Bookmark and Share


As Nvidia Corp. demonstrated its next-generation GeForce code-named Kepler in action at Game Developers Conference (GDC) and got applauds for impressive performance, a Chinese web-site published an image of what it claims to be GeForce GTX 670 Ti graphics card.

Epic Games, a major game designer, astonished the GDC12 attendees with its Samaritan demo, which depicts next-generation video-game graphics effects using Nvidia GeForce GTX "Kepler" hardware. Created in a dark, futuristic setting, the demo utilized a host of advanced rendering techniques that smoothly tessellated and morphed facial features, created realistic street scenes using point light reflections, and replicated the work of the best movie directors through the use of fine-tuned out of focus bokeh filters.

Back at GTC11, a similar demo was unoptimized and was run on three Nvidia GeForce GTX 580 graphics boards in multi-GPU mode. This time, however, at GDC12, the Samaritan (which is based on the tweaked Unreal Engine 3 demo) was rendered using one Nvidia GeForce GTX "Kepler" graphics cards. Potentially, this proves that the next-gen Kepler architecture is two or more times faster than the existing Fermi architecture in its model 580 form. The demo video is available below.

In the meantime, RedQuasar web-site has published an image of a graphics card that carries the GeForce GTX 670 Ti moniker and which is supposed to be one of the first Kepler GPU-based offerings.

The board has two MIO connectors, which are found only on high-end products, to provide three-way or four-way multi-GPU SLI configurations. The new adapter has DisplayPort, two dual-link DVI-I and HDMI display connectors. In addition, the card seems to have tricky PCIe power connectors previously found on an alleged Kepler-class graphics board.

Nvidia has very high expectations for its Kepler generation of graphics processing units (GPUs). The company claims that it had signed contracts to supply mobile versions of GeForce “Kepler” chips with every single PC OEM in the world. In fact, Nvidia says Kepler is the best graphics processor ever designed by the company. Unfortunately for Nvidia, yields of chips made using 28nm process technology at TSMC, including Kepler, are lower than the company originally anticipated and therefore their costs are high and Nvidia's chief executive recently complained that the company cannot get enough wafers.


Tags: Nvidia, Kepler, Geforce, Tesla, Quadro, 28nm


Comments currently: 40
Discussion started: 03/08/12 08:12:12 PM
Latest comment: 11/30/15 11:10:18 PM
Expand all threads | Collapse all threads


2-3 Times Faster than Fermi?

^ Most Definitely not true, at least not for the GK104 launching this month.

The demo this time was run using FXAA 3, not using Deferred MSAA as was the case for Tri-SLI 580s. FXAA 3 implementation is significantly less demanding than MSAA in deferred game engines (see Starcraft 2, Battlefield 3, for example).

Also, we have no information about the actual framerates or the resolutions Re: Tri-SLI 580s vs. 1 Kepler card to be able to conclude with confidence that Kepler is 2-3x faster.

For example, if the Demo was tested at 2560x1600 with 4x MSAA, GTX580 might have been memory swapping often due to only 1.5GB of VRAM. Secondly, there is at least a 30% hit in performance when using MSAA over FXAA 3. Thirdly, GTX580 Tri-SLI do not have 100% scaling. Therefore, 3x GTX580 does NOT equal 3x the performance from a single GTX580. Fourthly, EPIC and NV had at least 1 full year to optimize the demo, as well as to optimize the drivers to make it run more smoothly. We can't underestimate the power of optimizing drivers and application (see Consoles!).

I quote:

"The developers at Epic have optimized the code for its engine apparently so far that it now runs much more efficiently, although they have used in the new version, just a simple full-screen anti-aliasing (FSAA)."

There are too many variables at this point.

All will be revealed March 12th!
14 2 [Posted by: BestJinjo  | Date: 03/08/12 08:12:12 PM]
- collapse thread

100% correct.

