node-shrink-power-draw-improvements with Kepler. Of course I don't have any numbers to back it up, but the fact that Kepler already has more design wins now than Fermi ever had at any time is pretty telling.

absolutely. If what i heard is anywhere close to true, kepler will impress. I believe it is much more than a die shrink. Its gonna be a monster in performance. But I am not sure when we will see the full blown Flagship kepler. There will be new chips soon though. They should be quite good too.

 

[Silverforce11]

I fully expect Kepler to be ~gtx590 performance. A full node shrink already gives potential for doubling, but add to that some efficiency tweaks and its obvious it will be very fast.

 

[ocre]

We often just look at shaders because it's the easiest # to remember/compare. As a result, it looks like a 336 SP card is barely faster than a 240 SP card. In reality, shaders are no more important than ROPs or TMUs are. The GPU needs to be well balanced or at any one point it can have "too much" of 1 thing and "too little" of another. Classic examples are bucket loads of memory bandwidth in 4870/4890 despite barely faster performance over HD5770 with nearly half the memory bandwidth. Another example is 5830 being way faster on paper than HD4890 was, but in real world, being completely ROP starved. And then we have HD6970 with 71% more texture fill-rate than a GTX580, but it can't really use that advantage because it's bottlenecked somewhere else.

In fact, ROPs can also be critical. When HD5830 cut the # of ROPs from 32 in 5870 to 16, its performance plummeted to barely faster than HD4890 despite having 1120 shaders @ 800mhz vs. 800 shaders @ 850 for the 4890. And despite having 56 TMUs and as much memory bandwidth, the 5830 was barely faster than the 4890. The GPU became almost entirely ROP starved.

GTX460 vs. GTX285
Pixel fill-rate (# ROPs x clocks) = 21600 MPixels/sec vs. 20736 MPixels/sec
Texture fill-rate (# TMUs x clocks) = 37800 MTexels/sec vs. 51840 MTexels/sec
Memory bandwidth = 115.2 GB/sec vs. 158.976 GB/sec

On paper, GTX285 should crush the GTX460 without question. And yet, GTX460 is pretty much as fast!

Clearly, Fermi is far more efficient than GT200b was. Or another way to look at it, GT200b had way too much memory bandwidth and texture performance than it could actually utilize. Comparing shaders across same generation makes a lot of sense (such as GTX460 vs. 470 vs. 570 vs. 580, etc.). But comparing shaders with previous generations can often lead to misleading / erroneous results since performance in ROPs and TMUs can have a dramatic impact on performance.

Then there are small improvements that often go unnoticed (because most games might not benefit from them) even in the same generation. GF110 improved FP16 performance. Some games benefited well from this such as Dirt 2 where GTX570 is 8% faster than a GTX480. Looking at specs of GTX480 vs. 570, there is no way the GTX570 should be leading by 8%.

Similarly, removing shaders may not necessarily reduce performance. HD6870 compensates with much higher clock speeds vs. the 5850 despite far fewer shaders.

+5
Cores/shaders are only for marketing packaging anymore. For consumers, they have reduce the architectures to just a few "comparable" numbers. its just no simple, it really cannot ever be.


Your very thought filled, informative post are a great treat. I enjoy seeing them in the seas of gibberish that has become the "tech" communities. Its nice to see post of a higher level. Very rare for sure.

 

[happy medium]

Soon very soon now......This is a FTW card.

Stock 797/1594/1950
http://3dmark.com/3dm11/2257244

Oc'd 868/1736/2200
http://3dmark.com/3dm11/2257278

 

[Fear No Evil]

I plan on picking one of these up and as soon as they are out. Hopefully Amazon will have them as I have a $100 gift card.

 

[MrK6]

I hope this thing falls in the $199-219 range. 2 of those in SLI will mop the floor with a GTX580. Still, considering 6950 2GBs were going for $230-240 in February/March....a $200 card that's barely faster is kinda meh. At this point, I am all ears about any Kepler or HD7000 news We are at the end of a 2.5+ year marathon since HD5870 launched in Sept 2009. The next generation of cards should be extremely powerful and overwhelm any game currently out, and make BF3 wish it had a high-rez texture pack!!

God I hope so. I was cruising for information on new cards and there still isn't a heck of a lot. I think the 58xx cards have easily replaced the 8800GTX as the longest running "best buy." Anyone who bought a a 5870 or 5850 on release for $400 or $300, respectively, has pretty much been set for the last 2+ years. Kind of sucks for us enthusiasts as the market hasn't had much to offer since then. Hopefully 28nm shakes things up.

If Kepler has 768 cores with 900mhz clock speeds, it's going to make GTX580's performance look mid-range. hehe

I'm hoping for 8800GTX-level peerless performance. From either camp.

 

[happy medium]

Soon very soon now......This is a FTW card.

Stock 797/1594/1950
http://3dmark.com/3dm11/2257244

Oc'd 868/1736/2200
http://3dmark.com/3dm11/2257278

A few words on temps and noise for the gtx560ti 448 core.

"So far so good, After 2 1/2 hours of BF3, temps maxed out around 65c with the fan never hitting more then 50% speed."

 

[toyota]

its simply just a cut down gtx570 so of course temps will only be a couple degrees lower at best.