HD 6000 - The rumor mill brewing

[Dark4ng3l]

Well I was hoping for about 20% improvement in a general sense with better improvements in situations where the current architecture is pretty limited(excessive unigine heaven type tessellation). I read the rumors about the 5 to 4 shader change and the additional shaders on the chip but as far as I know details have been really scarce and most of it has to be speculation.

Lets all hope that it's 40% faster but I think something along the lines of 20%(including some additional clock speed) is more realistic.

 

[nemesismk2]

I don't trust anything on fudzilla, I stick to websites I trust like Anandtech for all my info. I just don't like fud lol

Thought I would recycle my post from the other topic about news from Fudzilla

 

[MrK6]

Well I was hoping for about 20% improvement in a general sense with better improvements in situations where the current architecture is pretty limited(excessive unigine heaven type tessellation). I read the rumors about the 5 to 4 shader change and the additional shaders on the chip but as far as I know details have been really scarce and most of it has to be speculation.

Lets all hope that it's 40% faster but I think something along the lines of 20%(including some additional clock speed) is more realistic.

Last I heard about tesselation, the 6xxx midrange cards were going to have similar tesselation performance to AMD's current models (which all have the same tesselation unit), 6xxx high-end will have double it, and the flagship model will have four times the current tesselation power. I have no idea how concrete that is though nor if it's changed.

 

[Lonyo]

Last I heard about tesselation, the 6xxx midrange cards were going to have similar tesselation performance to AMD's current models (which all have the same tesselation unit), 6xxx high-end will have double it, and the flagship model will have four times the current tesselation power. I have no idea how concrete that is though nor if it's changed.

Well, based somewhat on the GTX460 testing from Anandtech, it looks like there's a limit to how much tessellation power is currently needed (the DX11 testing thing in the GTX460 article shows the GTX460 and GTX480 getting the same score), so really AMD only need to improve their tessellation capability up to a certain point.

 

[ModestGamer]

Without even looking it up, my guess is that intel sees much higher profit margins per each processor sold. market share is only half the story.

this is true. for now. but they charge alot for minimal performance differences. but hardcore users alway pay for the best.

 

[evolucion8]

The performance gain from the 4870 over the 3870 was ~92% the increase in total ALU count was 2.5 fold (64 to 160). The performance gain from the 5870 over the 4870 was ~68% and the increase in ALU count was two fold (160 to 320). This can give point to measure theoretical performance of the next gen; my personal expectation is a gain of ~40-45% over the 5870 as a result of ALU count increasing from 320 to 480. The new chip's rumored size is 40nm @ ~395mm^2 compared to the 334mm^2 Cypress and the 535mm^2 Fermi.

But the HD 4870 usually proved to be at least twice faster than the HD 3870 and even faster than that in many situations with lots of eye candy due to the AA bug in the HD 3870 and its weak texturing performance. It was more like three times faster, also it's TMU were beefed up, the implementation of Hardware AA resolve and had a more efficient memory controller.

But since the HD 4870 was such a strong performer, that explain why the HD 5870 wasn't twice faster considering that the HD 5870 is about twice of everything compared to the HD 4870 theorically, because the HD 5870 shader processors are less efficient compared to the HD 4870 (Look at the Anandtech's HD 4550 vs 5450 review), means that a nice performance jump can be obtained with the HD 6870 if your information is correct, plus I think that a nice 10%-20% additional performance can be gained if all the optimizations used in the HD 4870 are implemented or improved in the HD 6870. One thing is certain, the chip will be bigger than Cypress if the 40nm process is used, and yet, will be smaller than Thermi.

 

[tviceman]

I'll be happy if the 6870 is on par with the 5970 or even a few clicks faster. This would put it 20% faster than a GTX480 and 10% faster than

 

[Kenmitch]

EDIT: I think it'll be a decent new part, I just do not believe at all that ATI will get 5970 performance from any single GPU on 40nm.

I doubt they could with a single chip also....But I could see a dual chip solution being able to surpass the 5970 with this new hybrid run.

Now once on the next node I see no reason why they wouldn't be able to spin a single chip solution that would surpass the 5970 performance level.

