Richland & Kabini rumours

[SiliconWars]

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

Yep it's 65nm.

 

[AtenRa]

Well then,

Even if at 28nm is 10mm2 + 75mm2 = 85mm

114mm2 is 34% larger than 85mm,
Graded Jaguar cores are 10% larger than Bobcat at 28nm but that would not translate for that die difference. I dont know but either they put 128-bit memory or more SPs.

I dont have another explanation

Edit: and i dont believe the larger iGPU Front End in Kabini will make for that much of die difference. Also, if at half the die size vs Trinity with smaller cores and smaller process (28nm vs 32nm)cant have at least half the Radeon Cores (384) they seriously messed up the design.

 

[itsmydamnation]

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

no idea, but I/O Logic normally doesn't shrink anywhere near as well as things like sram so it could be 65nm if it was more cost effective but an assumption of 40nm isn't a bad one.

also GCN cores are a fair bit bigger core/ALU then VLIW5 if normalized on the same process.


edit: also remember they might have added aditional I/O, 8X/16 PCI-E if they are going to push a crossfire solution

edit2: my quick math tells me the following
bobcat core 4.6 mm
jaguar core 3.2 mm
jag is 10% bigger then bob @ same node
4.6 *0.9 =4.14
4.14/3.2 = 1.29

so a 40nm to 28nm shrink has bought them a 30% reduction in size for CPU logic. The caches have likely scaled better.

 

[Haserath]

Well then,

Even if at 28nm is 10mm2 + 75mm2 = 85mm

114mm2 is 34% larger than 85mm,
Graded Jaguar cores are 10% larger than Bobcat at 28nm but that would not translate for that die difference. I dont know but either they put 128-bit memory or more SPs.

I dont have another explanation

Brazos+FCH at 28nm would be around 45-50mm, right? So accounting for non-perfect scaling, 2 CUs, andm4 bigger cores, Jaguar sounds to be about the right size.

All this extra would seem like it would need more bandwidth though...

 

[AtenRa]

Brazos+FCH at 28nm would be around 45-50mm, right? So accounting for non-perfect scaling, 2 CUs, andm4 bigger cores, Jaguar sounds to be about the right size.

All this extra would seem like it would need more bandwidth though...

Well, you might be right. Its getting late, ill take a fresh look in to it in the morning again

 

[Phynaz]

no it doesn't, single vs dual chip solution

It's more costly to produce one large chip than two small ones.

 

[Phynaz]

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

You are not serious....or clueless.

Or an Intel stockholder / pumper, right?

 

[itsmydamnation]

It's more costly to produce one large chip than two small ones.

well that depends on the yield curve, which we have no idea what its like for bobcat or jaguar based SOC's. but either way its not going to be twice the price. The MB's are going to be simpler so they can sell at a increased price relitive to bobcat without affecting OEM's margins or RRP.

 

[Haserath]

It's more costly to produce one large chip than two small ones.

It should make an easier cooling system and a cheaper system in general. Thus they get to charge more or maybe not.

 

[Nemesis 1]

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?

That a matter of perspective . My dad bought me my first car . LOL With blown engine. We completely overhauled the motor to better than new specs . It was an inline six 59 impala . When we were all done it averaged 17 miles to a gallon about 20 miles pg highway.
I could fill the 22 gallon tank for less than 20 dollars . gase pricies varied between 26 cents a gallon 30 cents . Best car I ever owned and than a drunk destroyed while it was parked. I would take that car over many modern cars that are speced for 30 + mpg that never comes close to those mpg rating . Honesty died along time ago . I would take the same $1.25 an hour also over a $30 dollar hour job to day if the pricies to day were the same as than . A new car $1800 $22oo for a musle car . I house sold for around 30 that today is $300.000 . Comparing today with yesterday . I take yesterday . People were better . Charity was great and it was charity without pay . not today . Everybody gets payed now adays . Ice cream cone double 10 cents single a 5 cents . Had they not scewed the cost of living numbers . that same dollar 25 cents today would be about 45 dollars an hour .

 

[Nemesis 1]

well that depends on the yield curve, which we have no idea what its like for bobcat or jaguar based SOC's. but either way its not going to be twice the price. The MB's are going to be simpler so they can sell at a increased price relitive to bobcat without affecting OEM's margins or RRP.

So everybody makes out but the consumer . I like that kind of thinking it has gotten us to were we are today.

