Epic demos Kepler GPU, could be faster than three GeForce GTX 580s

[Shawn Knight]

A recent demo by Epic Games at the Game Developer Conference in San Francisco suggests that Nvidia’s upcoming Kepler GPU could be faster than three GeForce GTX 580 graphics cards. An Nvidia blog post notes that the graphics in the Unreal Engine 3 tech demo are so advanced that it was recently regarded as a fringe demo and required no fewer than three GTX 580 cards to run.

The Samaritan demo is no doubt impressive but Epic Games vice president Mark Rein said that the team is able to get so much more out of the card that what was shown in Samaritan. Behind closed doors, Epic showed off the Unreal Engine 4, reportedly powered by a single Kepler GPU.

Epic Games unveiled the Samaritan demo at last year’s GDC. The tech demo impressed onlookers as advanced rendering techniques smoothly tessellated and morphed facial features and created realistic scenes using three GTX 580 cards.

Kepler includes Fast Approximate Anti-Aliasing (FXAA), an anti-aliasing technique designed to improve upon Multisample Anti-Aliasing (MSAA) that’s commonly found in many games today. 4x MSAA in Samaritan’s lighting pass consumed close to 500MB of GPU memory. As a shader-based technique, FXAA doesn’t require additional memory which makes it much more performance friendly in the demo. The extra memory allows developers the option to reinvest it in additional textures or other niceties.

FXAA is also noticeably smoother than MSAA, as shown in the demo embedded above.

Permalink to story.

https://www.techspot.com/news/47761-epic-demos-kepler-gpu-could-be-faster-than-three-geforce-gtx-580s.html

 

[MrTomTom]

Pictures or it didn't happen.

 

[captainawesome]

that video is impressive but no way the in-game graphics is gonna be near that!

 

[Guest]

You could buy a Kepler GPU card to run Samaritan demo at a smooth frame rate or instead with the same money you could buy the new iPad and run Samaritan demo on it at 0 FPS. hahahahaha

 

[Guest]

What happened to the paper launch that was supposedly today? Or is it later today?

 

[Cinders]

Wow, faster then three GTX 580's. That would be a huge leap forward if it's true. I wonder what the power requirements would be. :|

 

[Guest]

1) Nvidia and EPIC had more than a year to optimize the drivers/game code to make the game run faster.

2) EPIC stated before that 2x GTX580s were used for Graphics and the 3rd was used strictly for Physics last year.

3) GTX580s in Tri-Fire do NOT have 100% scaling. Therefore, in principal that's not 3x the performance of a single 580.

4) FXAA 3 is FAR less GPU intensive than deferred MSAA Mode used in the last year's presentation (just look at 40%+ performance loss in Battlefield 3 with 4x MSAA deferred).

FXAA is a post processing filter, not "real" anti-aliasing. It blurs textures.

Anyway,

The specs for GTX680 are right here:

http://videocardz.com/30951/nvidia-geforce-gtx-680-to-be-released-on-march-22nd-available-for-549

Based on the $549 price, only 320mm^2 die size and 256-bit memory bus, this card was only intended to be GTX670Ti. Since it's fast enough to compete with HD7970, TSMC is having 28nm capacity/yield issues and NV is late with the Large-die flagship, they relabelled it to a GTX680.

The "flagship" is delayed and likely won't be available until at least the summer. For now, NV will simply cut the prices of GTX560Ti and GTX570 as there is no replacement for them for the time being.

Overall, meh since NV is essentially selling upper midrange performance for $550. However, given how much more efficient this 320mm^2 256-bit card is, we can expect amazing performance from GTX780 or whatever the "large-die" Kepler is called. Too bad, it's way out on the roadmap...

 

[amstech]

I didn't think Nvidia would swing back this hard until the GTX 780.

 

[MrAnderson]

For this comming cycle of generation of PC gaming, it will take lots of work to get games to look like that espcially with CPU cycles and even GPU cycles possibly going to other things...