I only want to add that the card in the photo carries a 570 Ti moniker not 670.
2 0 [Posted by: john_gre  | Date: 03/09/12 01:30:31 AM]
Quick update to my post above as new information has become available:

Per Epic's comments:

1) Tri-SLI GTX580s achieved equivalent framerates in 4x MSAA as did 1 Kepler card in 4xFXAA. (wow!)

2) Epic is able to achieve the greatest performance from the upcoming architecture as opposed to AMD's Radeon HD 7xxx series, an indication which bodes well for NVIDIA.
2 2 [Posted by: BestJinjo  | Date: 03/09/12 07:28:53 AM]
Or did Epic "fix" their code as implied to make Nvidia cars faster....its been done before.
0 0 [Posted by: DEADTIME  | Date: 03/16/12 07:49:54 AM]
"All will be revealed March 12th!"

Excuse my ignorance, but what's on March 12th?
1 0 [Posted by: sanity  | Date: 03/09/12 03:49:44 PM]
"NVIDIA will reveal details about GTX 680 in an event on 12th March, with a hard launch rumoured for 23rd March - days after HD 7800's release."

Read more:"[/i]
0 2 [Posted by: BestJinjo  | Date: 03/09/12 06:08:39 PM]

Let's see what Monday brings.
1 0 [Posted by: sanity  | Date: 03/09/12 07:01:29 PM]

All i'm seeing so far is a lot of smoke and mirrors from nvidia. whenever they are actually ready to provide an actual 600 series gpu so that we can actually see the benchmarks on their new gpu then i'll actually come to a conclusion on it.
5 0 [Posted by: SteelCity1981  | Date: 03/08/12 08:19:59 PM]
- collapse thread

Based on 95% of current leaks, GK104 = GTX670Ti OR GTX680, GK110 is delayed and will most likely be renamed as GTX7xx series or GTX685, w/e.

GK104 = 1536 SP @ 1400 mhz shaders = 4.3 Tflops vs. 3.8 Tflops for AMD. But NV's card only has 192GB/sec memory bandwidth, crippled by 256-bit bus.

GK104 isn't meant to destroy HD7970 and it won't, not with HD7970's 30% overclocking, 3GB of VRAM and 384-bit memory bus!

If you want 30-50% more performance over 7970, this card is likely to disappoint you. You have to keep in mind GK110 is the flagship, but that isn't coming until the summer, at the earliest.

My guestimate is that HD7970s will hold their own with overclocking. 30% overclocking on HD7970 is insane!
7 3 [Posted by: BestJinjo  | Date: 03/08/12 08:35:41 PM]
Well until i see actual benchmarks from nivida all i keep hearing is bark and seeing no bite. if they continue to hype the nvidia 600 series then it better live up to that hype or else anything less then that of all of this hype will be dissapointing to say the least.
2 0 [Posted by: SteelCity1981  | Date: 03/08/12 09:35:15 PM]
GK104 is not the flagship card of the Kepler series since those are designated for 500mm^2 die GPUs. It will be technically a flagship in the interim since NV has delayed GK110 way out into the summer (at least based on all the rumors). So if a GK104 that was never meant to go head to head against HD7970 can trade blows with HD7970, that's insane! That's like having a GTX560Ti beating an HD6970 by 10%. That also means the flagship Kepler coming later in the year will blow GK104 out of the water. That means AMD will bring an even faster HD7970 card. Good competition ahead.

The most critical aspect is the price. If NV launches this at $399-449, it's a game changer. If NV launches this at $549, then it's lame.
0 2 [Posted by: BestJinjo  | Date: 03/09/12 06:47:00 AM]
if and buts at this point. less barking and more biting from nvidia. put up or shutup for nvidia enough with the hype. Hype means nothing if you can't actually see the results to deliver.
2 0 [Posted by: SteelCity1981  | Date: 03/10/12 05:43:10 PM]
Nvidia have history of changing card specs at final time

GTX 480 / 590 is perfect example of this

480 have been refreshed till a part where clock speed of the gpu touch a part where it beat HD 5870

with 590 it is even worse

590 clock speed was taken to a part where it trade blows ** DO NOT BEAT HD 6990 ** with HD 6990 and when you try to overclock it . It either dies or become unstable .