 

[piesquared]

So everybody makes out but the consumer . I like that kind of thinking it has gotten us to were we are today.

And yet this is the first time you've seemed to favor consumers, previously claiming, intel for example, was awesome for charging consumers as much as possible. Besides, his point was that OEM's could maintain margins without increasing consumer cost.

p.s. unless you were penning sarcasm.

 

[Olikan]

Graded Jaguar cores are 10% larger than Bobcat at 28nm but that would not translate for that die difference. I dont know but either they put 128-bit memory or more SPs.

I dont have another explanation

it's a quadcore vs a dual core

 

[MightyMalus]

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?

I only ment that it helped, as it was shown on past slides from AMD that they found ways that it could make things tighter. Maybe it was partially done, maybe it wasn't at all. If not, how does that explain Richland!?

And cars today are averagely more fuel efficient, not as much as some other unique ones but there is an oil business and all. Same way how Intel is not using the GT3 on the desktop. Why would they NOT do such a thing?

 

[Nemesis 1]

And yet this is the first time you've seemed to favor consumers, previously claiming, intel for example, was awesome for charging consumers as much as possible. Besides, his point was that OEM's could maintain margins without increasing consumer cost.

p.s. unless you were penning sarcasm.

Link to such a claim by me . In 2006 Intel lowered pricies across the board . What you want intel to lower pricies by $ 100 on the 5/7 . I would love that . AMD not so much, Intel unlike me wants amd around . I would love to see richland beat the k series from intel . It would be a price cut across the board by intel with haswell . I have no shares in intel. But intel lower its pricies across the board wouldn't hurt so much as some would expect . Intel is selling high dollar knight corner . They can't keep up with demand yet today. Intel is finely in phones and that will expand rapidly .

Henery Ford a great man . Did say I would rather build a million cars at $1 profit than 1 car at a million dollar profit . A great man , Henry understood volumn sales . IN phones chips there are volumns of chips to made . Intel now has a small % of that volumn in a low margine part . Volumn makes up for margines per part . Margines are great but the bottom line is the all be. Profits . I fully expect haswell to be lower priced than present IB . if not it will happen with broadwell . Intel is AFRAID of hurting AMD threw pricing . But if intel entered into AMD pricing . AMD would be done gone bye bye . The monopoly people would be at intels doors . Monopolies can hurt consumers as much as help them . Intel has to keep its pricies high . but ya intel really needs to come down from pricies set on 65nm.

 

[Homeles]

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

It's a two-way street. You don't have any support for your claim either.

 

[pablo87]

The die size is not surprising - you have to add the FCH. Also, if kabini was designed for portability between TSMC and Global Foundries (which it should be given the WSA), wouldn't this cause the die size to be larger? You also have to think they've designed in some additional redundancy on the GPU side.

Most important, small die makes no sense considering the WSA. Half of the annual commitment is already 45MM APUs. If anything, it would have made sense to make it even larger though that would mean no free lunch for Global Foundries and that just can't be allowed.

 

[Idontcare]

That a matter of perspective . My dad bought me my first car . LOL With blown engine. We completely overhauled the motor to better than new specs . It was an inline six 59 impala . When we were all done it averaged 17 miles to a gallon about 20 miles pg highway.
I could fill the 22 gallon tank for less than 20 dollars . gase pricies varied between 26 cents a gallon 30 cents . Best car I ever owned and than a drunk destroyed while it was parked. I would take that car over many modern cars that are speced for 30 + mpg that never comes close to those mpg rating . Honesty died along time ago . I would take the same $1.25 an hour also over a $30 dollar hour job to day if the pricies to day were the same as than . A new car $1800 $22oo for a musle car . I house sold for around 30 that today is $300.000 . Comparing today with yesterday . I take yesterday . People were better . Charity was great and it was charity without pay . not today . Everybody gets payed now adays . Ice cream cone double 10 cents single a 5 cents . Had they not scewed the cost of living numbers . that same dollar 25 cents today would be about 45 dollars an hour .

If you were living in this fantasy of yours then you'd actually have been dead long ago because your health sucks and modern medicine did not exist for you to use as a crutch to keep both feet out of the grave.

