Official AMD Radeon 6000 Series Discussion Thread

[dj-electric]

hmm... i dont think its bad, i mean... the juniper is a nice little core

 

[SNICK]

Barts die shot

Guys over chiphell have posted die shots of barts


die area:171.09*134.61=230.30 mm^2

OFFICIAL RADEON HD 6800 SERIES DRIVER:http://www.station-drivers.com/page/ati catalyst.htm

 

[Benetanegia]

i don know

seems juniper did really get rebranded though.

just compare those slides with these:

and what turned out to be relity, HD5870 for $400 and only 25% faster than GTX285.

I do think they are just being more realistic this time around, mostly because the expectations seen in these charts are the smilar to what we found in the leaked benchmarks, but my comment was more towards the fact that most people were always and still expecting Barts to be as fast/faster than Cypress and Cayman beating Hemlock, which couldn't be farther from truth. Of course, in reality everybody knew it was not going to happen.

 

[Lionheart]

1GB of memory better not be

 

[a_ump]

um.....yea i wont say or judge anything till its released. I honestly can't imagine that with the new shader arch and 960SPU's performing roughly on par with the 1600 in HD 5870 in vantage, that cayman is only going to be a notch/hair faster than the HD 5870. I say....a cayman delay to see what nvidia does, n that the 6970 is going to be roughly 40% faster than HD 5870

 

[cheezburger]

1GB of memory better not be

again that was another slide....cayman is suppose to be 3/2 of barts 48 rops means it's going to be 1536mb (128mb x12). 1gb will either be 64 rops (64mb x16) or 32rops (128mb x8) however these number won't match to 480 ALU and 96TMU/3RPE. i doubt it would be impossible to just adding shader without increase the rest of spec. other thought chiphell already given so many false informed which that's point completely opposite natural gpu design...(480 ALU on 32rops and 256bit bus +6gbp cayman + r6870 barts was only 128bit with also ridiculous shader number but only 16 rops rumor were quitely rumor for a while in chinese website then it spread more non sense about cayman been 48rops but 256bit bus (12x4 config that rumor through forum for a while but it is against the 3RPE config that cayman had)which these people don't know anything about gpu engineering and non of them turn to be real except the rebrand part.

I don't see the point why should we trust a chinese website because they were coincidentally predict the launch of evergreen. they were rumor that cypress was going to be 2.2x more powerful in last 2 months but end up only 20~25% above gtx 285.....i'd go with fud more than these chinese people that were merely just form of bunch gamer rather than serious engineer with serious attitude and some of them don't even know the tech inside the gpu itself and only knowing how to overclock.

um.....yea i wont say or judge anything till its released. I honestly can't imagine that with the new shader arch and 960SPU's performing roughly on par with the 1600 in HD 5870 in vantage, that cayman is only going to be a notch/hair faster than the HD 5870. I say....a cayman delay to see what nvidia does, n that the 6970 is going to be roughly 40% faster than HD 5870

barts is not 960sp but 1280sp which is same number of ALU as cypress and all spec are exactly same except new architecture and and less number of TMU.

 

[inferKNOX]

again that was another slide....cayman is suppose to be 3/2 of barts 48 rops means it's going to be 1536mb (128mb x12). 1gb will either be 64 rops (64mb x16) or 32rops (128mb x8) however these number won't match to 480 ALU and 96TMU/3RPE. i doubt it would be impossible to just adding shader without increase the rest of spec. other thought chiphell already given so many false informed which that's point completely opposite natural gpu design...(480 ALU on 32rops and 256bit bus +6gbp cayman + r6870 barts was only 128bit with also ridiculous shader number but only 16 rops rumor were quitely rumor for a while in chinese website then it spread more non sense about cayman been 48rops but 256bit bus (12x4 config that rumor through forum for a while but it is against the 3RPE config that cayman had)which these people don't know anything about gpu engineering and non of them turn to be real except the rebrand part.

I don't see the point why should we trust a chinese website because they were coincidentally predict the launch of evergreen. they were rumor that cypress was going to be 2.2x more powerful in last 2 months but end up only 20~25% above gtx 285.....i'd go with fud more than these chinese people that were merely just form of bunch gamer rather than serious engineer with serious attitude and some of them don't even know the tech inside the gpu itself and only knowing how to overclock.



barts is not 960sp but 1280sp which is same number of ALU as cypress and all spec are exactly same except new architecture and and less number of TMU.