I just hope that the power consumption for performance is going down. Having Killiowatt power supplies is getting rediculous. I cannot wait to see what demo Nvidia puts out. I miss those days when Nvidia and ATI used to create short realitime films to show off their tech.

 

[dividebyzero]

Wow, faster then three GTX 580's. That would be a huge leap forward if it's true.

It isn't. Last year's demo (w/ tri-SLI GTX 580) used multisampling antialiasing which imposes a large penalty on video RAM buffer. FXAA is basically free.
SLI also does not scale linearly, and the game is likely much better optimized as the Guest posted below you.

Based on the $549 price, only 320mm^2 die size and 256-bit memory bus, this card was only intended to be GTX670Ti.

Sounds like opinion being paraded as fact. Isn't it more likely that the 670Ti is a salvage part, and that the 680 is the fully operational at target voltage? Or do you expect all Nvidia's die's to be either 100% functional or non-working. Using that logic, the GTX 570 and 560Ti(448SP)/ 560Ti OEM wouldn't exist.
Die size isn't strictly relevant in relation to previous (Fermi) design, or possibly Tahiti, since at least one of those (and possibly both) used automated tools to layout the chips. This process isn't noted for delivering an area efficient design (see AMD Bulldozer)

TSMC is having 28nm capacity/yield issues

Really? You do realise that AMD's HD 7970, 7950, 7870, 7850, 7770, 7750 as well as Qualcomm (S4), Xilinx and Altera are shipping for revenue on TSMC's 28nm processes (28nm HP, HPL and LP)...it hasn't stopped any of them producing a chip. The fact that Nvidia's large chip (GK110) is still first-and-foremost for the professional market (remember also that Cray has first dib's on a sizeable percentage of the first production run), it's more likely that die area, memory controller, cache, and performance per watt are more problematic than capacity or process.

For now, NV will simply cut the prices of GTX560Ti and GTX570 as there is no replacement for them for the time being.

Since the GTX 680 is a direct replacement for the GTX 580, you'd think that the 580's would be the ones being price cut/EOL'ed -since there is plenty of room to lower 580 pricing without upsetting the $US 320-360 pricing of the next-tier SKU (GTX 570). Strangely enough, this is exactly what seems to be happening

Since it's fast enough to compete with HD7970,....Overall, meh since NV is essentially selling upper midrange performance for $550.

So what's more meh? AMD's single fastest non-dual card (HD 7970) supposedly offering only "upper midrange performance", or Nvidia matching that performance and price ?

What happened to the paper launch that was supposedly today? Or is it later today?

There was no launch scheduled for today. The only thing I've seen referencing the 12th March and Kepler, was a "story" from Charlie "Can't spell Charlie without L-I-E" D over at SemiArticulate. My guess would be that Charlie put up the story in order to accumulate page-hits and get his name in print, and once the story was repeated often enough by secondary sources it went from being parsed as "satire"/"ramblings cobbled together from musings on Chinese tech sites" to "fact" by some people. Charlie then could follow up the first "story" with an "Nvidia misses 12th March deadline, is late, deathknell only a matter of days away" front page article.....then again, if you're a devotee of the Cult of Charlie, you just might believe that the worlds largest supplier of discrete graphics cards deliberately held back a launch to make Charlie "seem wrong"

 

[St1ckM4n]

That video didn't impress me. At all.

I thought we had all this stuff 2 years ago. And the textures were poor.

 

[soldier1969]

I'll take a single 4GB GTX 680 to replace my 2 GTX 580s 3GB each please...

 

[hahahanoobs]

Guest said:

Based on the $549 price, only 320mm^2 die size and 256-bit memory bus, this card was only intended to be GTX670Ti. Since it's fast enough to compete with HD7970, TSMC is having 28nm capacity/yield issues and NV is late with the Large-die flagship, they relabelled it to a GTX680.

Word is, 670 Ti was an internal name.
http://www.fudzilla.com/home/item/26289-nvidias-upcoming-card-is-gtx-680

Same Guest said:

Overall, meh since NV is essentially selling upper midrange performance for $550.

