AMD: R9xx Speculation

Please oh god, do not make Dave Baumann start renaming cards.. please oh god hope he does not get bitten by the jensen bug

 

But why would AMD rename Juniper when they have Turks and Caicos planned? It doesn't make much sense, you rename parts when you don't have anything else, or when you can't supply enough new parts.

 

This makes about as much sense as fake Unigine benchmarks.

 

You're taking some sour note, sprinkling it onto some rumour and then saying that because it tastes bad, ATI as we know it are domed.

 

Doomed! DOOMED!

 

Dave Baumann won't allow it.

 

That, and the fact that even the 5770 is hard to come by, why rebrand a product that already sells like hot-cakes?

 

Sorry, but I think you guys really divorced the core part of that ati-forum news from its context:

(1) Basically, their sources only told them that Barts will trade under the name of the HD68xx series - a move which I (and some of you guys before me, obviously) already speculated about further above in our discussion.

(2) How can they "rebrand" a product which hasn't even been officially announced yet? Think about it.

(3) That entire part of the article discussing a putative HD5xxx series rebrand is pure misinformation. The only thing they actually "heard" was that Barts (which they originally ASSUMED to go under the name of the HD67xx family) will actually come to life as 68xxx family. From that, they INFERRED that AMD MIGHT be going to rebrand their entire line-up (which I think is total bs).

Now:

If that chart is to be trusted, doesn't it support my earlier speculation that:

(a) Barts (originally 32nm mid-range part) was bloated (after the 32nm node had been cancelled) to become the new 40nm high-end chip (i.e. the heart of the HD68xx cards).

(b) 2xBarts will team up for a HD6990 card (putatively about 70-80% faster due to cross-fire scaling)

(c) Cayman will be a 28nm part (shrunk from its original 32nm dimensions). It will launch as soon as mass production on the 28nm process is ready, maybe late H1 2011. Its performance will clock in about 60% over Barts/HD6870 (Cayman basically will be "Barts times 1.5" PLUS clock bump; 1280 -> 1920SPs etc., you get the image). It will perform some 10ish% under 2xBarts (HD6990) - and will COHERENTLY trade under the name of the 6970 series.

How would that kind of naming scheme be confusing with respect to the relative performance levels of the final cards?

 

http://forum.beyond3d.com/showpost.php?p=1466338&postcount=1434

 

It may be true, but that means that Bart is faster than Cypress, and it's also smaller.

 

@Mindfury

Yeah, basically a good point.

But I still don't think that whatever was supposed to be the high-end chip of the "original" 32nm HD6xxx series will come to life on the 40nm process.

And I still don't think AMD went to the pains to do a complete "half-step" redesign of the original NI architecture instead of aiming for (probably way more time-efficient) 40nm/28nm "adaptations" (shrinks/bloats plus some small optimization) of the originally planned 32nm chips.

Maybe (while adjusting their designs for the new process nodes) they just decided to keep the codenames bound to the performance levels of the final cards to be released with them, i.e. 32nm Barts became 40nm Cayman; 40nm Cayman became 28nm "something-else"?

I don't know, but maybe - with all that hence-and-forth caused by the cancellation of the 32nm process - we'd be better off not attaching too much weight to names.

It all hinges on the question of what happened to the original 32nm designs. If they really kept the basic designs (like I suggest) - with everything below the original high-end chip (call-it-what-you-like-but-too-big-to-be-feasible-on-a-larger-process-node) being ported to 40nm - I don't see how Cayman32nm-whatever-its-called-now could make a "sweet-spot-compatible" appearance on 40nm (let alone within an x2 card). Cayman32nm-whatever-its-called-now would HAVE to be scheduled for 28nm production - and Barts32nm-whatever-its-called-now would HAVE to step in with respect to the "new" 68xx high-end cards.

If AMD actally DID a complete redesign of their 32nm chips (basically trashing the entire 32nm design range in favour of an "only-slightly-faster-but-at-least-feature-driven" half step, all my speculation is utterly invalid, of course.