Their proximity to Taiwan makes me believe that they could get a glimpse/picture of the GFX cards through some connection of theirs perhaps involved in production... maybe non-working models are snatched, quickly taken a pic of, then returned to the bin or something? It's possible IMO...

 

[Benetanegia]

again that was another slide....cayman is suppose to be 3/2 of barts 48 rops means it's going to be 1536mb (128mb x12). 1gb will either be 64 rops (64mb x16) or 32rops (128mb x8) however these number won't match to 480 ALU and 96TMU/3RPE. i doubt it would be impossible to just adding shader without increase the rest of spec. other thought chiphell already given so many false informed which that's point completely opposite natural gpu design...(480 ALU on 32rops and 256bit bus +6gbp cayman + r6870 barts was only 128bit with also ridiculous shader number but only 16 rops rumor were quitely rumor for a while in chinese website then it spread more non sense about cayman been 48rops but 256bit bus (12x4 config that rumor through forum for a while but it is against the 3RPE config that cayman had)which these people don't know anything about gpu engineering and non of them turn to be real except the rebrand part.

I don't see the point why should we trust a chinese website because they were coincidentally predict the launch of evergreen. they were rumor that cypress was going to be 2.2x more powerful in last 2 months but end up only 20~25% above gtx 285.....i'd go with fud more than these chinese people that were merely just form of bunch gamer rather than serious engineer with serious attitude and some of them don't even know the tech inside the gpu itself and only knowing how to overclock.



barts is not 960sp but 1280sp which is same number of ALU as cypress and all spec are exactly same except new architecture and and less number of TMU.

I think it's time you all finally realize that Cayman is simply 256 bit and 32 ROPs, like I've been saying for a long time.

32 ROPs are more than enough:

What you are looking at is 16 ROPs @800 Mhz (12.8 GPixel/s), beating 16 ROPs @850 Mhz (13.6 Gpixel/s) by 11%, that is, HD5830 beating HD5770. ROPs are clearly not the bottleneck. But let's say that the bottleneck for 16 ROPs is around there, around the 100% relative performance that HD5830 represents... well ok, then the bottleneck for 32 ROPs @800 Mhz would be twice as high, or in the chart above it would represent 200%, much higher than the 163% that the HD5870 represents, and that's at 800 Mhz, while Cayman XT will probably be clocked at 900 Mhz...

OH and btw, ROPs are NOT going to be inside the RPEs, most probably. For various reasons ROPs and MC are better outside your compute engines. RPEs are most probably the same as Nvidia's GPC's, shaders+tmu+tesselation+(maybe) vertex setup+(maybe) rasterizer.

 

[cheezburger]

I think it's time you all finally realize that Cayman is simply 256 bit and 32 ROPs, like I've been saying for a long time.

32 ROPs are more than enough:

http://tpucdn.com/reviews/MSI/HD_5830_Twin_Frozr_II/images/perfrel_2560.gif

What you are looking at is 16 ROPs @800 Mhz (12.8 GPixel/s), beating 16 ROPs @850 Mhz (13.6 Gpixel/s) by 11%, that is, HD5830 beating HD5770. ROPs are clearly not the bottleneck. But let's say that the bottleneck for 16 ROPs is around there, around the 100% relative performance that HD5830 represents... well ok, then the bottleneck for 32 ROPs @800 Mhz would be twice as high, or in the chart above it would represent 200%, much higher than the 163% that the HD5870 represents, and that's at 800 Mhz, while Cayman XT will probably be clocked at 900 Mhz...

if you realize it can be clok up really high then what is the point release hd 6xxx? i dont see the more most of spec are exactly the same from cypress then they can just release a higher clock cypress r890 instead and cayman is completely unnecessary....why do you need a card when you can just clock up yourself with the lower version?....like 4890 was unnecessary for its existence. also it's shader number most no be 1920 but 1280 or 2560 or some even number. and i'd like to see how 3RPE can be divide into 32rops...

hell no, if cayman turns out to be only 32rops and 256bit bus while charging for 400~500 bulks that may really attach to the recent news of intel's huge revenue and why amd desperately need to earn cash because of to covering the loss in cpu market. if this turn out to be true then i'd just have to say: amd please go bankrupt and let some qualified people to taking care of business.....

