ATI's next gen GPU in next half of 2010

[wahdangun]

Also increase heat and power useage.

not really look at HD 5870 its uses less power than HD 4870 but have twice computing power,


that's why people keep saying fermi fail because it's use more power (double it) although it's still have double computing power.

 

[a_ump]

maybe double computing power but the actual performance increase from GTX 285 to GTX 480 is less than the increase from the HD 4890 to HD 5870. The power decreased with the HD 5870 compared to the HD 4890, the power increased for the GTX 480. It just shows that ATI went about their chip in a much more efficient way. Nvidia will i feel have to make quite a few changes to fermi within the next 2 gens in order for it to continue building on it as even die shrinks aren't going to dramatically change power draw and heat output.

Shame there hasn't been anymore leaked info on the HD 6XXX series. blah be so sweet to get some specs. like say:
2400SPU's
48ROP's
256-bit w/ 7ghz GDDR5
950mhz core

and i also remember reading something about the HD 5870 has the same like distributor for its SPU's as the HD 4870 and that it's due for an update. So if it got better one of those or maybe 2 of em that would be a bomb revision.

 

[ucanmandaa]

here is my theory on new cards. We all know Ati likes to test new production technologies with low mid range cards before releasing any high end cards with it. Like 9600pro (130nm), (1950pro 80nm* dongleless crossfire) and last of this trend was 4770(40nm).
We also know Ati haven't produce a single gpu highest end card since 1950xtx. They leave it to either crossfire cards or multi gpu single cards (3870x2,4870x2 and 5970)
This approach proved very simple yet highly profitable as far as i can see. Downside of this is the crossfire being dependent on profiles and even when a game has a crossfire profile then it may have scalablity issues.
Now comes my theory... seeing the code names Southern Island(s) and Northern Island(s) what if these cards are dual-core cards. Two gpus duct taped together.
Southern Islands >>> two 5670 gpus built on 28nm with shared memory and profile independent
Northern Islands >>> two 5870 gpus again on 28nm, shared memory and profile independent.

What do you think? Is something like this possible? also being profile independent should mean we always get x2 peformance right?

*i am might be wrong on that.

 

[ToTTenTranz]

I don't really think the next GPU will have many more shaders.

Shader processing is already oversized in Cypress.

What they do need is to improve the tesselation processing (i.e. adding more tesselation units), increase the nemory bandwidth and improve the schedulers (maybe increase the scheduler/shader-unit ratio).

 

[Super XP]

I don't really think the next GPU will have many more shaders.

Shader processing is already oversized in Cypress.

What they do need is to improve the tesselation processing (i.e. adding more tesselation units), increase the nemory bandwidth and improve the schedulers (maybe increase the scheduler/shader-unit ratio).

I am sure Northern Islands is going to be based on something we've never seen before and perform like a piece of raw meat strapped onto a running guys bottom being chased by a hungery dog.

 

[ToTTenTranz]

I am sure Northern Islands is going to be based on something we've never seen before and perform like a piece of raw meat strapped onto a running guys bottom being chased by a hungery dog.

Erm.. was that a praise to the Northern Islands GPUs?

 

[a_ump]

hmmm it would be cool to see HD 7870(NI) with say 156 SPU's with crazy performance lol

 

[Super XP]

Erm.. was that a praise to the Northern Islands GPUs?

Yes in a twisted point of view.

 

[_33]

Well, here's a sign that Northern Islands is going to happen: http://www.dvhardware.net/article42311.html

 

[erocker]

Already read about it here http://forums.techpowerup.com/showthread.php?t=120504

 

[afw]

Ok found this ....

ATI IS ON A ROLL, having taped out its next generation GPU family called Southern Islands. ATI might call it the HD6xxx series, and it could be out before Nvidia gets its GTX4xx line of GPUs fully fleshed out and to market.

The news is straightforward enough. Southern Islands (SI) taped out recently, and is now moving through TSMC. Although the schedule is very tight, if all goes really well we could see a demo or two at Computex in just over a month. This puts SI slightly behind where Evergreen was a year ago, but not by much.

Southern Islands is said to be a hybrid between the all new Northern Islands (NI) and the current HD5xxx Evergreen family. Those in the loop say that Northern Islands was meant to be on TSMC 32nm before that process was killed. Due to the ever-slipping nature of TSMC's 28nm high-K metal gate node, the SI hybrid GPU was slapped together to be fabbed on TSMC's current 40nm process.