Keep in mind: I'm not speculating from the point of view of an engineer - and I'm not saying I get all the techinal details involved here. But I'm speculating from the point of view of a business economist - and as such, I don't see any option as feasibly as:

(a) "devide our 32nm Northern-Islands designs in two groups according to die size"

(b) "bring the smaller designs to the 40nm node (keep the NI name), bring the bigger designs to the 28nm node (call them something similar but yet different, like Souther Islands)

(c) "release the 40nm parts as soon as they're ready - and fit in the 28nm parts as a healthy refresh as soon as you can"

 

Bart could be comparable to Cypress or slightly slower and despite it they could call it HD6800. Remember the R520-R580-RV570 (RV570 was slower than R580 and despite it it was part of the X1950 family)... Bart could be used for HD6830 and 6850 and Cayman for 6870 and 6890 (or something like that)

Are you implying, that they released driver supporting a GPU, which will be released 3Q later?

 

Given their recent GPU history, my basic assumption would have been that AMD designed their orignally-planned 32nm-high-midrange GPU (whatever they call(ed) it) with a performance aim at least on-par with their "last" high-end-GPU (Cypress), while keeping its die size below 200mm2.

If I'm doing my maths right, a (worst-case!) 200mm2 orignally-planned 32nm-high-midrange GPU would roughly result in a 320mm2 (312.5mm + some headroom for adjustments and performance tweaks allowing for the highest clocks possible) 40nm-new-high-end-GPU. Maybe it could even be less than that, given they know the 40nm process very well by now and might have decided to remove some of their "we-don't know-32nm-yet-safety-nets".

So, as far as die size is concerned, an "orignally-planned-32nm-high-midrange-GPU-ported-to-40nm" would safely clock in some mm2 below Cypress.

As far as power is concerned, though, I'd expect a HD6870 based on such a chip to consume a bit more than the current HD5870. It would be hard enough to reach the clock-speed they aimed for @32nm with the chip ported to 40nm - and they probably would want to even increase clocks a bit in order to make sure the final HD6870 card's performance sits some 20-30% over HD5870.

10-layer high-performance PCB anyone?

====

Baseline: Take your originally-planned-32nm-high-midrange-GPU, port it to 40nm (preferably implementing some kind of 4890esque clock-speed-optimizations on the way) - and you probably get a nice 40nm GPU which renders any alternative investments in a new "half-step" architecture on the same node invalid.

Again, that's the econimist speaking. Feel free to shred my speculation with technical insight

 

No, I actually recognized that point and tried to come up with a (possibly naive?) explanation:

I'd expect that AMDs internal codenames are more techincal either way?

BTW: Why can't I edit my posts on this forum? It really pains me to read my own typos ...

 

That means you are not using Firefox or you're posting more than you're thinking!

Or you can wait until you have 50 posts.

 

Remember, Cypress was supposed to be substantially larger one year ago, i.e. there was a chip with more features and/or performance than Cypress which was intended for 40nm.

 

This.

It is pretty well-known that Antilles Pro and XT is very much based on Cayman at this point, not Barts.

 

Be that as it may, I still got a feeling that - spec-wise (!!) - today's Cayman originally was 32nm's Barts ...

What about this train of thought: When the chip which used to be called Cayman on 32nm was pushed back for 28nm production, it basically became a Southern Islands (28nm) part, hence losing it's original Northern Islands (now 40nm) codename.

So "Cayman" basically became "vacant" as a NI codename - and was rightfully applied to the "old" 32nm high-midrange, but new 40nm high-end part: 32nm's Barts alias 40nm's new Cayman

Well, I'm hard at work trying - I just hope I won't be banned for spamming before I reach my 10th post

 

It is still in the Northern Islands family in the Mac OS X kext, which at this point appearing in the newest beta would just be plain weird.

I don't think they shuffle codenames like that but we have seen more odd things happen in the past.

This is silly season after all.

 

Well, with 32nm's Barts becoming 40nm's Cayman, 32nm's Turks would move up to become 40nm's Barts, of course ... xD

Well, they either

(a) re-sorted the codenames of their original 32nm designs for the 40nm adaptions in the way I suggested

(b) sticked to the codenames - and are about trump Nvidia's monolithic-mania with a fully-blown 32nm-to-40nm Cayman chip that was designed to hit the "sweet-spot" on 32nm - and now sits at around 450mm2 @ 40nm

(c) sticked to the codenames - and castrated the original 32nm Cayman design to make the final chip < 400mm2 @ 40nm

(d) really went for an extensive, unnecessary dev-time-intensive redesign to fill the gap created by the cancellation of TSMC's 32nm process. In that case - what were the codenames used for the originally planned NI chips @32nm?