 

[Benetanegia]

if you realize it can be clok up really high then what is the point release hd 6xxx? i dont see the more most of spec are exactly the same from cypress then they can just release a higher clock cypress r890 instead and cayman is completely unnecessary....

hell no, if cayman turns out to be only 32rops and 256bit bus while charging for 400~500 bulks that may really attach to the recent news of intel's huge revenue and why amd desperately need to earn cash because of to covering the loss in cpu market. if this turn out to be true then i'd just have to say: amd please go bankrupt and let some qualified people to taking care of business.....

Well first of all, because Evergreen couldn't be clocked higher, not enough to make a difference really, unless they wanted to destroy any OC capabilities. Secondly because they want to give the impresion that they release a new genaration every year or so. AMD has always counted HD3xxx series as a new generation, fact.

Third, shaders are not the same and they want/need to test them before they move to 28nm. A RV740 cannot be made this time around because the process is the one thing that comes later than expected.

And finally, you can be sure that Nvidia is going to release something before 28nm and that is going to be miles better than Fermi in regards to perf/watt, so they need something better than Evergreen to fight that, even if it's only slightly better, to stay ahead. Like I said, ROP count has nothing to do with actual performance, so Cayman being 32 ROPs has nothing to do with being the same as Cypress, not at all.

 

[cheezburger]

Well first of all, because Evergreen couldn't be clocked higher, not enough to make a difference really, unless they wanted to destroy any OC capabilities. Secondly because they want to give the impresion that they release a new genaration every year or so. AMD has always counted HD3xxx series as a new generation, fact.

Third, shaders are not the same and they want/need to test them before they move to 28nm. A RV740 cannot be made this time around because the process is the one thing that comes later than expected.

And finally, you can be sure that Nvidia is going to release something before 28nm and that is going to be miles better than Fermi in regards to perf/watt, so they need something better than Evergreen to fight that, even if it's only slightly better, to stay ahead. Like I said, ROP count has nothing to do with actual performance, so Cayman being 32 ROPs has nothing to do with being the same as Cypress, not at all.

i'd agree with you at first and second as they want to make impression and continue tradition releasing new line every years and testing new architecture and release line as guinea pig, like 40nm on 4770. i'd agree too. but still having cayman is completely unnecessary when barts can just do the good job as tech demo and pretty much able to clock higher than cypress, with much lower price. they may will also encounter problem because of 3RPE design cannot happen on 256bit bus and 32rops because 1920sp is apply to that design like formerly r870 did before tape out and shrink as cypress. despite of early rumor in january that suggest of uncore hybrid it still won't happen on cayman to allow have 1920sp while only 32rops and 256bit bus such blizzard configuration which is impossible to have . it's just common sense and entry level of understanding in gpu architecture.

raster pipe is as important as shader, as you can see these amd card on bench above were clock insanity high compare to nv counterpart. of cause you can have 16rops without battleneck when it's on super clock. but then again do you expecting people would pay for more money on these card that has little different than cheaper stock clock version or even the previous gen when you can just tweak it yourself with these card? cypress can't clock higher is depending on "how high" in you standard. to me, having ghz speed gpu is no as good comparing having powerful spec. that is what you pay for. not just some tweak but you pay for quality that you feel worthy. that was why i pay for my gtx 275 rather than choose 4890. at least i feel the money is worthy and 4890 has no different than 4870 to me and no worth to spend $299 when i could just buy a 4870 for $229 back to the day. if you think you can just define market position by just measure by clock I think you are terribly mistake on misunderstand consumer's intelligent. same applies to amd if they are just selling a card that has little different except shader while cost twice as barts will make consumer look like fool. if you just want better clock over cypress than barts is good enough.


like i stated before cayman will be a huge failure unless it cost only $50 dollars higher than barts if the spec is 32rop and 256bit bus. again i would rather spend on a huge chip for 600~700 dollar than spend 400~500 dollar on a chip that's merely just a mid range with insane clock that you can just do it yourself. you can change the clock but you can never change the spec, that is the point why it's worth to pay for larger chip/powerful spec. if barts's price does not set below 250 dollars it will be tragic event for amd. period. do not expecting people would go buy a overclock version cypress for 300+ dollars after one years of release.


let's go back to rops, 16rops in real time operation has to be 2x clock to be competitive to 1/2 clock 32rops. a 32 rops has to be 1.5x clock to equal to 48rops at 2/3 clock. that's been say if you can oc it really high you still not able to out pace larger rops when it also overclock. what you replying rops does not bottle neck has completely no basis. stop bring me these bench that how 4870 compete with 260 and 4890 to 275 while ignore their clock are nearly twice as high....also cayman is having 3 RPE and i really like to see how they can accomplish such amazing achievement by dividing 32 rops on 3RPE raster cluster. which i'd rather believe cayman is not real and barts will replace cypress with better overclock ability.