It's priced the same as what nVIDIA said is its AMD equivalent - the 7970.

 

[hahahanoobs]

I have a hunch we will eventually see a GTX 685 in addition to the GTX 680.

 

[Guest]

dividedbyzero,

"Die size isn't strictly relevant in relation to previous (Fermi) design"

If you look back at Nvidia's GPU history, you'd note:

1) Huge memory bandwidth increases over previous generation high-end chips (GTX680 aka GK104 brings none over GTX580)

2) Large die size chips designate high-end parts, a part of their large monolithic die strategy (i.e., 450mm^2+)

3) Performance increase on average of 50-100% vs. previous high-end. GK104 is unlikely to beat GTX580 by an average of 50%

ALL rumors even going back 6 months ago pointed that NV will counter HD7970 with a GK104 style chip, and release a much larger GK110 (or some called it GK112) much later in 2012. It seems this is becoming true. NV is unable to launch their flagship on time. They have no choice but to launch GK104 as a placeholder.

The fact that NV's GK106/107 are also no show just clearly goes to show NV isn't ready to launch Kepler series. 28nm yield issues and capacity constraints are widely documented and have even been sighted in Investor Financial Analyst calls by NV.

Go read some sources and you'll know that TSMC will not ramp up production of 28nm Fab until later in the year, around Q3. It's simply TSMC's own timeline. The fact is both NV and AMD held back on performance due to slow ramp up of 28nm process. Prices are also pressured since 28nm wafers cost MORE than 40nm wafers did. But all that means is consumers are better off waiting for 6 months, which shouldn't be hard to do given that 99% of games in 2012 for PC are console ports.

The fact that GK104 will hardly beat HD7970 by more than 10% is a 100% indication that it's not NV's high-end because HD7970 beats GTX580 by only 25%! Nvidia simply couldn't deliver the real flagship and now consumers are going to be stuck with $550 cards that normally would only be $399 or so given their performance levels vs. previous high-end.

 

[dividebyzero]

dividedbyzero,
"Die size isn't strictly relevant in relation to previous (Fermi) design"
If you look back at Nvidia's GPU history, you'd note:
1) Huge memory bandwidth increases over previous generation high-end chips (GTX680 aka GK104 brings none over GTX580)

Immaterial to the argument. Moreover, bandwidth isn't an accurate measure of performance. Let me illustrate:
GTX 480 : 384-bit / 8 x 3696 MHz effective = 177.4 GB/sec
HD 6970 : 256-bit / 8 x 5500 MHz effective = 176 GB/sec
and if the performance difference isn't readily apparent with that, try this:
HD 2900XT: 512-bit/ 8 x 2000MHz effective = 128GB/sec
GTX 560Ti : 256-bit/ 8 x 4008MHz effective= 128.27GB/sec

2) Large die size chips designate high-end parts, a part of their large monolithic die strategy (i.e., 450mm^2+)

And ? I don't believe anyone is arguing that Nvidia doesn't, at this point in time pursue a big-die strategy. That is pretty much self evident with the need for a wide buswidth memory controller (more important for compute than gaming), on-die cache, provision for 72-bit ECC memory etc.

What I would argue is that having been burnt on a big-die first strategy on 40nm with the exceedingly late to market (and with no full 512 shader die) GTX 480/470, Nvidia probably erred on the side of caution by going with a mid-sized die on a new 28nm process-likely given TSMC's problems (and eventual abandonment) of 32nm. It's hard enough ensuring a new architecture on a new node works out well, without having to factor in possible process problems with your sole foundry partner. Why complicate matters with a huge die that would take a correspondingly long time to debug.

Performance increase on average of 50-100% vs. previous high-end. GK104 is unlikely to beat GTX580 by an average of 50%

How often do you see a 100% increase in performance- practically never. Making up arbitrary numbers and calling them fact doesn't make the point (however vague) you're trying to make valid.
8800GTX/Ultra to 9800GTX, 9800GTX to GTX 280, GTX 285 to GTX 480..none represent a 100% increase in performance. You could possibly argue for a 7900GTX/GTO to G80 in some circumstances...but kind of goes against your "big die" hobbyhorse ( G71 being 196mm²)

They have no choice but to launch GK104 as a placeholder

Wrong. GK104 seems to have always been a part of the lineup.