(yes that costly and heavily subsidized industry that makes all the difference between living to 50 versus living to 80 for most males nowadays versus 50 yrs ago)

The difference between then and now pretty much all comes down to mortality rates, both infants and adults. With the reduced mortality rates and increased lifespan, combined with folks still wanting to "retire" (something else that didn't exist 50yrs ago as it does today) has resulted in the ballooning of the burden placed on the working by those who are no longer contributing a net-plus back into the society from which they now suckle.

The tax burden on today's working class has never been higher, their medical expenses have never been higher, their cost of living has never been higher...and it is directly because everyone has better health (even today's diabetic obese have longer expected lifespans than the average fit-and-trim adult in 1950), living longer but with a vastly larger economic footprint on society when they retire (and live longer while retired because the medical benefits continue to be subsidized).

Your fantasy is great, but you would not have lived to enjoy it. My grandfather lived to be 45, my father lived to be 56, if I live to be 65 it will only be because today's society will continue to provide me with access to medicine that was not available to my ancestors.

I only ment that it helped, as it was shown on past slides from AMD that they found ways that it could make things tighter. Maybe it was partially done, maybe it wasn't at all. If not, how does that explain Richland!?

And cars today are averagely more fuel efficient, not as much as some other unique ones but there is an oil business and all. Same way how Intel is not using the GT3 on the desktop. Why would they NOT do such a thing?

Oh yeah, no question it helps if you are in AMD's situation. They simply cannot afford to hire and maintain a big expense design team tasked with doing custom hand-layout. For the budget AMD is operating within, automated layout provides the best bang-for-the-R&D-buck they could hope to obtain.

And what is great about automated layout is that it is something that only gets better with time in the same way that software compilers continually get better and better at compiling the same old code.

 

[Nemesis 1]

My dad is 82 years old never been to hospital other than a shoulder injury . My grandmother 87 died . never been in hospital never took meds . Oh my dad never had a tooth cavity till he was 68 years old . Its how you live . I had many friends went to nam . all dead threw our government use of agent orange . Some towns in USA used by gooberment as test subject . Who do you suppose was testing the results . By the way the only thing I ever got in the usa for meds did nothing , Now mexico had the good stuff . But I not using that , True I more dead than alive . Cigerates please explain why we can by them over the counter today . Faith missplaced is not wise.

 

[CTho9305]

What is not known is just how "sub-optimal" a modern sythesized layout is compared to a modern hand-layout IC. And once you factor in budget constraints, a synthesized layout may actually fare better than a hand-layout IC if you can't afford to hire enough people to do the hand-layout job decently.

(cut too many corners and you can make anything be the crappiest way to go )

In general though it is widely accepted that if you are operating with an appropriately resourced layout team then the layout job when done by hand will yield superior results to that achieved by an optimal synthesis layout (however this is a moving goalpost because synthesis tools are advancing at a pace comparable to Moore's law whereas the pace of advancements in hand-layout is much more human-like).

I'm very skeptical that hand design buys you anything at all for many design points. In my personal experience, automated designs were no slower, but much smaller. You can also look to ARM: Cortex A8 had a bunch of structured datapaths, but as performance of their subsequent cores improved, automation increased and manual intervention decreased. I think there's also a paper from one of the companies involved with Cell (Toshiba?) showing that they saved area and gained performance by switching to automated design. For something like POWER7 with a 32KB, 2-cycle L1 cache at >3GHz (holy crap!), sure, maybe that has to be designed by hand...but most things? I don't think so.

 

[NostaSeronx]

SB7x0-9x0 = 130-nm
FCH = 65-nm

Sorry for the late post.

 

[Idontcare]

I'm very skeptical that hand design buys you anything at all for many design points. In my personal experience, automated designs were no slower, but much smaller. You can also look to ARM: Cortex A8 had a bunch of structured datapaths, but as performance of their subsequent cores improved, automation increased and manual intervention decreased. I think there's also a paper from one of the companies involved with Cell (Toshiba?) showing that they saved area and gained performance by switching to automated design. For something like POWER7 with a 32KB, 2-cycle L1 cache at >3GHz (holy crap!), sure, maybe that has to be designed by hand...but most things? I don't think so.

It all comes down to how well resourced the hand design project was.

Would you agree that humans could replicate an automated design if you simply hired enough of them and gave them enough time to get there? (in the academic argument, the infinite monkey theorem)

The infinite monkey theorem states that a monkey hitting keys at random on a typewriter keyboard for an infinite amount of time will almost surely type a given text, such as the complete works of William Shakespeare.