 

[20mmrain]

Or these are the most conservative marketing (read horrible marketing) slides ever released or I smell dissapointment.

Or because it's a new architecture.... they are the ones being very conservative. Meaning the last time they released a new architecture it was a disaster at first.

A possibility that they start off lower then it's real capability saving it's full realization for the 7000 series and a die shrink on the thermal side.

 

[Benetanegia]

i'd agree with you at first and second as they want to make impression and continue tradition releasing new line every years and testing new architecture and release line as guinea pig, like 40nm on 4770. i'd agree too. but still having cayman is completely unnecessary when barts can just do the good job as tech demo and pretty much able to clock higher than cypress, with much lower price. they may will also encounter problem because of 3RPE design cannot happen on 256bit bus and 32rops because 1920sp is apply to that design like formerly r870 did before tape out and shrink as cypress. despite of early rumor in january that suggest of uncore hybrid it still won't happen on cayman to allow have 1920sp while only 32rops and 256bit bus such blizzard configuration which is impossible to have . it's just common sense and entry level of understanding in gpu architecture.

raster pipe is as important as shader, as you can see these amd card on bench above were clock insanity high compare to nv counterpart. of cause you can have 16rops without battleneck when it's on super clock. but then again do you expecting people would pay for more money on these card that has little different than cheaper stock clock version or even the previous gen when you can just tweak it yourself with these card? cypress can't clock higher is depending on "how high" in you standard. to me, having ghz speed gpu is no as good comparing having powerful spec. that is what you pay for. not just some tweak but you pay for quality that you feel worthy. that was why i pay for my gtx 275 rather than choose 4890. at least i feel the money is worthy and 4890 has no different than 4870 to me and no worth to spend $299 when i could just buy a 4870 for $229 back to the day. if you think you can just define market position by just measure by clock I think you are terribly mistake on misunderstand consumer's intelligent. same applies to amd if they are just selling a card that has little different except shader while cost twice as barts will make consumer look like fool. if you just want better clock over cypress than barts is good enough.


like i stated before cayman will be a huge failure unless it cost only $50 dollars higher than barts if the spec is 32rop and 256bit bus. again i would rather spend on a huge chip for 600~700 dollar than spend 400~500 dollar on a chip that's merely just a mid range with insane clock that you can just do it yourself. you can change the clock but you can never change the spec, that is the point why it's worth to pay for larger chip/powerful spec. if barts's price does not set below 250 dollars it will be tragic event for amd. period. do not expecting people would go buy a overclock version cypress for 300+ dollars after one years of release.


let's go back to rops, 16rops in real time operation has to be 2x clock to be competitive to 1/2 clock 32rops. a 32 rops has to be 1.5x clock to equal to 48rops at 2/3 clock. that's been say if you can oc it really high you still not able to out pace larger rops when it also overclock. what you replying rops does not bottle neck has completely no basis. stop bring me these bench that how 4870 compete with 260 and 4890 to 275 while ignore their clock are nearly twice as high....also cayman is having 3 RPE and i really like to see how they can accomplish such amazing achievement by dividing 32 rops on 3RPE raster cluster. which i'd rather believe cayman is not real and barts will replace cypress with better overclock ability.

More ROPs are only necessary when they ARE necessary. That's what you constantly fail to understand, despite your vast knowledge on GPU architecture... And yes I'm being sarcastic. HD 5870 with 32 ROPs is NOT bottlenecked by its ROP count by any means. In fact, its 32 ROPs clocked at 850 Mhz are nothing but overkill. Nvidia cards? Even more overkill for games (still necessary for GPGPU and CAD/CAM though). You could add 2497 ROPs if that's what you wanted and NO it would not make any difference in gaming performance. You are speaking with a book in your hands " 16rops in real time operation has to be 2x clock to be competitive to 1/2 clock 32rops. a 32 rops has to be 1.5x clock to equal to 48rops at 2/3 clock" and bla bla bla, but that only applies when those ROPs (or any other computing element for that matter) are going to be put into use. Since that is not the case...


16 ROPs no longer cut it for the kind of performance that Barts needs, that is completely true, but at the same time 32 are OVERKILL for Barts, more than sufficient/still overkill for Cypress and enough for Cayman. Why does Barts use 32 ROPs then? This is what you fail to understand, apparently, and what has you completely confused thinking they are necessary... but I have already explained why, 16 ROPs just don't cut it, so 32 were "needed", because AMD is not going to use odd numbers.