What if GK 104 come at incense clock of speed of 1.1GHZ with 1.25V and no overclock headroom of any kind????

560 TI trade blows with 580 @ 1GHZ at 955MHZ it beat 570 with ease
0 0 [Posted by: pratkal9  | Date: 03/12/12 01:46:33 PM]
Yep 1200MHz overcklocking is insane, but then it's thee reasonable slurry selling point for DAMN and the reason why they kept its original specs at moderately low 925MHz. This way they can lure OCers into their camp and every HD7970 when the price drops cames wont be able to do same thing (or they would sell them as cheaper 7950s to keep oc customers happy 800MHz+30% ~1040MHz (ONLY) and even the heavily crippled chips could easily reach that Some 7930 won came very soon or would even 7930 when oced be good enough to leave GK104 in cloud of dust?!

Lets just praise that GK104 an nvidia could lower prices in tryouts to out compete DAMN. Thats what we really need not yet another king of the hill by some abysmal margin holding their price higher than HD7970
3 0 [Posted by: OmegaHuman  | Date: 03/08/12 10:41:51 PM]
If gk104 will perform in OC as gf104gf114 there is nothing to hold for a 7970. Also because AIB manufacturers will have the option to provide a 4GB version of the 680.
1 1 [Posted by: Marburg U  | Date: 03/09/12 03:42:14 AM]

show the post
0 3 [Posted by: pogsnet  | Date: 03/09/12 05:27:43 AM]
- collapse thread

Let HD7970 and GK104 battle it out at $450-550 now. By Q4 2012/Q1 2013 these cards will move into upper midrange segment and GTX780/HD8000 series will become new $550 champs. There is no need to assume that videocards will cost $1000.

It simply means current high-end cards will become mid-range as even faster cards are released, and so on. This is like GTX480 plummeting to $300 when the brand new 15% faster GTX580 launched at $499.
2 2 [Posted by: BestJinjo  | Date: 03/09/12 07:32:04 AM]

Kepler will be awesome.
2 2 [Posted by: beck2448  | Date: 03/09/12 07:51:50 AM]
- collapse thread

I have my fingers crossed this thing is a GPGPU/bitcoin mining monster. If true, I am ready for 2x GK110s in SLI. 2304 SP @ 1400mhz x 2 ops/clock x 2 GPUs = 12.9 Tflops of single-precision.

And if it overclocks another 20-30%, I'll **** ma pants.

3 2 [Posted by: BestJinjo  | Date: 03/09/12 09:37:47 AM]
Preliminary 3DMark 11 Scores on Extreme Presets:

GTX680 -> X3200 to 3300
HD7970 1240mhz/6620 GDDR5 -> X3300 (
HD7970 -> X2700 to 2800
GTX670Ti -> X2500 to 2600
HD7950 -> X2200 to 2300
GTX580 -> X2100

GK110 -> X4200+, to follow later in the year.

GTX680 beats a stock HD7970 in every DX11 game. GTX670Ti is rumored to be a $349-399 card, making $449 HD7950 look out of place.
2 2 [Posted by: BestJinjo  | Date: 03/10/12 08:14:20 AM]

2-3 times the performance of fermi lol!
6 2 [Posted by: redeemer  | Date: 03/09/12 12:59:43 PM]
- collapse thread

Not gaming, obviously, but single/double precision, definitely possible!

GTX580 only has 1.581 Tflops of performance (512 SPs @ 1544 Shader clock x 2 Ops/clock.