 

[Dark4ng3l]

Then prepared to be disappointed because I don't think there is any way they're going to come to market with a new high end GPU on the same node process that is as fast or faster than their current X2 card. The power envelope at 40nm, coupled with increased die sizes and temperatures and smaller yields just make this completely unrealistic. We are talking about a new GPU on the same node process as the previous generation. Considering they went from a hd4890 @ 55nm to a 5870 @ 40 nm but increased die size by 27% , and got around a 50% performance increase should put your idea's into perspective. Your expectations are just way too high.

EDIT: I think it'll be a decent new part, I just do not believe at all that ATI will get 5970 performance from any single GPU on 40nm.

One thing that needs to be kept in mind when comparing Evergreen to previous generations of R600 derived cards is that there is quite a bit of die space going towards support of DX11 features that take up some of the space that would normally be used to just add more shaders or whatever. If they had just shrunk the 4890 they could probably have made a 3200sp version of it or something. The fact that Evergreen is less efficient per clock than the previous generation probably has to do with the hardware needing to be able to do more stuff than before.

 

[tviceman]

I doubt they could with a single chip also....But I could see a dual chip solution being able to surpass the 5970 with this new hybrid run.

Now once on the next node I see no reason why they wouldn't be able to spin a single chip solution that would surpass the 5970 performance level.

Well if they are to make a 2X part based on the 6000 series cards, remember that the TDP still has to remain at or under 300 watts. And if the new chips are going to be bigger and more power hungry, it's not a given that they'll be able to combine two chips which will stay under the power envelope and still noticeably outperform a 5970. Then again, they may be able to but it also might not make much sense if the gains vs. costs don't justify it.

One thing that needs to be kept in mind when comparing Evergreen to previous generations of R600 derived cards is that there is quite a bit of die space going towards support of DX11 features that take up some of the space that would normally be used to just add more shaders or whatever. If they had just shrunk the 4890 they could probably have made a 3200sp version of it or something. The fact that Evergreen is less efficient per clock than the previous generation probably has to do with the hardware needing to be able to do more stuff than before.

That's fine I accept that, but I still 100% believe that to expect a 6x00 card on 40nm node to perform equal to or better than a 5970 is just setting yourself up for a big disappointment.

 

[MagickMan]

I thought we already knew that SI was just a refresh with a little better performance and a few new features. :\ NI is the "next big thing".


woot, 1000 posts.

 

[evolucion8]

One thing that needs to be kept in mind when comparing Evergreen to previous generations of R600 derived cards is that there is quite a bit of die space going towards support of DX11 features that take up some of the space that would normally be used to just add more shaders or whatever. If they had just shrunk the 4890 they could probably have made a 3200sp version of it or something. The fact that Evergreen is less efficient per clock than the previous generation probably has to do with the hardware needing to be able to do more stuff than before.

Not only that, the fact that they needed to keep transistor budget low while implementing DX11 functionality, it left out no space for additional improvements without making the die too big to fit a certain size. Because if it was only related to DX11, the card should perform like supposed to with other API's like DX10 and DX9, and yet, those shows that the Evergreen architecture is slower in a per clock basis.

 

[bryanW1995]

I thought we already knew that SI was just a refresh with a little better performance and a few new features. :\ NI is the "next big thing".


woot, 1000 posts.

well, NI was supposed to be coming out this fall. unfortunately, tsmc didn't get the memo and wasn't ready with 28nm yet. So amd is incorporating parts of NI into this falls refresh part SI. Nobody knows (or at least those who know aren't talking) which parts of NI will be incorporated into SI. if they just wanted to keep the same gpu size then yes it would be tough to get more than 10-20% more out of it; however, as they are going to have some/many/most of a new arch and appear willing to increase the die size a bit then it's anybody's guess what high SI will look like.

 

[MentalIlness]

Link added in OP.

 

[IntelUser2000]

well, NI was supposed to be coming out this fall. unfortunately, tsmc didn't get the memo and wasn't ready with 28nm yet. So amd is incorporating parts of NI into this falls refresh part SI.

28nm was always H2 2011. It was the cancellation of 32nm that caused the trouble.

110
90
80
65
55
40
32=dead
28

Radeon HD5870 did double everything over the 4870 except the memory bandwidth. That would factor in high resolutions as well.