Sources tell SemiAccurate that SI uses some of the NI uncore (unshader?), and wraps that around a mildly updated Evergreen shader. RAM is more of an open question. NI was set to use GDDR5+, but since DRAM makers might not be ready, we may end up with only GDDR5 on SI.

In any case, Nvidia has come out fighting with its GTX470 and GTX480, shipping hundreds, some say thousands, of units since ATI shipped Evergreen last September. If all goes well, and it appears to be doing just that, ATI might have Southern Islands on sale before Nvidia gets all of the GTX4xx variants out the door. The end of 2010 will see Nvidia fighting a new ATI line with a year late part that doesn't work within the promised specs, and can't be manufactured.

By the time Nvidia gets to 28nm next year, or possibly in 'GTX480' quantities this year, it is going to be fighting yet another new generation of ATI parts, Northern Islands. If the rumors are true, and the upcoming Fermi II is on TSMC's 28nm bulk process, and NI comes out on GloFo's 28nm high-K metal gate process, Nvidia won't even be in the game.

At this point, you really have to ask yourself if Nvidia will survive. Its chipset business is gone because it didn't understand Moore's law. Its GPU business is like a floating goldfish circling the drain. Its GPU compute business is wounded. And Tegra is being laughed out the door. Meanwhile, ATI is on a roll.S|A

mildly updated shaders ... and more tesselators ... and we're going to see it by september ... hmmm

WTF is happenning @ ATI ... cant they have a break till NV catches up ... if this is true ... NV must be surely feelin the heat .... :shadedshu

 

[Fourstaff]

WTF is happenning @ ATI ... cant they have a break till NV catches up

That, my friend, was what Nvidia did after releasing G80. Look at their position now.

 

[afw]

hmmm ... i guess ur right ... but were'nt they (NV) busy ... doing all those re-branding and die shrinks ... ...

 

[T949]

GO ATi, they are raping the competition lol

You do know thats bad for the gpu market. More competition = lower prices. Fanboism aside I hope both ati and nvidia have similar preforming products so their can be a price war.

 

[Nick89]

I hope nvidia can keep up other wise we will be paying an arm and a leg for an HD6800 series.

 

[phanbuey]

I hope nvidia can keep up other wise we will be paying an arm and a leg for an HD6800 series.

or we will have a really cheap gtx 470

 

[D007]

I finally believe the time has come for me to try ATI.
Good work ATI, keep it up.

 

[the54thvoid]

You do know thats bad for the gpu market. More competition = lower prices. Fanboism aside I hope both ati and nvidia have similar preforming products so their can be a price war.

This is sad but true. Look at how ATI have kept their prices as they are following Fermi's less than golden reception. If only Fermi had performed as it does with better thermals/power then ATI would have probably reduced price.

No, it would be very bad for ATI to run away with the performance crown. I love my radeons but i enjoyed buying them at 75% the price they go for now.....

Ying needs Yang just as ATI needs NV.

 

[ToTTenTranz]

I hope nvidia can keep up other wise we will be paying an arm and a leg for an HD6800 series.

I wish for Intel to launch high-performing Larrabee derivatives and for S3 and PowerVR to rise from the ashes with truly competitive mid-end solutions.


nVidia, on the other hand, has been hurting its customers too much with all the renaming and trying to force closed standards. I hope they burn down until they get different management and marketing teams.

 

[cadaveca]

I hope they burn down until they get different management and marketing teams.

Want to hear something interesting about nVidia marketing?

A couple of years ago(July 2007, to be exact), Brian Del Rizzo, Nvidia’s Senior PR Manager of Consumer Products, took me out for dinner, and asked me what I thought of how nV was doing, where they were headed, and what I thought they could improve, to make the customer experience better.

Sounds screwed up, doesn't it. One the top nV guys asking me what I thought...

T'was my brithday, too, but I dunno that he knew.

I wonder what he'd say to me now...if he even remembers, and how he thinks things would be different, if they had done what I said.


I knew they were in trouble then...he couldn't even look me in the eye when I told him I thought they were doing good. He had the same look that Hector had, 2 years before the merger.

 

[_33]

But if you look at the Fermi chip and take into perspective the technology that's been put into it, Southern Island is basically like the Fermi that Nvidia is trying to mass produce. Only, Nvidia is at the marketing stage, where ATI is at the tape out stage. So in essence Nvidia is about 4 to 6 months in advance. I don't know if anyone is following me on this. Nvidia has a technologically more advanced chip still than ATI's 5xxx generation, but of course it's a big mess with the 40nm node (lack of experience). And I would bet that Fermi 2 and Northern Islands will be the ultimate battle down the line in 2011 as both chips should include MIMD shaders by then. (I know I'm annoying with my MIMD shaders but that is the next big step in GPU, and that is ultimately where Larabee would come into play if it were a fully functionnal solution.)

