Fusion die-shot - 2009 Analyst Day

so "x4 Gen 2 + DP" would mean four PCIe lanes + displayport..

 

http://www.computerbase.de/news/hardware/prozessoren/amd/2010/september/amds-fusion-apu-auf-der-ifa-im-einsatz/

 

Seems to me that previous "die-shots" should fool competition. This one is significantly different...

 

The previous one(s) were more-or-less-PhenomII-based Llano, too, not Ontario, which is Bobcat-based

 

Well 9W TDP sounds ok as a dual core atom is 8.5W. I expect Ontario to have similar clocks as well so it should beat Atom in CPU performance and absolutely thrash it in GPU performance. Desktop Zacate should be > 2 Ghz.

Things will get even more interesting when the 28 nm shrink comes out(maybe 18 months later, Q2 2012 sounds reasonable)

 

Zacate, Cedar and Mobile Phenom II Quadcore

http://www.hardware-infos.com/img/startseite/klein/img_3957.jpg

From this report:

http://www.hardware-infos.com/news.php?news=3681

 

They say Zacate is a bit smaller than cedar. Dunno looks about same die size to me. In any case, something along 60mm² would be quite good - pinetrail atom is 66mm² for single and 87mm² for dual core. Now of course this is a different manufacturing process (40nm vs 45nm for starters) but still for something that should be faster sounds ok to me.

 

Hmm is this for real? There's at least one big mistake in the pineview diagram, there's nothing "PowerVX" in the GPU this is still old intel gma 3150 (the one with powervr gpu would be moorestown but this one isn't directly comparable).
The bobcat cores look too small to me. Half the size of the atom cores with more performance?
I guess it would basically be cedar with 2 bobcat cores bolted on (the die size would match, since cedar is 59mm^2 - with the cpu cores added that would be 75mm^2 which is close enough).
Seems a bit odd that the gpu is a full blown cedar if the cpu cores are that tiny (there's at least two ways cedar could be scaled back easily, either remove the second simd, or remove the second simd and make the other one 16-wide to compensate a bit but judging by size it would be a full cedar - not that scaling it back that way would change die size more than 5-10mm^2)...

 

the design methodology used in the atom could explain the large size of the cores. intel really doesnt care much about atom's performance because they have the market cornered.

 

If Johan did that assessment I'd wager it's correct, he's quite good at it. That being said I can't find the picture @ chip-architect.

 

Mr de Vries have been forum posting this picture. Perhaps its work in progress.

Here's another variant together with the released floor plan for Ontario. Link

 

It blows my mind just how small a proportion of the overall die space one of those Bobcat cores are, they're absolutely tiny! Integrated graphics are serious business these days.

 

When you think about it, Llano isn't all that different: 225mm², and each core takes less than 10mm².

 

i'm an amd fan, but it's hard to think that they could include ooo performance and low power in an architecture that take half the space of the atom
look at the L2 density, i belived that intel was heavely ahead

 

It's also interesting, that quad-core model with current GPU would be only 90mm² large and quad-core model with 160 SPs / 16 TMUs wouldn't probably exceed 110mm²...

fehu: it seems, that Atom CPU cores have a lot of blank areas (compared to any other CPU die-shot)

 

question #2
why don't they put an array of bobcat cores in a sea of cache for a niagara style server cpu?

 

Good question, they should be evaluating this now, I believe they could put four Bobcat cores where there is one Bulldozer core, the question is if four Bobcat cores are faster enough than a single Bulldozer core and if they scale well enough to pay for this design.

Niagara didn't had a lot of cache but had an advantage Bobcat doesn't have, threading.

Intel still ahead, it's cache is a bit denser:
http://citavia.blog.de/2010/04/14/a...f-westmere-sandy-bridge-and-llano-2x-8371390/

But that's comparing Intel 32nm vs GF 32nm not Intel 45nm vs GF 32nm or 28nm.

 

AMD is "definitely in the process of examining this as a design point - it would be foolish not to."

 

Wonder if we'll see Bobcat turn up in any of the next generation consoles, ATI are probably favourites to provide the GPU for all three consoles so I can't help but think they'll try and leverage that relationship to show the merits of a single chip design, that would have been a difficult task if all they had to sell was large and inefficient desktop x86 cores but Bobcat is much more like the sort of CPU you'd want in a console. Good console design is all about being area efficient and Bobcat looks to be utterly fantastic in that area, though maybe the lack of SIMD throughput in comparison to a PowerPC design may be an issue but if you've got a beefy GPU on the same die does that matter anymore? Bobcat supports SSE4.1 at least doesn't it? How does that instruction set compare to the revision of AVX found in the 360?


There definitely seems to be a lot of potential for this core, it'll be interesting to see where it ends up. AMD have a real window for opportunity if Intel aren't going to significantly increase Atom's performance until 2013. The single threaded performance of an Atom always struck me as too low and the latest revision was seriously underwhelming. At the very least, I know I'm already interested in an Ontario based netbook, and I think many others will be as well if you can demonstrate to consumers how it manages to "fix" most of the issues the current crop of Atom based netbooks have. Not including full 1080p hardware decoding was a serious own goal for Intel, I don't know what they were thinking by leaving that out of the latest revision, its such an easy feature to sell especially now Flash offers full hardware acceleration.