Official AMD Radeon 6000 Series Discussion Thread

[SNICK]

3DMark Vantage (Performance-preset)/3DMark06:

NVIDIA

GeForce GTX 480: P18376/19671
GeForce GTX 460 (256-bit & 1 Gt): P13623/18601
GeForce GTX 460 (192-bit & 768 Mt): P13386/18259
GeForce GTS 450: P9792/15793
AMD

ATI Radeon HD 5870: P17924/19433
AMD Radeon HD 6870: P16270/19480
ATI Radeon HD 5850: P15593/18762
AMD Radeon HD 6850: P14872/18750
ATI Radeon HD 5830: P14014/17298
ATI Radeon HD 5770: P11017/16358
ATI Radeon HD 5750: P9124/14966

 

[cheezburger]

Picture of Radeon 6850 right here:

http://www.presence-pc.com/image/AMD-Radeon-6850-Barts,0101-264573-0-2-3-1-jpg-.html

ati is dead long time ago....

Thoughts on this thread from January?

Most interesting comments are from 265586888. This, this, this, and this.
:

265586888 >intel 4004:

Not correct, Cypress started as 1920 SPUs/384-bit GDDR5, not 2400 SPUs, but the plan changed to 1600 SPUs/256-bit GDDR5 after AMD did their homework with TSMC's 40 nm.

so the cayman was actually the original r870(not cypress) with 4D ALU and some tessellation tweak +better 40nm fab after all....

3DMark Vantage (Performance-preset)/3DMark06:

NVIDIA

GeForce GTX 480: P18376/19671
GeForce GTX 460 (256-bit & 1 Gt): P13623/18601
GeForce GTX 460 (192-bit & 768 Mt): P13386/18259
GeForce GTS 450: P9792/15793
AMD

ATI Radeon HD 5870: P17924/19433
AMD Radeon HD 6870: P16270/19480
ATI Radeon HD 5850: P15593/18762
AMD Radeon HD 6850: P14872/18750
ATI Radeon HD 5830: P14014/17298
ATI Radeon HD 5770: P11017/16358
ATI Radeon HD 5750: P9124/14966

source?

i think(if it's real) the reason why barts can't be any faster than cypress was because of TMU and bandwidth bottleneck. consider cypress had 80 while barts only have 64 and bandwidth 153 compare to barts's 134. but still i expecting better driver and higher performance over cypress.

 

[crazyeyesreaper]

it was already proven more then once tho that the 5800 series did not have a memory bottleneck they were starved for more ROPs the new gpus seem to have the same issue from the looks

 

[pantherx12]

it was already proven more then once tho that the 5800 series did not have a memory bottleneck they were starved for more ROPs the new gpus seem to have the same issue from the looks

Is that actually proven though?

Could of been the shader architecture, hence the change?

Correct me if I'm wrong.

But would of seemed a rather stupid thing to overlook a problem like for the engineers, lots of them all very well trained.

 

[crazyeyesreaper]

wolf benet and others ran alot of tests a 30-40mhz core clock gave more performance then 150-200mhz on the memory if the memory was the issue then scaling would have been much better but .5fps from 200mhz on the ram is a bottleneck not lol

 

[pantherx12]

wolf benet and others ran alot of tests a 30-40mhz core clock gave more performance then 150-200mhz on the memory if the memory was the issue then scaling would have been much better but .5fps from 200mhz on the ram is a bottleneck not lol

All that proves is that memory is not the bottle neck, not that rops are the bottle neck by the way.

 

[Steevo]

I did the same, the memory is not the issue, the setup engine and shaders arrangement are the bottleneck. One triangle per clock.