But essentially, I'm absolutely on edge as to what will Southern Island end up to be performance wise and what the changes will be in the uncore section, as this is where all the new stuff gets implemented. It should help beat Fermi down the road in most benches, well I hope it does anyway. Southern Islands is where my money will go for my next card, for sure. Just anxious to get the details on what it will be in final form, the specs!!!

 

[a_ump]

But if you look at the Fermi chip and take into perspective the technology that's been put into it, Southern Island is basically like the Fermi that Nvidia is trying to mass produce. Only, Nvidia is at the marketing stage, where ATI is at the tape out stage. So in essence Nvidia is about 4 to 6 months in advance. I don't know if anyone is following me on this. Nvidia has a technologically more advanced chip still than ATI's 5xxx generation, but of course it's a big mess with the 40nm node (lack of experience). And I would bet that Fermi 2 and Northern Islands will be the ultimate battle down the line in 2011 as both chips should include MIMD shaders by then. (I know I'm annoying with my MIMD shaders but that is the next big step in GPU, and that is ultimately where Larabee would come into play if it were a fully functionnal solution.)

But essentially, I'm absolutely on edge as to what will Southern Island end up to be performance wise and what the changes will be in the uncore section, as this is where all the new stuff gets implemented. It should help beat Fermi down the road in most benches, well I hope it does anyway. Southern Islands is where my money will go for my next card, for sure. Just anxious to get the details on what it will be in final form, the specs!!!

hmmm maybe, but either way, nvidia ahead or not, i think ATi went about it the smarter way. Use an architecture you understand with a new process. You should always know at least 1 variable really well, which is what ATi seems to do. They either understand the process node very well or understand the architecture on a new process very well. Nvidia this time went with both a new and very complex architecture, as well as a process node they knew very little about it seems or else all their delays woulda been avoided.

I do agree with NI and Fermi 2 being a good showdown as i think Nvidia will do some serious tweaking and changes to the current fermi to streamline it and once that happens i think we'll see the architecture they developed take off. However.....we have no idea what ATi's new architecture is going to be putting up performance wise. The 5 series is keeping up with fermi just fine...so I don't think its too far fetched to think ATi's new architecture with NI will make nvidia stumble.

 

[mastrdrver]

But if you look at the Fermi chip and take into perspective the technology that's been put into it, Southern Island is basically like the Fermi that Nvidia is trying to mass produce. Only, Nvidia is at the marketing stage, where ATI is at the tape out stage. So in essence Nvidia is about 4 to 6 months in advance. I don't know if anyone is following me on this. Nvidia has a technologically more advanced chip still than ATI's 5xxx generation, but of course it's a big mess with the 40nm node (lack of experience). And I would bet that Fermi 2 and Northern Islands will be the ultimate battle down the line in 2011 as both chips should include MIMD shaders by then. (I know I'm annoying with my MIMD shaders but that is the next big step in GPU, and that is ultimately where Larabee would come into play if it were a fully functionnal solution.)

But essentially, I'm absolutely on edge as to what will Southern Island end up to be performance wise and what the changes will be in the uncore section, as this is where all the new stuff gets implemented. It should help beat Fermi down the road in most benches, well I hope it does anyway. Southern Islands is where my money will go for my next card, for sure. Just anxious to get the details on what it will be in final form, the specs!!!

The reason SI and eventually NI will end up with a similar setup to Fermi is because DX11. Not because of some foresight.

nVidia designed Fermi for DX11!!!!
ATI designed Evergreen for dx11.

(Yes I know Fermi is more than just DX11 but I'm illustrating here)

While Fermi is a DX11 architecture, it's way overkill for anything in the next year at least. ATI on the other hand took Evergreen and made it to be playable on what little DX11 will be used in the early DX11 games. NI, with SI in between, will be the move to give ATI a DX11 gpu that will be ready for more gpu demanding DX11 games.

The problem, for nVidia, is that the TSMC 40nm process is not that good causing Fermi to not be what it could have been had TSMC had a better 40nm process.

 

[Velvet Wafer]

i dont know, if someone already posted it, but i found especially the pictures interesting
Zero-Defect Wafers,32nm SOI and 28nm bulk pics

 

[Super XP]

Great link.
O.K. so we know GF is doing the 28nm Bulk Silicon just right and so they also showe some 32nm SOI which was already stated by AMD's CEO somewhere in an article that the 32nm is a no go.