New rumor about ATI Southern Islands

[Sn0wm@n]

cant wait for benchy to flow through

[Chumbucket843]

This was pretty much what I was hoping for from AMD, not on the nvidia side though that's just going to mean higher prices and less real choice. I don't really see anything in nvidia's strategy that allows for a fight back until 28nm.

Both sides were expecting to have a chance at 32nm TSMC and GF dump it so both sides are stuck on 40nm, ATI/AMD still have more room to wriggle with their architecture.

Anyone else thinking nvidia is throwing more effort into their 28nm for a comeback?

i think they will have a decent 40nm gpu sometime H1 2011. i think they might announce something at GTC which is in a few weeks. this would make sense because it complies with the rumors that nvidia is ordering less 40nm lots.

i think all of the 32nm node rumors are charlie's BS. all of that is made up so he can create elaborate stories to explain why ATi doesnt have a refresh out yet. tsmc did the same thing with 45nm->40nm and publicly announced they would do the same thing with 22nm->20nm. see a pattern? why would 32nm->28nm be any different? it seems tsmc goes straight for the half node, possibly because that is part of their technique for designing processes or perhaps their nodes mature fast enough that they might as well delay it a few months and add an optical shrink. furthermore, i dont see any technical reason that 32nm would be cancelled, especially if they can make a 28nm node.

[LordEC911]

So the die is approx. 68.5 mm² (9.5*7.2). Any estimates on transistor count? If the transistor density is anything like in Cypress, the transistor count would be about 440.9M ((2150 M/334 mm²)*68.5), however compared to similar sized Cedar(59 mm²), the transistor count would be approx. 339 M((292 M/59 mm²)*68.5).

If the only change from Cedar is UVD 3.0 + 3D Bluray playback(Which are more or less just additional logic) then the transistor count of ~340 M seems most accurate.

I believe the die size is ~66.5mm2.

[Calmatory]

I believe the die size is ~66.5mm2.

With which core width & height measurements?

[Iconyu]

i think they will have a decent 40nm gpu sometime H1 2011. i think they might announce something at GTC which is in a few weeks. this would make sense because it complies with the rumors that nvidia is ordering less 40nm lots.

i think all of the 32nm node rumors are charlie's BS. all of that is made up so he can create elaborate stories to explain why ATi doesnt have a refresh out yet. tsmc did the same thing with 45nm->40nm and publicly announced they would do the same thing with 22nm->20nm. see a pattern? why would 32nm->28nm be any different? it seems tsmc goes straight for the half node, possibly because that is part of their technique for designing processes or perhaps their nodes mature fast enough that they might as well delay it a few months and add an optical shrink. furthermore, i dont see any technical reason that 32nm would be cancelled, especially if they can make a 28nm node.

Thinking back on how both of them have worked since the DX8 days you're probably right about 32nm, I guess it was just a dream for us that didn't jump HD5870 on launch day.

nvidia do tend to have a high end refresh between processes, but this time around I'm not sure how far they can go.

[Ejizz]

Anyone else thinking nvidia is throwing more effort into their 28nm for a comeback?

Maybe, I think they need to sort out their performance per mm2 before they stand a chance at a comeback.
I think they need to make some major architecture changes, else they will metamorphosize into the new Matrox or worse 3DFX...

[-Sweeper_]

Anyone else thinking nvidia is throwing more effort into their 28nm for a comeback?

me..
they'll probably try to bring a ''RV770'' on 28nm...

[FischOderAal]

Well. If all what our Mr.Salt says is true I hope AMD has some surprises for SI as well and won't just deliver a "simple" shrink. But seeing what they did lately, I'm a tiny tad optimistic.

Maybe, I think they need to sort out their performance per mm2 before they stand a chance at a comeback.
I think they need to make some major architecture changes, else they will metamorphosize into the new Matrox or worse 3DFX...

I don't know. Maybe Big Green should have two teams working on different chips, one chip for HPC/GPGPU and one for gaming only. Thing is we only see the gaming side of the end.

me..
they'll probably try to bring a ''RV770'' on 28nm...