And again by all means 32 ROPs are more than enough for the kind of performance that Cayman is going to offer with 1920 SPs, no matter what your distorted perception tells you, like that a 32 ROP/256 bit card is midrange.

 

[yogurt_21]

matching rop's for the midrange and highend doesn't match ati/amd's trend for prior series when comparing xt core to xt core. (doing it only once in the past with the 4770 to 4890 difference)

9600xt 4 pp, 9800xt 8pp (couldn't find rop data so used pixel pipeline)
x700 pro 8rop, x800xt 16 rop (first appearance of the 7 as a midrange, missing in next few)
x1600xt 4 rop (12pp), x1800xt 16 rop
x1650xt 4 rop, x1950xtx 16 rop
2600xt 4 rop, 2900xt 16 rop
3670 4 rop, 3870 16 rop
4670 8 rop, 4870 16 rop
4770 16 rop, 4890 16 rop <- exception (first reappearance of the 7 as a midrange)
5770 16 rop, 5870 32 rop

now here's what we know regardless of what the card is going to be called barts is the juniper replacement tagged for beating the gtx460 1gb.

the 5830 trails the 460 by 12%, the 5850 beats the 460 by 12% for the summary of all resolutions in w1z'a review. http://www.techpowerup.com/reviews/Zotac/GeForce_GTX_460_1_GB/31.html

we also know that the gap shrinks to 10% between the 5830 and the 460 at the highest resolution and the 5850 gains a healthy lead at 23%

with a shader design change, a tmu bump, and clock speed bumps over the 5830 I'm pretty confident that 16 rop's is more than enough to make barts take out the gtx460 1gb.

amd isn't wasteful, if 32 is enough for caymen, 16 is enough for barts.

 

[cheezburger]

More ROPs are only necessary when they ARE necessary. That's what you constantly fail to understand, despite your vast knowledge on GPU architecture... And yes I'm being sarcastic. HD 5870 with 32 ROPs is NOT bottlenecked by its ROP count by any means. In fact, its 32 ROPs clocked at 850 Mhz are nothing but overkill. Nvidia cards? Even more overkill for games (still necessary for GPGPU and CAD/CAM though). You could add 2497 ROPs if that's what you wanted and NO it would not make any difference in gaming performance. You are speaking with a book in your hands " 16rops in real time operation has to be 2x clock to be competitive to 1/2 clock 32rops. a 32 rops has to be 1.5x clock to equal to 48rops at 2/3 clock" and bla bla bla, but that only applies when those ROPs (or any other computing element for that matter) are going to be put into use. Since that is not the case...


16 ROPs no longer cut it for the kind of performance that Barts needs, that is completely true, but at the same time 32 are OVERKILL for Barts, more than sufficient/still overkill for Cypress and enough for Cayman. Why does Barts use 32 ROPs then? This is what you fail to understand, apparently, and what has you completely confused thinking they are necessary... but I have already explained why, 16 ROPs just don't cut it, so 32 were "needed", because AMD is not going to use odd numbers.

And again by all means 32 ROPs are more than enough for the kind of performance that Cayman is going to offer with 1920 SPs, no matter what your distorted perception tells you, like that a 32 ROP/256 bit card is midrange.

32rops are overkill? i expecting it may be overkill because to average people 100fps and 200 fps are no different. that is the reason why is it more then sufficient in some meaning. term 200+fps may seem to be overkill in actual gaming but high end consumer demanding it which still have its own basis. you can say that 200+fps are useless but many people like me just want real time performance. however amd's market strategy has confuse general audience since cypress came out. it's 32 rops only perform on the side with gtx 285 while cost 399 dollar as much but just had the spec of last gen high end. its 32rops are perfom side with gtx 285 and a bit faster due to clock rate butit's main advantage is only in heavy load like higher shader/AA and higher resolution and the cripple directx 11 feature but generally due to the number of rops its fps are pretty much limited to evenly with gtx 285 in maximum frame rate. cypress would have be great if as a mid range card but place it high end with such low spec is just wrong. and cypress is just as good as g104 in spec looking but cost 50% more while die size is a lot smaller than g104. why people have to pay for something that cost so cheap the price so high?

