[3dcenter] GK104 specs

[Silverforce11]

Also Kepler != Kepler. GK110 might be much more HPC oriented than GK104. Everything from GK104 down seems to have been engineered for maximum efficiency in perf/mm and perf/W. Again, why make a successor only 30-40% faster? It is impossible that GK104 is only on par or below GTX580. That would be to little improvement for a tweaked architecture and a node jump.

You mentioned the focus on perf/mm and perf/w.. if they made gk104 to be excellent perf/w, they may not target it for outright performance. Thus, gtx580 performance but at half or less the power use is a major winning design and expected of a full node shrink.

Next gen mid-range offers previous gen high-end at significantly less power use, why is that bad?

Edit: gk104 is not the successor, gk110 will be the true heir.

 

[boxleitnerb]

GK104 is the successor to GF114 as the codename implies.
You have to compare GK104 with GF114. If it only scratches at the 580 and has a similar TDP as a GTX560(Ti) that would put the perf/W only at +30-40%.

It's not bad, it's just a too small step in this case since GF114 itself is quite powerful. GK104 has more than twice the texture fillrate, more pixel fillrate, almost twice the compute power as a 580. And with all this it should at best equal this card? Simply no, even at reduced efficiency per unit. We're not talking GTX580+50% here. 15-20% are not unreasonable, though.

 

[Crap Daddy]

GK104 is the successor to GF114 as the codename implies.

It's not bad, it's just a too small step in this case since GF114 itself is quite powerful. GK104 has more than twice the texture fillrate, more pixel fillrate, almost twice the compute power as a 580. And with all this it should at best equal this card? Simply no, even at reduced efficiency per unit. We're not talking GTX580+50% here. 15-20% are not unreasonable, though.

Right. If we'll found out that NV will name this GTX680, which more and more rumors imply, then the performance will be better than that of the 7970. I don't think NV can afford to launch a GTX*80 that's slower than AMD's top dog.

 

[Silverforce11]

Don't forget Cayman had more compute power than Gtx580 if you measure SP. DP they are close.

Looking at gk104, it looks very similar to Cayman. More shaders but less complex, takes less transistors, so they can pack more of it on. Thus, going from gf114 -> gk104 you get a smaller die but nearly 4 times the shaders. The assumption im making here is these shaders are not up to scratch per clock vs fermi, nowhere near it, just like VLIW4 wasn't thus it had heaps more but didnt beat gtx580. I could well be wrong, but its not as stary-eyed an estimate as some of you expecting >>gtx580 performance.

ps. i would love it if gk104 was super fast, really, competition is always good as these 7950s are way too expensive.

 

[BallaTheFeared]

In short you're comparing AMD graphics cards per/nm to graphics cards that were much more versatile and did far more than their AMD counterparts.

Compute power doesn't mean anything in that case because AMD had no real programming language to use their cards in a real GPGPU compute program. AMD has been designing for half the market, 7xxx changed that but still has no support afaik.

 

[boxleitnerb]

Right. If we'll found out that NV will name this GTX680, which more and more rumors imply, then the performance will be better than that of the 7970. I don't think NV can afford to launch a GTX*80 that's slower than AMD's top dog.

AMD changed their naming scheme, why can't Nvidia do the same? I don't think GK104 will beat Tahiti XT overall. Maybe in some cherry picked cases and in general it may come close but beating it is unlikely in my opinion. BigK is not ready/not planned for now, so what do you do? At least make the name sound highend, cherry pick benchmarks so that you barely beat the 7970 by a hair's breadth. But this could backfire because then people would think that Kepler is only 30% faster than Fermi and forget that BigK is still coming. It will be interesting how marketing spins this one.

 

[Silverforce11]

In short you're comparing AMD graphics cards per/nm to graphics cards that were much more versatile and did far more than their AMD counterparts.

Im not arguing the reasons for the observed facts.

Merely stating the previous trend whilst comparing gk104s perf based on leaked specs vs NVs previous gen. I just don't think its going to be faster than gtx580 given its specs and die size. Would be great if it was and i'll be there to buy it if its cheaper than 7950.

 

[BallaTheFeared]

Word play.

I'd say agree to disagree in so much as the fact that while Nvidia needed more nm to beat AMD, their cards weren't one trick pony's that were worthless in the most lucrative markets.

This rumored card has nothing in common with gk104, nothing. It's a total redesign, stop using false logic based on incomparable information.

 

[Silverforce11]

Word play.