The fact that NV's GK106/107 are also no show just clearly goes to show NV isn't ready to launch Kepler series

You probably need to pay closer attention to whats going on. GK107 has been in the wild somewhat longer than GK104 (was going to link to the earlier video benchmarks that have been doing the rounds, but they've been pulled)

28nm yield issues and capacity constraints are widely documented etc, etc...

And I repeat. It.Hasn't.Stopped.Anyone.Releasing.A.Chip. Being constrained on wafer production and/or yield is one thing- being unable to produce a chip at all is entirely another. Considering TSMC's 28nm production (including low-power) is around 2% of their total wafer output I'd be a little surprised if they could provide for the whole market. You might also want to consider that Nvidia surrendered a portion of it's 28nm wafer starts. If GK110 was ready to go now, then they would quite simply sacrifice the GK104 to go with the high-value compute card. AMD sure as hell don't have more wafers being baked at TSMC than Nvidia- it hasn't stopped people from being able to buy HD 7000 cards.
Everything points to a 23rd March launch for the GK104, so Nvidia's AIB's obviously have boards up and ready to go

Nvidia simply couldn't deliver the real flagship and now consumers are going to be stuck with $550 cards that normally would only be $399 or so given their performance levels vs. previous high-end.

Does.Not.Compute.
Nvidia will price relative to performance. If performance is broadly equal to an AMD card then they price accordingly, with probably a tariff on top allowing for the brand. Rory Read and AMD have set the price/performance bar as far as 28nm goes.Nvidia simply wouldn't sell a card that outperforms a card already on the market for less. It makes no sense, any more than AMD releasing the HD 7970 at the same (or lower) price as a GTX 580 which it has a comfortable performance margin over....and since people are buying HD 7970's at $549+ it proves the market exists for performance at that price. Q.E.D....and the only way that changes is if one company decides to institute a price war and slashes prices. The consumer has determined what these cards will sell for. AMD and Nvidia will exploit that.

You might note that Nvidia's recent second-tier cards have had only one SLI connector, with a maximum of two cards supported in SLI (GTX 460, 560). The 680 features two SLI connectors, which indicates at least three, and probably, four card SLI supported. As is usual for the top-tier SKU's. When was the last time a top tier Nvidia card debuted at $399 MSRP ?

BTW: A GK110 at the rumoured specs. No way it retails at $550 or anywhere close. The market/signed contracts for Quadro and Tesla versions of the GPU would dictate that few would see the light of day as gaming cards, and I really wouldn't be surprised at a $700+ pricetag when they arrive.
I still haven't seen any actual convincing argument that GK110 was ever intended to be released as a 600 series card. From all accounts, the initial 28nm offerings were going to follow AMD/ATi's recent model (i.e. mid-sized die with the top SKU being a dual card**) with the GK110 to follow at a later date -probably as the GTX 780 to combat the HD 89xx cards.

EDIT: ** Dual-card GTX 690 in May according to 3DCenter, with GK110 somewhat later-indicating that it would be closer to being a 700 series than 600.

 

[emmzo]

Single 6 series card vs 3 580s? We`ve seen this kind of advertising over and over again. Even if such tech is available, marketing dictates small leaps for maximizing profits. The 680 will only be 10 to 15% faster than the 7970, so probably not even 40% over the 580 and the price will also get bigger since AMD has a stable market and won`t cut prices anymore.

 

[indiangamer]

it looks good but not that much that it requires 3 gtx 580 to run. this shows how weak our software development is.consoles are 10 times weaker than current generation gpus but they don't look 10 times bad.
and epic only shows demo i think the hires some 100s of software engineers to work for years and make a 2:46 min long tech demo. see every game using unreal engine 3, they looks fool.

 

[Guest]

Wrong. [link] .