plus putting "32rops&256bit bus vs 48rops&384bit vs 64rops&512bit" bus aside. it is the price that matter/driven the cayman/barts. do you think a smaller >300mm^2 chip is going to sell any price range beyond 300 dollars? only amd dare to do this. nvidia might had such price tape in the past but that was applied to the chip/pcb cost while cypress are has lower quality than gtx285 and even gtx470/480 while sell at the same price range. do you think these high end enthusiast going to buy them because of price tab that look high end while the spec are poorly compare its counterpart? because the sake of overclock? no way. paying the same high end price on the relatively low spec mid range card for just overclocking? that sounded idiotic to me. nvidia had finish the transition of refine its fermi architecture, that is gtx460. of cause barts is the replacement of cypress&older juniper with cheaper cost newer 4D shader/better clocking. and that is all we need to expecting the hd 6xxx to be while it won't be shitty pricing as cypress were. a over number shader of cayman is unnecessary and won't boosting frame rate that much even it had many shader. which i had pointed it few pages ago. rops are necessary when they ARE necessary....yes it is necessary to me because i want 300+fps gaming


well, explain to me how do you going to put 1920sp on 32 rops? no matter how you explain this but it is true that it will be 3RPE and 8x6 config rops and since barts is 2RPE i wonder how this will turn out to be when 3RPE only have 32 raster pipeline..... unless the actual card come out in november i wouldn't take chiphell's slide too serious. first to be 480sp then 800sp then 1280sp and now 960sp.......while fud still shows barts will be the 1280sp...

 

[Benetanegia]

amd isn't wasteful, if 32 is enough for caymen, 16 is enough for barts.

Remember that Cayman is not twice Barts, unlike all your other examples (except HD4770 vs 4890, interesting?). IMO 16 is not enough to ensure good scaling for Barts, but 24 even 20 would probably be enough, but AMD does not do odd numbers, so it must be 32. It's just as easy as this, if you want some performance improvement, ANY performance above HD5830, you need more than 16 ROPs. The HD5870 is not as constrained as the 5830, in fact it is not constrined at all, so there's room above it.

Or look at it this way, HD5870 has around 50% more shader/tmu than HD5830, so if HD5830 is not bottlenecked with 16 rops, that means that HD5870 has enough with 24 (+50%), and hence Cayman with 20% more shader has enough with 32. Necessarily.

 

[Super XP]

Need I say more!

http://img151.imageshack.us/f/nextgenhd6000.jpg/
http://hardforum.com/showpost.php?p=1036227013&postcount=44

 

[entropy13]

@yogurt: 9600xt 4 rop, 9800xt 8 rop

 

[Super XP]

I still think they should stick with ATI Radeon HD 6770 and not HD 6870 as a midrange card.

 

[cheezburger]

Remember that Cayman is not twice Barts, unlike all your other examples (except HD4770 vs 4890, interesting?). IMO 16 is not enough to ensure good scaling for Barts, but 24 even 20 would probably be enough, but AMD does not do odd numbers, so it must be 32. It's just as easy as this, if you want some performance improvement, ANY performance above HD5830, you need more than 16 ROPs. The HD5870 is not as constrained as the 5830, in fact it is not constrined at all, so there's room above it.

Or look at it this way, HD5870 has around 50% more shader/tmu than HD5830, so if HD5830 is not bottlenecked with 16 rops, that means that HD5870 has enough with 24 (+50%), and hence Cayman with 20% more shader has enough with 32. Necessarily.

it's not about 32rops enough of not. remember this, 5830 was cut down version of cypress which can not apply to cayman unless cyaman has 64 rops and 2560sp to cut down to 32rops/1920sp to support your theory on 5870vs5830. but it seems unlikely happen. 5830 is a cripple die and cayman is not. it's just the how gpu work that don't allow any illogical design to be like this happen. unless they decide to make it 24sp per SIMD then original plan of 20sp per SIMD but such speculation is not happening today because of 3RPE and that design only comes with 2RPE/4RPE but cayman is 3/2 of bart that means it has to be 3RPE. needless to say more.