I'd say agree to disagree in so much as the fact that while Nvidia needed more nm to beat AMD, their cards weren't one trick pony's that were worthless in the most lucrative markets.

This rumored card has nothing in common with gk104, nothing. It's a total redesign, stop using false logic based on incomparable information.

So whats wrong with my logic compared to those who expect >>gtx580 performance?
Or rather.. could you explain the logic behind that IMO unrealistic expectation.

On paper it looks like Cayman, very high SP output, 1/4th DP. 256 bit and low bandwidth >>gtx580??

 

[boxleitnerb]

Nvidia still has a superior frontend, so it should be able to harness more performance from a fixed amount of resources than AMD. And it has higher clocks than Cayman. Cayman comes quite close to the 580, sometimes even beating it (SSAA, high res). So beating a 580 is not out of the question, it's rather a given.

But you still cannot compare Cayman and Fermi/Kepler. The devil is in the details.

 

[Silverforce11]

A few leaks now have it in early March for a review launch. Not too long to go.

 

[BallaTheFeared]

So whats wrong with my logic compared to those who expect >>gtx580 performance?
Or rather.. could you explain the logic behind that IMO unrealistic expectation.

On paper it looks like Cayman, very high SP output, 1/4th DP. 256 bit and low bandwidth >>gtx580??

People are just speculating, nobody has enough information to say more or less really other than what they personally expect.

Calling it unrealistic puts you in the same boat as those people having those expectations, no better, no worse.

Again you're basing your information off information you simply cannot use.

You have no idea how the ROPs or TMUs perform, nor the shader cores compared to Cayman, you see similar specs and automatically assume based on you know about those cards when those cards have nothing to do with this card.

 

[Crap Daddy]

AMD changed their naming scheme, why can't Nvidia do the same? I don't think GK104 will beat Tahiti XT overall. Maybe in some cherry picked cases and in general it may come close but beating it is unlikely in my opinion. BigK is not ready/not planned for now, so what do you do? At least make the name sound highend, cherry pick benchmarks so that you barely beat the 7970 by a hair's breadth. But this could backfire because then people would think that Kepler is only 30% faster than Fermi and forget that BigK is still coming. It will be interesting how marketing spins this one.

Let's think different. These specs translated to Fermi as in a die-shrink only, would be 768 CUDA cores, that's 50% more than the GTX580. Based on Fermi CUDA core count and performance how faster do you think this chip will be compared to a GTX580?

A hint: GTX 580 has 75% more CUDA cores than the 560Ti and has 35% better performance on average. Considering the same abstract comparison a chip that has 50% more CUDA cores than the 580 will gave an average performance increase of 25%. Stock per stock clocks the 7970 is up to 20% faster than the GTX580 across the board, some reviews place this difference even lower. So waht do you think?

 

[BallaTheFeared]

Cuda cores aren't the only thing that matters here, GF110 had double pumped shaders, so they ran at 1544Mhz, which is faster than these will "reportedly" run.

You also need to consider Poly Engines, ROPs, and TMU's and the performance per unit.

 

[Crap Daddy]

Cuda cores aren't the only thing that matters here, GF110 had double pumped shaders, so they ran at 1544Mhz, which is faster than these will "reportedly" run.

You also need to consider Poly Engines, ROPs, and TMU's and the performance per unit.

Of course. GK104 has the same ROP and TMU count as Tahiti XP.

 

[boxleitnerb]

Let's think different. These specs translated to Fermi as in a die-shrink only, would be 768 CUDA cores, that's 50% more than the GTX580. Based on Fermi CUDA core count and performance how faster do you think this chip will be compared to a GTX580?

A hint: GTX 580 has 75% more CUDA cores than the 560Ti and has 35% better performance on average. Considering the same abstract comparison a chip that has 50% more CUDA cores than the 580 will gave an average performance increase of 25%. Stock per stock clocks the 7970 is up to 20% faster than the GTX580 across the board, some reviews place this difference even lower. So waht do you think?

I think GK104 will be held back by its bandwidth. Also, the 20% for the 7970 are to conservative for my taste. Almost all the reviews I have read put this card 30% above the 580. Drivers could also add 5-10%, in some cases more. For one card to beat another, I expect a lead of at least 10% across the board. I doubt that will happen with GK104, especially at high res settings.

 

[Crap Daddy]

I think GK104 will be held back by its bandwidth. Also, the 20% for the 7970 are to conservative for my taste. Almost all the reviews I have read put this card 30% above the 580. Drivers could also add 5-10%, in some cases more. For one card to beat another, I expect a lead of at least 10% across the board. I doubt that will happen with GK104, especially at high res settings.