 

[cheezburger]

it was already proven more then once tho that the 5800 series did not have a memory bottleneck they were starved for more ROPs the new gpus seem to have the same issue from the looks

judging the specification between barts and cypress only have little different rops (32 vs 32) bus (256bit vs 256bit) and ALU(320 vs 320). the mirror performance loss between barts and cypress can only be determined by less bandwidth (134 vs 153, 1050mhz vs 1200mhz GDDR5) and texture filling unit (64 vs 80) which these are really important under dx9/10 games

 

[crazyeyesreaper]

its not memory speed at 850/1000 my in game frame rate is 1fps difference from 850/1200 so 200mhz gave me a 1 fps in most games in others i get 2-3max where as the 125mhz on the core gave me a 6-8fps boost in general ram speed or bandwidth was never really the issue

but since they changed the shader arrangement its probably far more efficient in doing the same work so lesser bandwidth shouldnt hurt performance and id much rather see lower avg frame rates but better minimum frame rates thats the biggest issue i have with ATi / AMD gpus minimum frame rate is usually far lower then a competing Nvidia gpu these days which probably relates more to what Steevo touched upon

 

[erocker]

Yeah, it's not memory bandwith holding things back at all. This can be easily proven by factual testing/benchmarking.

 

[cheezburger]

its not memory speed at 850/1000 my in game frame rate is 1fps difference from 850/1200 so 200mhz gave me a 1 fps in most games in others i get 2-3max where as the 125mhz on the core gave me a 6-8fps boost in general ram speed or bandwidth was never really the issue

but since they changed the shader arrangement its probably far more efficient in doing the same work so lesser bandwidth shouldnt hurt performance and id much rather see lower avg frame rates but better minimum frame rates thats the biggest issue i have with ATi / AMD gpus minimum frame rate is usually far lower then a competing Nvidia gpu these days which probably relates more to what Steevo touched upon

that might explain why barts has less max fps in most case. however in 3dmarks bandwidth takes differently than real time gaming even thought bandwidth may seem merely nothing in gaming performance gain(of cause you can't go too low either...). other major issue is less TMU because less TMU/rop = less fps.

 

[Super XP]

I really don't agree with AMD's changing the naming scheme, I think it's a big mistake and will cause customer confusion, but that is rumour. Another thing that is rumour is the rumoured fact the HD 6800 series will be based on Northern Islands and Evergreen.

If this is in fact true, that may very well explain the name change. Think about this for a sec? HD 6700 & HD 6800 is a modified HD 5800 with tweaks and such. The HD 6900's may very well be based on the Hybrid design NI & Evergreen combined

 

[mastrdrver]

See that is one of my problems I have with the rumors. If Barts is what it is rumored then the die will end up about the same as Cypress. You can't sell a card with that big of a chip at the price point that the GTX 460 is at which is what is rumored Barts is aiming for.

I still wonder how soon TSMC or GF 28nm is coming as I'm beginning to believe that there might be a 4770 (if you will) Barts on 40nm that is just for testing architecture for the parts coming on 28nm which some rumors say 1Q 2011.

This slide vs this slide show two different Barts.

 

[cheezburger]

I really don't agree with AMD's changing the naming scheme, I think it's a big mistake and will cause customer confusion, but that is rumour. Another thing that is rumour is the rumoured fact the HD 6800 series will be based on Northern Islands and Evergreen.

If this is in fact true, that may very well explain the name change. Think about this for a sec? HD 6700 & HD 6800 is a modified HD 5800 with tweaks and such. The HD 6900's may very well be based on the Hybrid design NI & Evergreen combined

it won't be hybrid, it's more like mixing line between generation. like rebrand of 6770/6750 for example some mid-range evergreen will be co-exist with northern islands until 28nm southern islands come with die shrink and another double up. this situation is more like g92/94 rebranded because of gt200 yield issue. keep in mind this is the only way they can do. go bybrid design will just delay the further new architecture development since their main uncore architecture hadn't change since r300.

these rumor about r6xxx will be hybrid design have no bases in today as most of these rumor are nearly 8 months ago and spurce hadn't update ever since .

See that is one of my problems I have with the rumors. If Barts is what it is rumored then the die will end up about the same as Cypress. You can't sell a card with that big of a chip at the price point that the GTX 460 is at which is what is rumored Barts is aiming for.

I still wonder how soon TSMC or GF 28nm is coming as I'm beginning to believe that there might be a 4770 (if you will) Barts on 40nm that is just for testing architecture for the parts coming on 28nm which some rumors say 1Q 2011.

This slide vs this slide show two different Barts.

28nm will not be available any time soon or else amd will wait for the new fab than rush NI this bad.