I still think they should stick with ATI Radeon HD 6770 and not HD 6870 as a midrange card.

barts is upper mid range, not mainstream. just like 3850/3870 when came out to stuff the position zone between 2900xt and 2600xt(PS: 2900xt only retire when when 4870x2 came out to replace it). same apply to 9600/8800gt and gtx 460 today. that is nothing new but happen to be successful to hold the market before the real next gen line release or happen when the midrange card had poor performance (ex 8600gt/2600xt) and these card are purposely to place in these empty price range.

matching rop's for the midrange and highend doesn't match ati/amd's trend for prior series when comparing xt core to xt core. (doing it only once in the past with the 4770 to 4890 difference)

9600xt 4 pp, 9800xt 8pp (couldn't find rop data so used pixel pipeline)
x700 pro 8rop, x800xt 16 rop (first appearance of the 7 as a midrange, missing in next few)
x1600xt 4 rop (12pp), x1800xt 16 rop
x1650xt 4 rop, x1950xtx 16 rop
2600xt 4 rop, 2900xt 16 rop
3670 4 rop, 3870 16 rop
4670 8 rop, 4870 16 rop
4770 16 rop, 4890 16 rop <- exception (first reappearance of the 7 as a midrange)
5770 16 rop, 5870 32 rop

now here's what we know regardless of what the card is going to be called barts is the juniper replacement tagged for beating the gtx460 1gb.

the 5830 trails the 460 by 12%, the 5850 beats the 460 by 12% for the summary of all resolutions in w1z'a review. http://www.techpowerup.com/reviews/Zotac/GeForce_GTX_460_1_GB/31.html

we also know that the gap shrinks to 10% between the 5830 and the 460 at the highest resolution and the 5850 gains a healthy lead at 23%

with a shader design change, a tmu bump, and clock speed bumps over the 5830 I'm pretty confident that 16 rop's is more than enough to make barts take out the gtx460 1gb.

amd isn't wasteful, if 32 is enough for caymen, 16 is enough for barts.

16rop can even take off the 48 rops like gtx 480, if clock ridiculous enough. but the product of heat are much greater than larger spec with lower clockrate. it already happen that 5830 ate more power and ran hoter than 5870/5850.

 

[Benetanegia]

it's not about 32rops enough of not. remember this, 5830 was cut down version of cypress which can not apply to cayman unless cyaman has 64 rops and 2560sp to cut down to 32rops/1920sp to support your case on 5870vs5830. but it seems unlikely happen. 5830 is a cripple die and cayman is not. it's just the how gpu work don't allow this happen. unless they decide to make it 24sp per SIMD then original plan of 20sp per SIMD but such speculation is not happening today because of 3RPE and that design only comes with 2RPE/4RPE but cayman is 3/2 of bart that means it has to be 3RPE. needless to say more.

Ok you don't have a clue of what your talking about. SIMDs and ROPs have nothing, read NOTHING to do with each other. They are a completely separete thing in every single GPU released to date. They are in a different part of the silicon and they communicate through a crossbar type bus. Nothing prevents anyone from putting 28000 ROPs and 12 SIMDs, except the fact that it would be pointless. Seriously at this point, I'd even bet my car if I had one that you don't even know what a ROP is.

It is true that there's been speculation of ROPs being included in the RPE in NI, but that does not mean it's going to be like that. Like I said, it makes more sense to put them outside of them and closer to the memory controler, so that your chip can be harvested in many different ways. Just like Nvidia does.

Cypress as in HD5830. Some SIMDs disabled, some ROPs disabled, at will, no relation between them at all, between how many have been disabled.

Fermi.

NI is most probably similar to Fermi in that GPC==RPE. But as you can see ROPs are independent (in blue, next to L2 cache in the middle, not the ones inside the SMs).

 

[Tatty_Two]

(PS: 2900xt only retire when when 4870x2 came out to replace it).

WTF? The 3870 (only marginally by 1%), the 3870x2, the 4850 and the 4870 were all faster than the 2900XT BEFORE the 4870x2 came out.

 

[cheezburger]

Ok you don't have a clue of what your talking about. SIMDs and ROPs have nothing, read NOTHING to do with each other. They are a completely separete thing in every single GPU released to date. They are in a different part of the silicon and they communicate through a crossbar type bus. Nothing prevents anyone from putting 28000 ROPs and 12 SIMDs, except the fact that it would be pointless. Seriously at this point, I'd even bet my car if I had one that you don't even know what a ROP is.

It is true that there's been speculation of ROPs being included in the RPE in NI, but that does not mean it's going to be like that. Like I said, it makes more sense to put them outside of them and closer to the memory controler, so that your chip can be harvested in many different ways. Just like Nvidia does.

http://tpucdn.com/reviews/ASUS/EAH_5830_DirectCu/images/arch.jpg

Cypress as in HD5830. Some SIMDs disabled, some ROPs disabled, at will, no relation between them at all, between how many have been disabled.

http://tpucdn.com/reviews/NVIDIA/GeForce_GTX_480_Fermi/images/arch.jpg

Fermi.