I do agree that with drivers (when they will come out...) and overclocking the 7970 is much faster than a 580 but everybody references official reviews and stock clock cards which place the 7970 within range of the supposed GK104 performance. So what I'm saying is that it's enough to be on 5% average across the board better and claim they have a winner. Depending on the final pricing one can determine if the 7970 is a better buy.

 

[BallaTheFeared]

Quick test... Crysis 2

900 core 1800 shader, on the core/shader side this is about equal to a 580 stock, maybe slightly faster.

First test is 104 GB/s

Second is 172 GB/s

Third is 160 GB/s

172 GB/s increased performance over 104 GB/s by 16.5%, however over 160 GB/s the increase was only 1.9%. My guess is memory bandwidth produces only negligible benefits past a certain point, until it gets to the point where it no longer produces any benefit at all. More than likely for GTX 580 performance the difference between 160 GB/s and 250 GB/s is only going to be a few percent.

I'd also wager increasing core speed to 950/1900 and keeping 160 GB/s would increase fps more so than 900/1800/172 GB/s did over the same clocks with the lower bandwidth.

 

[Lonyo]

Memory speed is usually diminishing returns.
You need to have *enough* memory bandwidth to feed the GPU core, too little and you will probably suffer a lot, ESPECIALLY in edge cases like high resolutions with AA/etc, but for most situations it would be fine.

Any chance of testing with settings that might push the RAM some more, like extra AA etc? And maybe some lower settings which will push it less (e.g. lower res and no AA).

It's been shown in previous generations that AMD closes the gap as you increase resolutions, with NV having their biggest lead at 1280x1024 and the smallest at 2560x1600 (assuming you keep AA/etc the same).

It may be that NV are tailoring this card to specific resolutions, e.g. 1920x1080 (if you look at past slides, they have indicated what sort of resolution/settings cards are for), and they have decided that they don't need masses of memory bandwidth to give the core enough at their target resolutions.
It may be that the GTX680 will perform noticeably worse than the 7900 series at 2560x????, but be close enough at 1920x1080 for example. All depends on how they have decided to tweak the card and what they want from it.

 

[BallaTheFeared]

4xAA is as high as the bench allows, the game has no AA setting only Post Processing which I believe is FXAA.

Toss me a couple settings/clocks you'd like to see and it shall be done.

 

[boxleitnerb]

I just tried to bench the impact of memory bandwidth as well, but alas, I get funny colored artifacts when I lower the memory clock. Weird.

@Balla: Do you have time to run some more tests? What kind of games do you have? Would be great if you could bench 2-3 games with SLI active, default GTX470 clocks and a memory bandwidth of 160, 240 and 320GB/s (note: SLI doubles memory bandwidth, so 80, 120 and 160GB per card).

 

[tviceman]

A few leaks now have it in early March for a review launch. Not too long to go.

We shall soon see then if you're

You guys who think a die smaller than Tahiti while being faster than it is dreaming.
At 342mm2, its going to be around gtx580 AT BEST.

My bet is gk104 with these die size and specs will struggle to match a gtx580.

The die size/perf gap is big, now you suddenly expect NV to turn it all around and have their 342mm2 gk104 faster than 7970? If that isn't dreaming..

Predictions are at all accurate.

 

[BallaTheFeared]

(note: SLI doubles memory bandwidth, so 80, 120 and 160GB per card).

Uh, are you sure about that? Can anyone confirm/deny that? I figured since the memory doesn't double, neither does the actual bandwidth.

My ssd and hdd died, I'm using a laptop drive and I didn't install many games on it...

I've actually been dinking with 51Mhz core low power state gaming with League of Legends.

My gaming temps are the same as my idle temps, lol and that's with 4/5 fans off.

 

[boxleitnerb]

Uh, are you sure about that? Can anyone confirm/deny that? I figured since the memory doesn't double, neither does the actual bandwidth.

Pretty sure. If you look at the specs of a dual-card, all things are doubled, also the memory interface. I know that the interface isn't actually twice as wide but as both cards work simultaneously and access the same data in their respective VRAM, the outcome is the same.

http://www.bit-tech.net/hardware/graphics/2011/03/24/nvidia-geforce-gtx-590-3gb-review/2

 

[BallaTheFeared]

Yeah but that's just marketing, the 590 is marketed as a 3GB card but it still only has 1.5GB usable same as a 580, and 580 SLI.