Radeon HD 7000 Revealed: AMD to Mix GCN with VLIW4 & VLIW5 Architectures

[Dimitriman]

NEW Launch date is Dec 22nd!


http://www.techpowerup.com/156876/AM...cember-22.html

December 22? Wow... AMD was really sneaky with this card... We will not even make 20 pages by launch date!

[SubZero.it]

yes, please. let him sink his own boat.

Stevil can you ban him from the world?

Going back to the topic. To me it seems a really interesting and innovative card such the 5870 was, that's the reason why I'll buy it. Considering also that the GTX 6xx is three quarters away (I think)

[Olivon]

Considering also that the GTX 6xx is three quarters away (I think)

Are you sure ?

[minpayne]

chiphell leak

A few numbers:
1. NDA ends on December 22
2. Beats GTX580 by 20% on average, not 3DMark. Beats GTX580 by around 30% in Crysis. Expected goal has been reached. There has been no goal of doubling the performance of previous generation cards as speculated.
3. Two cards will be (paper) launched: PRO + XT

[Vardant]

Are you sure ?

I'm 99% sure he's wrong.

[Picao84]

A few numbers:
1. NDA ends on December 22
2. Beats GTX580 by 20% on average, not 3DMark. Beats GTX580 by around 30% in Crysis. Expected goal has been reached. There has been no goal of doubling the performance of previous generation cards as speculated.
3. Two cards will be (paper) launched: PRO + XT

Number 2 fits my earlier expectations and reasoning:

It wouldnt surprise me. We have to keep in mind the change in architecture, with hardware schedulers, instead of a software one on VLIW. That has to account for higher sized die (compared to one in VLIW) as well as higher power consumption. However, if the upgrade in performance is little, I will also say they didnt went overboard with die size. They are following the carefull strategy of not doing big chips. To keep the die size low, they traded a big increase in performance for the scheduller, not risking doing both this generation.

[Picao84]

I'm 99% sure he's wrong.

Can that be because he said GTX 6xx? :P
A GTX 6xx doesnt really need to be GK100 :P

[spursindonesia]

Stevil can you ban him from the world?

Going back to the topic. To me it seems a really interesting and innovative card such the 5870 was, that's the reason why I'll buy it. Considering also that the GTX 6xx is three quarters away (I think)

Are you sure ?

Take it with a grain of salt, and this is only unconfirmed information from one of my friend working in one of the AIB distro in my country -as a distribution manager, or something like that. GK 104, NVidia's answer to Tahiti, will be here no later than end of Q2 2012, but he didn't expect it arriving before late March or early April next year too. He's quite confident talking about it, and his info has been quite accurate AFAIK.

Then, if we're talking about GK 110, things get much more blur, since this will be one big monster chip, we're gonna see it when NVidia is confident enough with TSMC's 28 nm process. He wasn't all that sure, but his boss has said to expect this monstrosity at early Q4 2012 at the earliest (when everything went totally perfect), and can easily slip to 2013.

No solid info regarding peformance, but he expects GK 104 to be trading blow with Tahiti -and he's quite even handed to both IHVs.

Ofcourse, perhaps all of that are just silly misinformations, can't be assure of anything.

[BoredByLife]

If Nvidia comes that late with their monster chip, AMD will have an answer ready.

[DilTech]

Stevil can you ban him from the world?

Going back to the topic. To me it seems a really interesting and innovative card such the 5870 was, that's the reason why I'll buy it. Considering also that the GTX 6xx is three quarters away (I think)

...try 2 months tops. Where are people getting this 2-3 quarters stuff from?

It's due IN 1Q 12, NOT end of the year.

Also, the 5870 wasn't really innovative. It was, for all intents and purposes, a 4890 with DX11 tacked on and a higher shader count. That's what makes the 7970 interesting, because it's the first time AMD have REALLY stepped away from the architecture they used in the R600.

Take it with a grain of salt, and this is only unconfirmed information from one of my friend working in one of the AIB distro in my country -as a distribution manager, or something like that. GK 104, NVidia's answer to Tahiti, will be here no later than end of Q2 2012, but he didn't expect it arriving before late March or early April next year too. He's quite confident talking about it, and his info has been quite accurate AFAIK.

Then, if we're talking about GK 110, things get much more blur, since this will be one big monster chip, we're gonna see it when NVidia is confident enough with TSMC's 28 nm process. He wasn't all that sure, but his boss has said to expect this monstrosity at early Q4 2012 at the earliest (when everything went totally perfect), and can easily slip to 2013.