 

[Super XP]

Well I am hoping for a 60% increase in performance with HD 6870 over the current HD 5870 and another 120% increase with HD 6970 over HD 5870. We shall see...

 

[p_o_s_pc]

Well I am hoping for a 60% increase in performance with HD 6870 over the current HD 5870 and another 120% increase with HD 6970 over HD 5870. We shall see...

I don't see it happening till they can shrink it. there is always that chance

 

[cheezburger]

Well I am hoping for a 60% increase in performance with HD 6870 over the current HD 5870 and another 120% increase with HD 6970 over HD 5870. We shall see...

shader performance will definitely increase, but less number of TMU and smaller bandwidth had become bottleneck. and high end like cayman and dual gpu like antilles won't be much faster than current 5970 in older title because of TMU didn't increase as much. cayman's 96 TMU is just little bit superior to cypress's 80 and antillies is just barts x2..which had little advantage competing hamlock which was cypress x2. tessellation will definitely increasing to insanely rate because of new shader power, may be 10x in comparison between cayman and cypress but under dx9/10 it will only be 35~40% increase due to 48 rop and 384bit bus cause bad scaling problem. so the performance increase not as good as we expecting since new architecture suppose to be increase 2x + in term of performance but it isn't.

however that's rumor since later source suggest that new architecture allow raster pipe able to do 4 pixel per clock which means 48 rops can be 192 rops effective and 32 rops can be 128rops effective...who knows it may be actually be far superior than current hd 5xxx evergreen family(cypress/huniper can only do 2 pixel per clock in each pipeline, same as nvidia counterpart.).

 

[jaredpace]

so the cayman was actually the original r870(not cypress) with 4D ALU and some tessellation tweak +better 40nm fab after all....

You know what, they might actually keep the VLIW-5 ALU design on N.I. Not saying 4D is ruled out completely as a possibility since I don't know for sure, but I'll wager it's VLIW-5 still, and that the entire front-end and setup engine has been redesigned.

Then they'll save the 4D shader redesign to go with that new front-end on the HD7k cards once 28nm is available.

The rumors have been so strongly leaning towards 4D, some numbers don't add up when considering wavefronts & texture mapping unit ratio, suggesting that 4D is not the case here.

Gispel wrote at 3DC that drivers suggested 4D, and then a rumor came from Taiwan saying 1920sp 4d < 400mm2. Charlie wrote shortly after in April that it was a new frontend with a larger number of Evergreen shaders (suggesting keeping VLIW-5). Then 4D rumor was back in full swing from April-Oct. Now weeks from launch Obermeyer says 4D is false, and nothing in his media supports that idea. I think 960 5-way on Barts and 1920 5-way on Cayman. The impressive performance a result of a new Ultra threaded dispatch engine.

O well, find out in a week whats what.

 

[mastrdrver]

You know what, they might actually keep the VLIW-5 ALU design on N.I. Not saying 4D is ruled out completely as a possibility since I don't know for sure, but I'll wager it's VLIW-5 still, and that the entire front-end and setup engine has been redesigned.

Then they'll save the 4D shader redesign to go with that new front-end on the HD7k cards once 28nm is available.

The rumors have been so strongly leaning towards 4D, some numbers don't add up when considering wavefronts & texture mapping unit ratio, suggesting that 4D is not the case here.

Gispel wrote at 3DC that drivers suggested 4D, and then a rumor came from Taiwan saying 1920sp 4d < 400mm2. Charlie wrote shortly after in April that it was a new frontend with a larger number of Evergreen shaders (suggesting keeping VLIW-5). Then 4D rumor was back in full swing from April-Oct. Now weeks from launch Obermeyer says 4D is false, and nothing in his media supports that idea. I think 960 5-way on Barts and 1920 5-way on Cayman. The impressive performance a result of a new Ultra threaded dispatch engine.

O well, find out in a week whats what.