28 nm might be the node where the architecture shows it's muscles RV770 showed R600's architecture wasn't so bad after all (imho RV670 already gave a little hint at that), just that they wanted too much too early (big chip etc). Kind of what NVIDIA did with GF100.

[Tenknics]

i think they will have a decent 40nm gpu sometime H1 2011. i think they might announce something at GTC which is in a few weeks. this would make sense because it complies with the rumors that nvidia is ordering less 40nm lots.

i think all of the 32nm node rumors are charlie's BS. all of that is made up so he can create elaborate stories to explain why ATi doesnt have a refresh out yet. tsmc did the same thing with 45nm->40nm and publicly announced they would do the same thing with 22nm->20nm. see a pattern? why would 32nm->28nm be any different? it seems tsmc goes straight for the half node, possibly because that is part of their technique for designing processes or perhaps their nodes mature fast enough that they might as well delay it a few months and add an optical shrink. furthermore, i dont see any technical reason that 32nm would be cancelled, especially if they can make a 28nm node.

yea Charlie made up tsmc cancelling its 32nm...



Kinda like how TSMC's 40nm problems and yield issues were made up too......oh wait...

[Tha Last Meal]

Originally Posted by informal View Post
Latest article by C. Demerjian, consume with salt of course:

LOL. C. Demerjian Again

[Chumbucket843]

Thinking back on how both of them have worked since the DX8 days you're probably right about 32nm, I guess it was just a dream for us that didn't jump HD5870 on launch day.

nvidia do tend to have a high end refresh between processes, but this time around I'm not sure how far they can go.

well i see it as two options:
1. full throttle on 28nm "fermi2"[huge rewards, but risky, and almost exactly what they did on 40nm, also 28nm could be a year away!]
2. silicon spin on 40nm with solid enhancements[decent perf & manufacturing cost gains, lots of manual effort but a mature process is always a safe play]

nvidia has a lot more information to base their decisions off of than we do. i'm going for #2.

yea Charlie made up tsmc cancelling its 32nm...



Kinda like how TSMC's 40nm problems and yield issues were made up too......oh wait...

or kind of like your excellent sarcasm. aside from that you misinterpreted my post. by 32nm node rumors i meant all of the godly things ATi would have done on 32nm if it werent for tsmc's incompetence. it's really poor journalism. ATi, the underdog, is going against the system which is holding them down. (i.e. nvidia and tsmc.)

tsmc doesnt have yield problems. it took them a year to ramp but after that the process was fine. the problem is they undershot capacity by a large margin which isnt as unusual as one might think. iirc in some thread i calculated defect density to be .115defects/cm2. that's pretty good.

also fermi's yields do not mean anything at all for other chips. they could have poorly designed the chip. when you have billions of transistors things get complicated. yields are no exception.

[Jowy Atreides]

i think this is what people are talking about.


and no that is not everything they could possibly change. take a computer architecture course or something of the sort.

Take a literary course if you can't read.

Quoting a sentence and leaving out three contextual words.

"that"
"affects"
"performance"

Re-branding unreleased parts isn't in any fashion redundant.

[Iconyu]

Reading other rumours I don't think the 5770 will be renamed to the 6770, I think the 6770 will have the same architectural changes the 6870 will enjoy on a smaller scale and will become the 6950 in x2 form, while the 6970x2 will be the full blown 6870 dual we'd be expecting from AMD.

I do think the 5770 might be tweaked and re-branded but as a much lower model.

Going back to this:

roadmap from chiphell:

These would have to be price brackets, and the only way it would be worth making the 6950 sell for less than the 5970 would be if they used smaller chips to start with. A chopped chip wouldn't use enough less current for it to be worth the effort.

It's also possible to make with the switch from 5 to 4 to make a smaller part with a similar number of shaders and higher clocks that would give impressive performance on the right benchmarks. If you only picked benchmarks that were heavily biased to complex shaders you'd get the results we've seen posted, however on benchmarks with simple shader bias there would be next to no change at all.