These new cards should be a god send for guys running distributing computing projects, bitcoin mining and other GPGPU tasks!
2 1 [Posted by: BestJinjo  | Date: 03/09/12 02:36:52 PM]
580 has 1024 SPs but only 512 of those SPs can process 64b/32b Scalar

256 64b Scalar "Floating Point"
512 32b Scalar "Floating Point"
512 64b Vector "Integer"
1024 32b Vector "Integer"

1024 * 2 * 1.544 => 3,162.112 GFlops
1536 * 2 * 1.8 => 5,529.6 GFlops

32b Vector^

1.8 GHz from the turbo boost rumour

64b Scalar - 680 GTX
512 * 2 * 1.8 => 1,843.2 DP GFlops

64b Scalar - 580 GTX
256 * 2 * 1.544 => 790.528 DP GFlops

1843.2 ÷ 790.528 => 2.33x
0 1 [Posted by: seronx  | Date: 03/10/12 12:42:07 PM]
Wait, what? All of your calculations are 100% wrong.

GF110 only has 512 SPs, each operate at double the data rate of the GPU. The GPU runs at 772mhz, which means the SPs run at 1544mhz.

Fermi does 2 Ops / clock, meaning it's single Precision is:

512 SPs @ 1544 x 2 = 1.581 Tflops Single Precision

Consumer Fermis have Double-precision performance artificially limited to 1/8th of SP.

So GTX580 has 198 Gflops of DP performance, while GTX480 has 1.35Tflops SP and 168 Gflops DP:

This explains 100% why Fermi tanks in all double precision distributed computing projects:

- Milkway@home
- Collatz Conjecture

It's single precision performance is also nowhere near 3.162 Tflops as you made it. Otherwise it would beat HD6970 in bitcoin mining. No...just no.
1 1 [Posted by: BestJinjo  | Date: 03/10/12 02:24:41 PM]
A "Fermi" CUDA Core has 2 ALUs. 1 INT and 1 FP
A "Kepler" CUDA Core has 3 ALUs, 1 INT and 2 FP*

*Assumed to be one more FP unit
0 0 [Posted by: seronx  | Date: 03/10/12 05:29:31 PM]
Let's just wait to see the specs. 1800mhz Shaders clocks @ 1536 SPs is very fast. Why would such a card barely beat HD7970 by 10% in BF3?

It would be hilarious if NV purposely fed wrong information and we see some insane monster card that lays down 1536 SPs @ 1800mhz shaders like you said. Just a day or 2 left to see!
1 1 [Posted by: BestJinjo  | Date: 03/10/12 09:12:24 PM]
LOL. Is that true?

It's crappytastic. Even DAMN and Intel running away from 3ALU design but nvidia are marching into club on 28nm level.

So now we'll have even more overheating friendly CPUs, pardon ex.GPUs.

JHH mast've found silicone holy grail and next time hell whip out space for two more ALUs and at least one more FPU in it's cuda (miracle/milagro) core.
0 0 [Posted by: OmegaHuman  | Date: 03/21/12 04:06:55 PM]

Is that photo real, How big is that fan?
All that power/performance and only "that" fan to cool it, I am getting a little worried folks, Really don't want another 480GTX jet engine!
I don't really care if the card costs an extra $50 just get the damn fan right so I don't have to modify or water cool it!
0 1 [Posted by: ozegamer  | Date: 03/09/12 05:43:45 PM]
- collapse thread

Don't worry. GTX570 and 580 use that same fan design and they are quiet.

The problem with GTX480 was not the fan, but the chip ran extremely hot, it had "leaky" transistors, and was manufactured on at that time immature 40nm process. As manufacturing node improves, the leakage of transistors improves and you can achieve similar frequencies at lower voltages. This card will be using a much more efficient 28nm process.

Also, the actual heatsink for 570/580 series replaced the heatpipes of the GTX480 with a Vapor Chamber Cooler:

Also, keep in mind this card is on 28nm process and the die size is about 320-340mm^2 vs. 529mm^2 for GTX480!

Also, it uses a far less complex 256-bit memory bus, which makes the memory controller inside the GPU less complex.

Power consumption should probably fall along the lines of GTX560Ti with performance faster than HD7970 in DX11 games.
2 2 [Posted by: BestJinjo  | Date: 03/09/12 06:17:51 PM]
Sounds like a winner, My 480 SLI set up was great in winter but cost me $2500 for an aircon for small bedroom in summer, Extreme cooling you might say!