That would make a lot of sense (about the shader part). The front end has always said to constrain the shaders on Evergreen and the rumor about the hybrid setup goes back to about the time Charlie broke TSMC canceling 32nm. Though I still think it is the same N.I. just back ported to 40nm:

1) Would explain 256 bus on the mid range 6770 Barts (more pad area)
2) Not enough time (die shot on base part has 28th week of 2010 which is not enough time to design anything let along have a working board in 28 weeks especially since 32nm was canned at the very end of November)

Now whether N.I. was a hybrid or next gen architecture idk and would explain the 5770 rebrand rumor.

With everyone rambling now that Barts is 6870 I'd be amazed to see that. You're telling me that a just above 150W part is going to land between 5850 and 5870 performance while being roughly as big and sell for just above $220? That's terrible performance for die size and that they actually lost! performance. Then they are going to make a generation out of this design?

If Barts was a little bigger than Juniper at say 200mm2 that would sound more logical.

 

[Benetanegia]

That would make a lot of sense (about the shader part). The front end has always said to constrain the shaders on Evergreen and the rumor about the hybrid setup goes back to about the time Charlie broke TSMC canceling 32nm. Though I still think it is the same N.I. just back ported to 40nm:

1) Would explain 256 bus on the mid range 6770 Barts (more pad area)
2) Not enough time (die shot on base part has 28th week of 2010 which is not enough time to design anything let along have a working board in 28 weeks especially since 32nm was canned at the very end of November)

Now whether N.I. was a hybrid or next gen architecture idk and would explain the 5770 rebrand rumor.

With everyone rambling now that Barts is 6870 I'd be amazed to see that. You're telling me that a just above 150W part is going to land between 5850 and 5870 performance while being roughly as big and sell for just above $220? That's terrible performance for die size and that they actually lost! performance. Then they are going to make a generation out of this design?

If Barts was a little bigger than Juniper at say 200mm2 that would sound more logical.

Wasn't Barts supposed to be sized somewhere around 280 mm^2? That's what I have always thought since I read it somewhere and in that case perf/die area is about the same we've seen from other 40nm AMD cards. It's Cayman which is bigger than Cypress, >400 mm^2.

Also the leaked specs said tdp was >150w for Barts XT and <150w for Barts pro. HD5850's power consumption is 150w while HD5870's is 200w+. It just fits. Barts' performance is somewhat lower than Cypress, it only makes sense that tdp is also a little bit lower, i.e 130w and 180w, which translated to marketing talk becomes <150w and >150w at the time of writing the leaked chart, because at the time there's no way they could know the specifics of the final cards, when the final specs were nowhere near finalized...

 

[mastrdrver]

Charlie says 10-15% larger than Cypress for top N.I. gpu

Use this for the 6770 chip which should be half of 6870 and you end up with ~200mm2 part. He also states this in the Barts thread on the S/A forum.

If it was as you suggested then what's the point of putting out an entire new line if performance is a little slower? Also if 6770 was 240mm2 it isn't going to sell for ~$200. Too big.

 

[Mindweaver]

Radeon HD 6800 performance figures leaked

Check out the full story here.

3DMark Vantage (Performance-preset)/3DMark06:

NVIDIA

GeForce GTX 480: P18376/19671
GeForce GTX 460 (256-bit & 1 Gt): P13623/18601
GeForce GTX 460 (192-bit & 768 Mt): P13386/18259
GeForce GTS 450: P9792/15793

AMD

ATI Radeon HD 5870: P17924/19433
AMD Radeon HD 6870: P16270/19480
ATI Radeon HD 5850: P15593/18762
AMD Radeon HD 6850: P14872/18750
ATI Radeon HD 5830: P14014/17298
ATI Radeon HD 5770: P11017/16358
ATI Radeon HD 5750: P9124/14966

 

[KainXS]

someone beat you to it

Official AMD Radeon 6000 Series Discussion Thread

 

[Mindweaver]

Well shit.. Sorry guys.. Can a mod delete this?

 

[JATownes]

Well, IF this is true, I am disappointed. An HD6870 scores lower than the 5870? That will lead to massive confusion in the retail market. Also with the scores, the 6870 only produces 19480 in 3DMark06? I break 20k with my overclocked 4850s. This is just sad, but then again, who knows if this is true.

Also, the pic in the link of the "supposed" 6850 shows an ATI label on it. If these are Barts, shouldn't they have AMD labels?