The point that is making me think this was is the rumour that AMD is launching the 6770 first and not the 6870, not even nvidia would launch a new generation of GPU's with a re-branded unit. I guess we'll find out if AMD starts dumping the price on the 5770 or leaves the price high in order to protect stock of the silicon they plan to re-brand.

[ajaidev]

I hope they dont rebrand i mean 6990 = dual gpu and 6970 = single gpu does not compute that is if 6990 is made of 2 cayman chips bu if its made up of 2 Barts then i think its ok.

[Mechanical Man]

markkinoille lokakuussa Barts Pro- ja XT

"first to market are Barts Pro and XT in october"

[flopper]

"first to market are Barts Pro and XT in october"

Midsegment then.
any idea whn nda is gone?

[LordEC911]

Midsegment then.
any idea whn nda is gone?

Release to the most important market where you are the weakest ASAP, slap the GTX460 around.

Sounds like there is an event on Oct 12th but the release isn't until the 26th(?).

[RPGWiZaRD]

Release to the most important market where you are the weakest ASAP, slap the GTX460 around.

According to sweclockers that's correct. http://translate.google.com/translat...el-med-cypress

I wouldn't put too much faith into their source as they keep calling it Northern islands but they expect Barts to be released first that is 256bit and gonna easily beat GTX 460 at price 2000 ~ 2500 SEK which should be around 230~$250 (tax differences and price variances taken into account).

[Iconyu]

Release to the most important market where you are the weakest ASAP, slap the GTX460 around.

Sounds like there is an event on Oct 12th but the release isn't until the 26th(?).

I'll just float this idea, I'm thinking if those benchmarks were real, they can only be from the Barts XT, it's the only chip that could match the GPUID profile we're seeing. If you look at the difference between the 3870 and the 4870 it would be possible for ATI to do this again. Especially if they planned to take their time with 6870.

Basically ATI isn't releasing a 'mid range' GPU their releasing a new top end and giving it a mid range name. It'll be priced as high midrange but it'll beat the 480, it'll be a huge psychological hit. Especially as nvidia's midrange chip is bigger than ATI current high end chip.

Sure it'll decimate their top end prices but they can just release the aforementioned 6950. They'll keep the 6870 name back for either a push into a big core in 2011 or 28nm depending on how nvidia react.

If it works as planned, ATI would have the DX11 market sewn up by the time nvidia can release on 28nm and still make a profit as they've already proved they don't have to drop prices in comparison to nvidia.

This is all theory, but the rumours I've read up until this point support it.

[mindfury]

I'm not sure if it's real,I can't find the source of this pic.

[MrMojoZ]

The point that is making me think this was is the rumour that AMD is launching the 6770 first and not the 6870, not even nvidia would launch a new generation of GPU's with a re-branded unit.

Hmmm, Geforce 9x00 series a new generation?

[Pantsu]

Most likely Barts XT will be close to 5850 and offer better price/performance than GTX 460. I'd bet in that case it'll also replace 5850 quickly since they'd be too close and people would be more inclined to buy a 6770. I wouldn't be surprised if 5850 price dropped down soon.

[Solus Corvus]

I still find it hard to believe that whipping up a frankenstein is easier then modifying an existing generation or un-shrinking a future one.

Whatever the case is, it sounds like AMD did the exact opposite of "not have a backup plan".

If they are going to rename some old cards or change the naming scheme, I can't really support that. It makes sense the way it is now. Don't try and confuse customers AMD.

[madcho]

I'm not sure if it's real,I can't find the source of this pic.

i think this is real, if you check the other picture, this side is the same than the other one, and the 8 pins is left too.

[Manicdan]

seems like a small PCB for a 6+8pin

if its really consuming 250+ watts, id expect the cooler would have to be much bigger, or they are giving you room to OC

size of the heatsink says alot more than the number of power pins