AMD: R9xx Speculation

[LordEC911]

It makes sense. ATi has never released 2 256bit GPUs for the same product line at the same process. It implies, that Bart was initially 128bit part, probably a direct replacement of Juniper, around 170mm² at 32nm. 170mm² GPU enlarged (what's the opposite of shrink? ) to 40nm would be 265mm² large. Maybe slightly larger due to the change of memory bus or slightly smaller if the architectural changes compensated it.

I'm quite curious if this GPU will be competitive with 100mm² bigger GF104...

Seeing as how everything is pointing towards Barts being the drop-in replacement for Cypress cards, I would imagine that performance is similar.

If my assumptions are correct, the "new" 5850 should perform as a 5870. Probably just Clocks&Optimizations.

Is Barts the "HD5850 replacement that performs like a 5870" that neliz was talking about back in June?

a 6700 would be faster than GTX480 and 5870. So a 6700 would have around 3TFlop @40nm and a 6800 would have 6TFlop at 28nm but not show up until somewhere in 2011.
I think "increased efficiency" is the keyword of the coming AMD designs.

Also, since he let slip the "6700 is about 3TFlops" in April, if the SP count is correct at 1920 that implies clocks of at least ~800mhz. Hmmm....
And since 28nm kept getting delayed, the "6TFlop 6800" became a dual GPU part?

Let the new silly season begin!

 

[caveman-jim]

It has inflated pricing compared to the previous generation product which can be purchased for a similar price.

Unless you're ferro's dupe account I would prefer to have a reponse from ferro; besides which, the previous gen product is the same price and the current gen product has an inflated price? Did you mean the previous generation higher performance model (that lacks features present in the current gen) is now a similar price?

 

[TKK]

Where did you find these in Cat 10.8's? the .inf's for sure don't have Cayman nor Barts nor Caicos - or NI anything for that matter
(edit: vista/win7 x64 .inf's anyway)

it's in a file called atiicdxx.dat.

223,CAYMAN GL XT (6700),NI CAYMAN
224,CAYMAN GL XT (6701),NI CAYMAN
225,CAYMAN GL XT (6702),NI CAYMAN
226,CAYMAN GL XT (6703),NI CAYMAN
227,CAYMAN GL PRO (6704),NI CAYMAN
228,CAYMAN GL PRO (6705),NI CAYMAN
229,CAYMAN GL (6706),NI CAYMAN
230,CAYMAN GL LE (6707),NI CAYMAN
231,CAYMAN GL (6708),NI CAYMAN
232,CAYMAN GL (6709),NI CAYMAN
233,CAYMAN XT (6718),NI CAYMAN
234,CAYMAN PRO (6719),NI CAYMAN
235,ANTILLES PRO (671C),NI CAYMAN
236,ANTILLES XT (671D),NI CAYMAN
237,BLACKCOMB XT/PRO (6720),NI BLACKCOMB
238,BLACKCOMB LP (6721),NI BLACKCOMB
239,BLACKCOMB XT/PRO Gemini (6724),NI BLACKCOMB
240,BLACKCOMB LP Gemini (6725),NI BLACKCOMB
241,BARTS GL XT (6728),NI BARTS
242,BARTS GL PRO (6729),NI BARTS
243,BARTS XT (6738),NI BARTS
244,BARTS PRO (6739),NI BARTS
245,WHISTLER XT (6740),NI WHISTLER
246,WHISTLER PRO/LP (6741),NI WHISTLER
247,WHISTLER XT/PRO Gemini (6744),NI WHISTLER
248,WHISTLER LP Gemini (6745),NI WHISTLER
249,ONEGA (6750),NI TURKS
250,TURKS XT (6758),NI TURKS
251,TURKS PRO (6759),NI TURKS
252,SEYMOUR XT/PRO (6760),NI SEYMOUR
253,SEYMOUR LP (6761),NI SEYMOUR
254,SEYMOUR XT/PRO Gemini (6764),NI SEYMOUR
255,SEYMOUR LP Gemini (6765),NI SEYMOUR
256,CAICOS GL PRO (6768),NI CAICOS
257,CASPIAN PRO (6770),NI CAICOS
258,CAICOS PRO (6779),NI CAICOS

 

[neliz]

Also, since he let slip the "6700 is about 3TFlops" in April, if the SP count is correct at 1920 that implies clocks of at least ~800mhz. Hmmm....
And since 28nm kept getting delayed, the "6TFlop 6800" became a dual GPU part?

Back then I was Foolish and mal-informed.

I don't think anyone is realistically expecting anything remotely spectacular for the coming quarter anymore. better, yes, spectacular? No.

 

[rpg.314]

I don't think anyone is realistically expecting anything remotely spectacular for the coming quarter anymore. better, yes, spectacular? No.

How much better is it?

 

[neliz]

How much better is it?