No solid info regarding peformance, but he expects GK 104 to be trading blow with Tahiti -and he's quite even handed to both IHVs.

Ofcourse, perhaps all of that are just silly misinformations, can't be assure of anything.

GK104 is the midrange part. GK100 is the monster part, and GK110 is likely the refresher part later in the year.

If Nvidia comes that late with their monster chip, AMD will have an answer ready.

Supposed leaked slides show the GK100 as around double the GTX580 (sometimes higher, sometimes lower). That'd put it at over 60% faster than the 7970. It's suppose to be here in Q1 2012. NVidia's midrange GK104 should be about 75-80% of the GK100 in terms of overall power, that's a scary situation.

AMD's answer to NVidia's high end has generally been a dual-gpu card. Anyone heard anything about a 7990 coming? Last few cards we heard about said card while we were hearing about the rest on the way, but this time all is quiet on the front. Is this chip too big to pull off that set-up?

[Dimitriman]

Supposed leaked slides show the GK100 as around double the GTX580 (sometimes higher, sometimes lower)...

I have never seen in history a release of a card that was more than 2x faster than the previous high end equivalent.

Just sayin...

[DilTech]

I have never seen in history a release of a card that was more than 2x faster than the previous high end equivalent.

Just sayin...

...

6800Ultra was 2x 5950 Ultra (higher than that in DX9)
8800GTX was OVER 2x 7900GTX (in some cases 2.5x, on the release drivers no less. By the end of it's life it was more like 3x)

You've seen it, you just needed a friendly reminder. Personally, I'm not listening to the slides from either company, but the specs match the percentages the slides are claiming. 1024 nvidia shaders (which are also reworked for kepler) with a 512bit memory bus is definitely enough to double the performance of the GTX580.

Either way, lets get back to the topic at hand here. AMD's 7xxx series.

[radier]

http://www.obr-hardware.com/2011/12/...his-enjoy.html

[Olivon]

I have never seen in history a release of a card that was more than 2x faster than the previous high end equivalent.

Just sayin...

8800 GTX was really something compared with 7900 GTX :



http://www.hardware.fr/articles/644-...formances.html

[Lanek]

[url] ]

You know you are not allowed to post things from OBR in this forum ? not cause it is OBR; but cause he's is banned, and this is a rules for all people who have been banned..

What does this slide tell us more of what we know allready ? outside the gcn he have copy in each CU ?

[Dimitriman]

...

6800Ultra was 2x 5950 Ultra (higher than that in DX9)
8800GTX was OVER 2x 7900GTX (in some cases 2.5x, on the release drivers no less. By the end of it's life it was more like 3x)

You've seen it, you just needed a friendly reminder. Personally, I'm not listening to the slides from either company, but the specs match the percentages the slides are claiming. 1024 nvidia shaders (which are also reworked for kepler) with a 512bit memory bus is definitely enough to double the performance of the GTX580.

Either way, lets get back to the topic at hand here. AMD's 7xxx series.

8800 GTX was really something compared with 7900 GTX :



http://www.hardware.fr/articles/644-...formances.html

But you are certain that GTX 680 or GTX 780 whatever it is called, will be over 2x faster than 580 at launch?. And btw 6800Ultra & 8800GTX were close to 2x faster at launch but they didn't quite make it that far. Please find evidence here and here. And please don't link things which shows it faster than 2x in under extreme circumstances and ideal scenarios. I'm talking in overalls here.

With each generation consuming more power and requiring more cooling, you can see why obtaining this yet again will be very hard with 28nm. Not to mention the normal diminishing returns, ie. even if you exactly double the GTX580 in every aspect, you will not get linear 100% improvement.

If you believe that then you are just buying into the hype too early...

[AKM]

AFAIR since the HD 4XXX gen at least, the fastest single GPU Radeon has been as fast as Nvidia's second fastest single GPU card, I bet we'll see the same thing with this generation.

[DilTech]

But you are certain that GTX 680 or GTX 780 whatever it is called, will be over 2x faster than 580 at launch?. And btw 6800Ultra & 8800GTX were close to 2x faster at launch but they didn't quite make it that far. Please find evidence here and here. And please don't link things which shows it faster than 2x in under extreme circumstances and ideal scenarios. I'm talking in overalls here.

With each generation consuming more power and requiring more cooling, you can see why obtaining this yet again will be very hard with 28nm. Not to mention the normal diminishing returns, ie. even if you exactly double the GTX580 in every aspect, you will not get linear 100% improvement.