?

That link itself showed that GK110 was a real replacement for GTX580 while GK104 was always intended to be a replacement for GTX570.

I don't know what you are arguing about. Whether NVidia launches GK110 as a GTX780 is irrelevant.

It's like Nvidia launching a GTX460 first and skipping GTX480 entirely due to issue and re-releasing it as a GTX580 --> aka 780.

GK104 is not the real high-end Kepler chip. It might only be the highest end single-GPU for 600 series, but that's just marketing and naming.

Also, you are dead wrong that previous NV's high-end generations didn't bring a 50-100% performance increase. They all did. GK104 will bring the least performance increase in that regard.

http://alienbabeltech.com/abt//viewtopic.php?t=21797

Comparing memory bandwidth across different brands just shows you are either not knowledgeable or trying to skew the argument by ignoring the point. You should ONLY compare memory bandwidth across the same brand. I never said anything about AMD vs. Nvidia's relation of memory bandwidth. In Nvidia's case it did increase memory bandwidth every time it released a new high-end GPU. Considering Kepler GK104 doesn't move that bar at all, it's upper mid-range. Nvidia is just launching it as high-end because they COULD NOT get out the real flagship card. And since GK104 is good enough to compete with HD7970, it's their "flagship" holdover card.

If you don't want to believe that, it's your choice. But that's a fact. Just speak to any Nvidia employee. *wink*

 

[EXCellR8]

Def a pretty badass tech demo...

 

[Adhmuz]

Guest said:
If you don't want to believe that, it's your choice. But that's a fact. Just speak to any Nvidia employee. *wink*

Because we know what Nvidia employees tell us is 100% fact... Don't be so naive.

Guest said:
Comparing memory bandwidth across different brands just shows you are either not knowledgeable or trying to skew the argument by ignoring the point. You should ONLY compare memory bandwidth across the same brand.

The point was more to illustrate that memory bandwidth doesn't determine the performance of a GPU, ever consider that the need for more bandwidth doesn't actually exist? At least not with current GPU technology, ie the GPU is now the bottleneck and not the memory.

As far as the video, the original running off the 580s looks better imo. This one looks muddy and not as sharp.

 

[dividebyzero]

Also, you are dead wrong that previous NV's high-end generations didn't bring a 50-100% performance increase. They all did. GK104 will bring the least performance increase in that regard.

GTX 480 + 39% over GTX 285
GTX 280 + 39% over 9800GTX
9800GTX -2% vs. 8800GTX

...Yeah, right, "they all did"......0 fer 1

Comparing memory bandwidth across different brands just shows you are either not knowledgeable or trying to skew the argument by ignoring the point.You should ONLY compare memory bandwidth across the same brand

So, you somehow think that comparing Nvidia-to-Nvidia will somehow validate your Walter Mitty world view:

GTX 280 : 512-bit / 8 x 2214 MHz effective = 141.7 GB/sec
GTX 560Ti : 256-bit / 8 x 4008 MHz effective = 128.27 GB/sec

Oops.....0 fer 2

I think I'll stop there, since the rest of your posting really isn't worth the effort, and provides no more than a textbook example of why I don't usually bother even reading the posts of the bulk of Guest posters.

 

[St1ckM4n]

...

Just speak to any Nvidia employee. *wink*

:O


Also I agree with someone above - this junk shouldn't max out even a single 580. I don't know if I can't see it through the YouTube, but is the entire focus on AA and texture filtering, along with some other filters/effects..?

Also, for consoles to look 10x worse, you really need to be looking for this stuff. From a couch position to the casual gamer, consoles are mind blowing.

 

[dividebyzero]

The fact that NV's GK106/107 are also no show just clearly goes to show NV isn't ready to launch Kepler.

Straight from Guest posting accuracy contest:

Stay tuned, we'll have a quick preview of Kepler GPU performance on the mobile site coming later tonight!!

Ryan Shrout
Review/preview here

 

[Guest]

I laugh at the silly kids who spend a fortune on computer hardware...so quickly outdated lol