That actually beats me, just what you guys heard already, UVD3/Eyefinity+ and I think I also already wrote about the new top-model being faster than a full GF100.. though with AMD that might just mean special driver X under circumstance(resolution) Y and that we won't see that translated to real world Fermi-besting around the block.

 

[trinibwoy]

That actually beats me, just what you guys heard already, UVD3/Eyefinity+

It can't be just Eyefinity and UVD3 though, NI/SI are supposed to shake things up a lot more than that. Unless the whole SI=40nm NI thing is a smokescreen.

 

[no-X]

In the case NI were a half-year refresh planned for Easter 2010 (delayed by the 32nm->40nm redesigning), I wouldn't expect significant architectural changes. What was the source of the 5D -> 4D rumour? Wouldn't it be to significant change for a six-mounths refresh?

I'm starting to believe a simplier theory. 32nm manufacturing would allow +50% tranzistors staying at the same die-size. E.g. RV840 replacement could have 1200 SPs + 60 TMUs. The problem: RV8xx building block consists of 800 SPs + 40 TMUs. Solution: desing smaller builing block of 640 SPs + 32 TMUs. That would solve even another problem - efficiency. The two SIMDs had very litle impact on gaming performance and consumed die area.

Bart would be based on two blocks - 1280 SPs + 64 TMUs. It was originally designed as a 128bit part, but larger die due to the 40nm process would allow 256bit bus. I'd expect 16 ROPs...

Cayman would be based on three blocks - 1920 SPs + 96 TMUs, 32 ROPs + 256bit bus with ~6GHz GDDR5.

 

[Mindfury]

In the case NI were a half-year refresh planned for Easter 2010 (delayed by the 32nm->40nm redesigning)

NI was planned for 2011 as a full lineup,not a refresh.

 

[Kaotik]

NI was planned for 2011 as a full lineup,not a refresh.

Full lineup, yes, but probably 2010, just like Evergreen came first out in 2009 but is listed for 2010 in the slide.

My guesstimate on the names:
Antilles = 6970 (=2x Cayman)
Cayman = 68xx
Barts = 67xx
Caicos = 66xx/65xx
Turks = 64xx

(the last 2 based on Caicos- and Turks -islands size)

 

[no-X]

NI was planned for 2011 as a full lineup,not a refresh.

That's platform roadmap, not product roadmap... Please notice, that platform roadmap listed R7xx family for ~ Q1-Q2/09, while it was in fact released during Q2/08:

 

[Gipsel]

Is Barts the "HD5850 replacement that performs like a 5870" that neliz was talking about back in June?

Also, since he let slip the "6700 is about 3TFlops" in April, if the SP count is correct at 1920 that implies clocks of at least ~800mhz. Hmmm...

Back then I was Foolish and mal-informed.

You should just pretend, that you was talking about the ASIC ID of Cayman, which is 6700

 

[neliz]

You should just pretend, that you was talking about the ASIC ID of Cayman, which is 6700

Barts, Turcs, Smurfs.. I want RV9xx back!

 

[ferro]

Unless you're ferro's dupe account I would prefer to have a reponse from ferro

Normally competition will bring prices down. In this case prices have only gone up (see charts from my previous post), which is unusual. I blamed the lack of competition, but like others have said, factors like capacity constraints, currency exchange rates, and component costs may also play a role.

 

[Bouncing Zabaglione Bros.]

Normally competition will bring prices down. In this case prices have only gone up (see charts from my previous post), which is unusual. I blamed the lack of competition, but like others have said, factors like capacity constraints, currency exchange rates, and component costs may also play a role.

It's pretty unbelievable that after the first month, prices went up and have stayed there for the rest of the life of the product. It looks like it's going to get EOL'ed before prices get down to their launch prices, let alone below that.

 

[Mindfury]

What was the source of the 5D -> 4D rumour?

http://www.chiphell.com/forum-viewthread-tid-108125-page-1-authorid-2.html

1920SP,最终版R600架构,但是不再是5D,而是4D,其它的大家自己计算.
nApoleon 发表于 2010-7-15 18:26

 

[Jawed]

5D->4D rumour goes back further, e.g.:

http://forum.beyond3d.com/showthread.php?p=1416950#post1416950

Not sure if it came up before that. I have a feeling it does.

 

[3dilettante]

I think the rumors have started to settle more on the T unit consuming one of the thin lanes, unless a new set of rumors has come along since then.

 

[Jawed]

I think the rumors have started to settle more on the T unit consuming one of the thin lanes, unless a new set of rumors has come along since then.

I don't think there's any further insight and we still have the ASIC_ALU_REORDER thing to ponder.

 

[Jawed]

Another 5D->4D reference

http://forum.beyond3d.com/showpost.php?p=1423337&postcount=672

It seems to me that Cypress becomes Barts, 1280 ALUs, 80 TMUs, 32 ROPs, 256-bit. The ALU reduction saves area, let's say 50mm² for the sake of argument, but other new stuff makes it bigger again.

This should have been a ~200mm² chip on 32nm though...