If you believe that then you are just buying into the hype too early...

http://vr-zone.com/articles/nvidia-g...ew/4216-1.html

Keep in mind, that's comparing it with the 7950 gx2 (the dual-gpu card). Now look at those AA results! Yeah...2x was childs play once you started cranking up the AA.

Anyway, no, gains aren't linear. Kepler is a new design though, it's not just Fermi with more shaders tacked on, much like how the 7970 isn't just a 6970 with more shaders tacked on either. Also, another thing to remember is that the GTX 580 was memory starved at the top-end of the spectrum (hence why the 6970 was able to over-throw it at ultra high rez with AA), while the GK100 should fix said problem (meaning, at 2560x1600 double should definitely be doable, which is what the slides claim).

This is why I keep saying I can't wait to see this battle go down. It'll be the first time we've seen a show-down between TWO new architectures since the G80 vs R600. Since then, AMD have mainly been tacking things on to their R600 design and trying to clean up it's flaws, so I'm actually extremely intrigued to see how this new card will fare against it's competition.

[Olivon]

A single GeForce 8800 GTX is more powerful overall than a 7900 GTX SLI configuration and even NVIDIA's mammoth Quad SLI. Although it's no longer a surprise to see a new generation of GPU outperform the previous generation in SLI, the sheer performance we're able to attain because of G80 is still breathtaking.

Anandtech 8800 GTX Review

[Dimitriman]

Anandtech 8800 GTX Review

:rolleyes:

Its about 2x faster.. its not a hallmark faster than that.. Yet the GK100 slide is claiming an average of 2.5x faster than GTX580, even with twice the hardware, you would not achieve this, go ahead and check your GTX580 SLI numbers and prove me wrong. Besides, even if GK100 is 2x faster than GTX580, we still need confirmation that it will also come in dual chip flavor, because its direct competition will not be the 7970 right?

Anyway, one can believe whatever he wants, we should not go off topic any longer, I think 40% over Caymen will be acceptable to most people provided that:
a) price is competitive
b) they have the option to purchase a 7990 dual chip card in case they need more raw power in the same slot.

[spursindonesia]

AFAIR since the HD 4XXX gen at least, the fastest single GPU Radeon has been as fast as Nvidia's second fastest single GPU card, I bet we'll see the same thing with this generation.

I second this opinion.

[Pantsu]

2x increases are mostly limited to specific scenarios where newer GPUs excel, like AA, tessellation etc. On avarage the increases have been far more modest in the last few generations (after 8800GTX and 4870). That's why I expect a max 50% performance increase for next gen Radeon/GeForce on avarage. There might be some cases that produce 2x increase of course, but that doesn't mean that gen is 2x faster overall. 5870 more than doubled everything and resulted in ~70% performance increase on avarage. Based on the leaked specs 7970 is nowhere near double 6970, so even 50% increase would be impressive. The same goes for GK104. Of course architecture brings its own little spice. For AMD it probably results in a disproportionate performance increase in tessellation, perhaps even 2x in some cases, but otherwise perhaps lackluster general performance due to immature drivers. Nvidia has a more mature architecture so it'll probably have a more even performance increase.

[Zoran]

...in about 50% of games, in the other 50% it is almost as fast as the fastest nvidia

[DilTech]

2x increases are mostly limited to specific scenarios where newer GPUs excel, like AA, tessellation etc. On avarage the increases have been far more modest in the last few generations (after 8800GTX and 4870). That's why I expect a max 50% performance increase for next gen Radeon/GeForce on avarage. There might be some cases that produce 2x increase of course, but that doesn't mean that gen is 2x faster overall. 5870 more than doubled everything and resulted in ~70% performance increase on avarage. Based on the leaked specs 7970 is nowhere near double 6970, so even 50% increase would be impressive. The same goes for GK104. Of course architecture brings its own little spice. For AMD it probably results in a disproportionate performance increase in tessellation, perhaps even 2x in some cases, but otherwise perhaps lackluster general performance due to immature drivers. Nvidia has a more mature architecture so it'll probably have a more even performance increase.

GK104 is the mid-range part. If the rumors hold true, it's a 768 shader 384-bit monster of a midrange part that should easily fly in comparison to even the GTX580. GK100 is the high-end, and it's supposedly 1024 shaders and a 512bit bus.

It's going to end up a clash of the titans for sure.

[paulbagz]

OBR Link

Just a friendly warning but you really need to stop posting this stuff now for the safety of your account =/

-PB