Richland & Kabini rumours

[Ventanni]

Temash isn't a chip you're going to want to play 64 player BF3 games on. It'll be excellent for the content it's designed for though: Movies, streaming, reading, writing, general productivity, and internet surfing. Maybe some light gaming.

 

[Hans de Vries]

Kabini/Temash die size: 11.1 mm x 10.2 mm = 114 mm2

http://www.theverge.com/2013/1/7/3848896/amd-temash-reference-tablet-hands-on-pictures#3913369
http://tweakers.net/nieuws/86510/amd-draait-dirt-showdown-op-full-hd-tablet-met-temash-apu.html


Hans.

 

[inf64]

Thanks Hans,great job as usual!

 

[Abwx]

That s quite a large die considering the 28nm node.

At the same node that a bobcat 1.0 it would be triple the area.

 

[Phynaz]

Makes Kabini roughly twice the cost to produce as Bobcat.

 

[Abwx]

Except that isn't coming until Excavator. Did you not read the AnandTech article on Steamroller?

So no, that is decidedly not where credit is due. And you should know that an automated process generally results in lower performance/clock speeds over a hand-tuned process.

Tools are rapidly improving and btw the said high density library cells
are first manually optimized before being used in an automated process
that will do the repetitive tasks/placements.

Parts of a design are already done this way with simulations
to check the global electrical properties.

 

[Abwx]

Makes Kabini roughly twice the cost to produce as Bobcat.

With 50% more surface..?...

Also , it is a quad core.

 

[Phynaz]

With 50% more surface..?...

Also , it is a quad core.

Yes.

And?

 

[Abwx]

28nm is not 35% more expensive than 40nm in a mm2 basis.

And being a quad core command higher prices and margins.

Does it even matter if it cost 15$ instead of 8$ if it can be sold
for 50 instead of 30 ?.....

 

[The Alias]

http://www.pcmag.com/article2/0,2817,2414008,00.asp

 

[inf64]

Does it even matter if it cost 15$ instead of 8$ if it can be sold
for 50 instead of 30 ?.....

That's it 🙂.

 

[Phynaz]

28nm is not 35% more expensive than 40nm in a mm2 basis.

Really...And what is GloFo's defect density for you to make claims like this?

 

[MightyMalus]

Well if these performance increases are not being helped by automated design somehow, someone in AMD sold their soul or did some crazy stuff for these improved chips.

I honestly don't think the cpu performance increase of the richland came from raising frequencies...how high would they be needed to increase to get that kind of performance? And that's a notebook richland!

I'm seriously wondering bout that 😵

 

[Haserath]

Kabini does have a better GPU, double the cores, and the chipset on there. Makes sense that it would be quite a bit bigger.

 

[SiliconWars]

Makes Kabini roughly twice the cost to produce as Bobcat.

Except it has the southbridge on-die whereas Bobcat required the Hudson FCH.

Also there will be ~520 die candidates per wafer instead of ~800. Take 80% of both and there are some 420 good die instead of 640, which is nowhere near twice the cost

 

[Abwx]

Really...And what is GloFo's defect density for you to make claims like this?

So if Glofo manage only 1% yield AMD would have to pay
the chips accordingly..?..

You are not serious....or clueless.

 

[itsmydamnation]

Makes Kabini roughly twice the cost to produce as Bobcat.

no it doesn't, single vs dual chip solution

 

[The Alias]

So if Glofo manage only 1% yield AMD would have to pay
the chips accordingly..?..

You are not serious....or clueless.

didn't amd strike a deal with glofo (and paid money) so they only pay money for working wafers/chips ?

 

[Idontcare]

Well if these performance increases are not being helped by automated design somehow, someone in AMD sold their soul or did some crazy stuff for these improved chips.

The other way around, what it is showing is just how poorly unoptimized the existing automated designs are.

Are today's automobiles ridiculously fuel efficient, or is it just that yesterday's automobiles were ridiculously fuel inefficient?

 

[Abwx]

didn't amd strike a deal with glofo (and paid money) so they only pay money for working wafers/chips ?

Surely that they pay in proportion of the targeted final yields.

It would make no sense to pay according to early yields manufacturing
prices since the foundry wouldnt even try to improve the thing given
that low double digit yields would be already profitable.

 

[Haserath]

Those are two different issues altogether, not to be correlated or connected in any sort of "cause and effect" linkage. When they do happen to occur at the same time, it is simply coincidence.

Lower "theoretical speed" is based on the inherent limitations within the design - sans the additional limitations imposed by less-than-ideal electrical parametrics from the process node itself.

Lower "actual clock speed" is based on inherent limitations imposed by process variability (or simply a failure to hit targeted parametrics entirely) at the hands of the fab.

At the end of the day both conspire together in a sort of multiplicative effect to reduce the final clockspeed envelope.

I assume Bulldozer gets limited thermally before anything else. Lower power allows the same heat sink to draw away more temperature from the chip at least( worded it strangely but hopefully you get the meaning).

Keeping temperature in check would seem like the best way to hit clock targets in the design.

I mean... Liquid nitrogen cooling allows Piledriver to go beyond 9Ghz in some cases.

So automated design to lower power use should possibly help in the real world application.

 

[AtenRa]

Makes Kabini roughly twice the cost to produce as Bobcat.

Makes Kabini have more performance at more than half the die size as ULV Trinity at 17-19W. 😉

2012(Brazos 2.0 E-1800 18W) : 2807 in PCMark Vantage
2013(Kabini): 5272 in PCMark Vantage

Trinity UltraThin A6-4455M (Dual Core Single Module 2.1GHz(2.6GHz Turbo) 256 Radeon Cores(SPs) 17W) PCMark Vantage score 3947

They are going to replace ULV Trinities with Kabini. Lower cost, higher margins.
Richland will be higher end products 25/35W and above.

 

[AtenRa]

Kabini/Temash die size: 11.1 mm x 10.2 mm = 114 mm2

http://www.theverge.com/2013/1/7/3848896/amd-temash-reference-tablet-hands-on-pictures#3913369
http://tweakers.net/nieuws/86510/amd-draait-dirt-showdown-op-full-hd-tablet-met-temash-apu.html


Hans.

Nice thx,

Would you say that the bigger die could incorporate 256 Radeon Cores (SPs) (4x 64)??

 

[itsmydamnation]

Brazos,
75mm cpu/gpu "SOC"
28mm FCH

total 103mm

Kabini
114 mm

CRYSIS AVERTED
people just need to count, 256 GCN cores would be extremely unlikely, nowhere near enough bandwidth on a 64bit memory system. the massive pin increase from brazos to kabini is likely all I/O that was on the FCH.

 

[AtenRa]

Brazos,
75mm cpu/gpu "SOC"
28mm FCH

total 103mm

Kabini
114 mm

CRYSIS AVERTED
people just need to count, 256 GCN cores would be extremely unlikely, nowhere near enough bandwidth on a 64bit memory system. the massive pin increase from brazos to kabini is likely all I/O that was on the FCH.

The FCH on Brazos was made at 40nm or 65nm ??

Edit: I believe its a 65nm part. If it is 28mm2 at 65nm, at 28nm it could be less than 10mm