NI is most probably similar to Fermi in that GPC==RPE. But as you can see ROPs are independent (in blue, next to L2 cache in the middle, not the ones inside the SMs).

this is so called the uncore architecture that being used exclusively on evergreen line, and which it works only with in the line. it is correct that in uncore confiuration the rop had really nothing to do with SIMD like it was on previous 4xxx. but it's strange that cypress had to disable rop while disable the SIMD cluster...however NI's RPE is similar to fermi's modules that amd wrap a gpu "core" as a single engine which each block already act as single gpu processor that everything had included. if you look at barts's core it's more like 2 juniper together except in new 4D shader so in NI you can't disable rops without disable the whole RPE which it is in debating right now. just imagine cayman as a tri core gpu then you're realize you can't make 32rop on it because it'd be 2 module instead of 3 module. but since the die map hasn't out yet so you may be right that NI will just adapting the uncore design from evergreen than what it suppose to be renderer processing engine. that is the only thing i can think of the possibility to have 1920sp and 32rops. of cause i know raster pipe that imply to gpu operation the higher pixel the faster frame rate& more benefit to shader and less bottleneck . but i'd just have to skip these raster pipe vs shader because it will be another endless debate and turns out to be max frame rate speed vs heavy load performance.....

WTF? The 3870 (only marginally by 1%), the 3870x2, the 4850 and the 4870 were all faster than the 2900XT BEFORE the 4870x2 came out.

sorry, i forgot 3870x2....my bad. 3870x2 was the replacement of 2900xt indeed

 

[yogurt_21]

Remember that Cayman is not twice Barts, unlike all your other examples (except HD4770 vs 4890, interesting?). IMO 16 is not enough to ensure good scaling for Barts, but 24 even 20 would probably be enough, but AMD does not do odd numbers, so it must be 32. It's just as easy as this, if you want some performance improvement, ANY performance above HD5830, you need more than 16 ROPs. The HD5870 is not as constrained as the 5830, in fact it is not constrined at all, so there's room above it.

Or look at it this way, HD5870 has around 50% more shader/tmu than HD5830, so if HD5830 is not bottlenecked with 16 rops, that means that HD5870 has enough with 24 (+50%), and hence Cayman with 20% more shader has enough with 32. Necessarily.

interesting, if you're right barts xt may just be nicking at the heels of caymen pro which is usally when you find a stellar deal on a gpu with an awesome price/perf while making caymen pro a horrible deal at the same time.

 

[Benetanegia]

however NI's RPE is similar to fermi's modules that amd wrap a gpu "core" as a single engine which each block already act as single gpu processor that everything had included. if you at barts's core it's more like 2 juniper together except in new 4D shader so in NI you can't disable rops without disable the whole RPE which it is in debating right now. just imagine cayman as a tri core gpu then you're realize you can't make 32rop on it because it'd be 2 module instead of 3 module. but since the die map hasn't out yet so you may be right that NI will just adapting the uncore design from evergreen than what it suppose to be renderer processing engine. that is the only thing i can think of the possibility to have 1920sp and 32rops. of cause i know raster pipe that imply to gpu operation the higher pixel the faster frame rate& more benefit to shader and less bottleneck . but i'd just have to skip these raster pipe vs shader because it will be and endless detabe and turns out to be max frame rate speed vs heavy load performance.....

Prove that you can't. NI is most probably like Fermi where front end and computing units are inside the same clusters, but rops are outside of them, because that's what makes most sense. Having all your architecture (ranging from memory bus, rop count, L2 cache) depend on how many clusters you can enable is stupid. Not to mention ackward, tell me where in NI is the L2 cache? It's is split between RPEs too? How do MCs communicate properly with the rest of the chip if there's many different L2? Is it outside the RPEs and is the same for all the NI chips, while only RPE number changes?

If absolutely everything is split between RPEs, performance is going to be horrible, because NI would be basically 3 small GPUs glued together and it would be subjected to 3-way crossfire scaling == bad.

Nah, I tell you that NI is going to be identical to Fermi, or in any case it would have even less units integrated into RPEs, not more. ROPs, L2 and MCs are going to be independent that's for sure, and seeing as how geometry performance is not much better than Cypress according to leaks, I bet that vertex setup/raster is still one single unit in the front end sitting just after the threading engine.