I'm not in any way arguing in any of my posts in this thread that hand-coded design bests automated design...I am saying there are obvious resourcing caveats that have to come into the discussion if there is to be any sort of apples-to-apples comparison made.

Under-resource any hand design team and you will surely get something far less optimal than that obtainable with the same resources enabling an automated design team.

Over-resource any automated design team and you will surely get something no better than that obtainable with the same resources enabling a hand-layout team. (since in the limit of infinite optimization loops - be it cpu cycles or humans - they both will end up solving the same "traveling salesman" problem)

And I argue that the switch-over point in that trade-off equation will forever be increasing towards favoring the automated design team. (meaning comparable funding for the automated team remains inferior to funding a hand-design team, the funding level for parity increases every year in favor of automated design)

 

[NostaSeronx]

AMD is pure hand-design except for the Cat Family. The R&D team has largely been unaffected by the layouts ranging from 2000 to 2013.

 

[CTho9305]

It all comes down to how well resourced the hand design project was.

Would you agree that humans could replicate an automated design if you simply hired enough of them and gave them enough time to get there? (in the academic argument, the infinite monkey theorem)



I'm not in any way arguing in any of my posts in this thread that hand-coded design bests automated design...I am saying there are obvious resourcing caveats that have to come into the discussion if there is to be any sort of apples-to-apples comparison made.

Under-resource any hand design team and you will surely get something far less optimal than that obtainable with the same resources enabling an automated design team.

Over-resource any automated design team and you will surely get something no better than that obtainable with the same resources enabling a hand-layout team. (since in the limit of infinite optimization loops - be it cpu cycles or humans - they both will end up solving the same "traveling salesman" problem)

And I argue that the switch-over point in that trade-off equation will forever be increasing towards favoring the automated design team. (meaning comparable funding for the automated team remains inferior to funding a hand-design team, the funding level for parity increases every year in favor of automated design)

There are two different questions... there's the infinite-ideal-humans question, and the maximum-real-humans-and-time-that-even-profitable-companies-can-afford question. Infinite ideal humans with infinite time will beat automation in every dimension. That said, I don't think I'd want that job (unless you're paying infinite $$); doing hand design of irregular structures is incredibly tedious, and I personally get demotivated when I don't feel like my work is making the end product meaningfully better.

I think you could hand-build something like a Bobcat core in a couple of years with, say, 100 implementation engineers. That seems like a big implementation team to me; a larger team starts to introduce serious management and communication challenges. If you take much more time, you run into issues like process nodes: a competitor will just auto-place something on the newer process node and get a better final product.

I don't know that you'd be able to get a result in as little area. The simple act of breaking a core into multiple pieces for engineers to build in parallel will introduce overhead: you will never correctly predict the area and shape each piece will need. Consider Apple's core in the A6 SOC:

I doubt they intended for left side to be so uneven - it's leaving a lot of wasted space. There's also a big gap between the cores and the L2 cache; is the L2 control logic a little too wide? At a finer level of detail, you can see whitespace inside and between all the subsections of the cores. Fixing that would require a lot of iterative redesign. That takes time--time in which you (Apple) are stuck selling clunky Cortex A9-based A5's (wow that's confusing!) while Samsung moves to Exynos 5 and Qualcomm moves to Krait.

Human engineers might be able to make each sub-block faster, but since you've broken the design up into small chunks, you now have to handle timing paths that cross between blocks. In practice, dealing with that tends to be difficult and again introduces the need to iterate repeatedly - and remember, the whole core is limited to the speed of its slowest block. If 99 blocks are 20% faster than the automated design, but one block is only 5% faster, you're only going to be 5% faster; you should really go back and slow down the 99 blocks to recover area and power. I don't think anyone has time to do that properly.

You also have the problem that some percentage of your engineers won't be as competent/skilled as the rest, and that impacts the quality of work you get from them. It's challenging to maintain team quality as team size grows.

Or maybe I'm underestimating the amount of work people are willing/able to put into their chips. Nehalem's floorplan is pretty impressive.

The R&D team has largely been unaffected by the layouts ranging from 2000 to 2013.

Assuming you typo'd "layoffs", your sources are terrible and you should look for different ones.

 

[AtenRa]

it's a quadcore vs a dual core

Yea but we have 28nm vs 40nm.