If the top end card later on this year is quiet and heaps more the performance of 580GTX then it's a no brainer,May even go dual again.
1 0 [Posted by: ozegamer  | Date: 03/09/12 06:44:27 PM]
$2,500 for air con + electricity costs? That sounds excessive. What is this an industrial style air con? Where do you live? In the summer time I just open the window about 4 inches and put the PC close to it. I have the tower raised off the ground floor so that it exhausts straight out of the 2nd floor window of my house. Helps a lot
1 1 [Posted by: BestJinjo  | Date: 03/09/12 07:45:23 PM]
Nahhh no good here in summer, 40-45c temps open the window and you just let in more hot air which is usually hotter than the air indoors,Has to be well sealed/insulated house. Aircon was just a bonus for cooling pc mainly bought it to help me get some sleep.
No not industrial, Just an inverter type pretty energy efficient,Then again the outlay of the unit kind of cancels that out,But at least it's quiet.

Looking forward to these new 28nm cards,2012 is going to be great for both AMD and Nvidia users, Just hope some developers out there use some of that GPU power instead of console ports.
Tomorrow is D-Day, Nvidia finally spills some beans on the upcoming card! yessssssssss excited as a kid at Xmas.
1 0 [Posted by: ozegamer  | Date: 03/11/12 06:41:14 PM]
How big is that fan?

It's OVER 9000!!!
0 1 [Posted by: BestJinjo  | Date: 03/10/12 09:12:50 PM]

And over 110dB too?
0 0 [Posted by: OmegaHuman  | Date: 03/21/12 04:02:12 PM]

Unfortunately for Nvidia, yields of chips made using 28nm process technology at TSMC, including Kepler, are lower than the company originally anticipated and therefore their costs are high and Nvidia's chief executive recently complained that the company cannot get enough wafers.

Always same poor old bastard story to keep the prices sky high. So should we expect more than insanely 699USD for this new envy monster?

C'mon envydio we need cool gpus not overpriced frying pans. You know how to make a GPU in 7900GTX days. At least rehire some people that know what they're doing.

And JHH you could shoot yourself in toe, both of them. We dont wanna pay for your obsessed poker debts.

Potentially, this proves that the next-gen Kepler architecture is two or more times faster than the existing Fermi architecture in its model 580 form.

You could only dream that this is true. Unfortunately nvidas previous SLI incarnations are not that scaling friendly and not that ubiquitous as evy's single GPU solutios. TripleSLI is even more fail. Mqany test show that only good performing series 500 craps are 570GTX when it comes to SLi

Too much hype about yet another overpriced frying pan. Three years in a row?! That's ultimately sad (if we dust under the carpet 8800GTX and crappy 280GTX)
0 0 [Posted by: OmegaHuman  | Date: 03/21/12 03:55:20 PM]


Add your Comment

Related news

Latest News

Wednesday, November 5, 2014

10:48 pm | LG’s Unique Ultra-Wide Curved 34” Display Finally Hits the Market. LG 34UC97 Available in the U.S. and the U.K.

Wednesday, October 8, 2014

12:52 pm | Lisa Su Appointed as New CEO of Advanced Micro Devices. Rory Read Steps Down, Lisa Su Becomes New CEO of AMD

Thursday, August 28, 2014

4:22 am | AMD Has No Plans to Reconsider Recommended Prices of Radeon R9 Graphics Cards. AMD Will Not Lower Recommended Prices of Radeon R9 Graphics Solutions

Wednesday, August 27, 2014

1:09 pm | Samsung Begins to Produce 2.13GHz 64GB DDR4 Memory Modules. Samsung Uses TSV DRAMs for 64GB DDR4 RDIMMs

Tuesday, August 26, 2014

10:41 am | AMD Quietly Reveals Third Iteration of GCN Architecture with Tonga GPU. AMD Unleashes Radeon R9 285 Graphics Cards, Tonga GPU, GCN 1.2 Architecture