AMD Unveils "Barcelona" Architecture

Quote:

---

Originally Posted by SOLDNER-MOFO64

Granted, but what success would they have had with a native quad-core and an even greater performance increase over opteron? Then they'd have less to worry about from AMD with theirs.

As it stands, the 5%-6% market-share they may have recovered is gonna all but dissapear overnight when AMD launch they're OPTY-QUADS.

So now INTEL have an even bigger fight on their hands than they would've had they gone the native route. IMO they'd have got there before AMD with all they're $$$$$ and know-how....and would have left AMD in a very precarious position. Alas they didn't and decided to rush out C2Q's still clinging onto the FSB like nobody's buisness and raking in as much $$$$ as possible....and as a result now find themselves in this position, by which I mean staring at BARCELONA.

---

Intel's decision to go with an MCM aproach was based on the following factors :

-costs
-performance
-time to market

Barcelona falls flat on all 3 :

-huge die , expensive to make
-looks like it will perform worse than C2Q
-no comment

Intel executives said that single die QC is too much for 65nm , the sweet spot is 80-150mm^2.Barcelona is 2x that and this will bite AMD dearly.

Secondly , AMD looks more and more like 3DFX preparing to release the 5000 and 6000 series ; huge monster that ultimately were the last nails in the company's coffin.

Quote:

---

Originally Posted by SOLDNER-MOFO64

Much the same as when it out-performs your pentium 6 (2xp3's) you'll moan "it's supposed to, ours are only pentium III's glued together in a fake quad-setup :hm: "

So what's the term for someone who buys a product which they are lead to believe is something which it is not? DUPED?

What's the term for someone who cries into they're cornflakes when BARCELONA is released? INTEL-USER :ROTF:

---

wth are you talking about ?

you pay for 4 cores on one cpu, you get 4 cores on one cpu. :confused:

Quote:

---

Originally Posted by SOLDNER-MOFO64

Granted, but what success would they have had with a native quad-core and an even greater performance increase over opteron? Then they'd have less to worry about from AMD with theirs.

---

This is where I think people need to really take some time and study the economics of this industry and think through the logic of what it takes to design a CPU.

The cost to manufacture a CPU is wrapped around a whole bucket of details, but ultimately it fits within die size and yield... if the die size is large yield is smaller, just by nature of the way CPUs are made. Example say you have 3 random particle defects that kill 3 random die on a wafer. Just as an example, say one die size puts 10 die to a wafer, but a different approach (smaller die) puts 20 die to a wafer, in one case you can kill 3 of 10 die, and yield is 70%... in the second case you kill 3 die or get 17 of 20 die working, yield is 85%. So to get the same through put and on larger die, a company (in this example) needs to increase capacity (tooling) by 15%... this runs into the hundreds of millions of dollars.

Making a monolithic die at 65 nm is relatively large, both AMD and Intel have the quad die of about 280 mm^2. But Intel will get higher yields as a killer only kills 1/2 the die.

Make no mistake, Intel will undoubtedly continue this trend... their monolithic quad will be at 45 nm, to go 8 core they will MCM it.... the monolithic 8 core will be at 32 nm and the 16 core will be MCM.... Intel takes the cost and design complexity and balances that approach to optimize yield and lower costs. Hence, we have a 270-300 buck quad core available to us.

Now, AMD will get a 250-300 buck quad to you as well... but they will not make nearly the money... this is good for you and me, but bad for AMD as their cost will be higher overall.

The compromise to time to market and costs is slightly lower performance scaling across threads.... in servers Intel is paying a larger price for this though as the BW of a shared bus does not compound with core count, this is the uniqueness of AMDs platform design/approach and has enabled AMD to get throughput performance way way up... just smart if you ask me, I think Intel is making a big mistake sticking with the parallel UMA approach this long.

In terms of how much better would they have done... it doesn't matter, pre-Barcelona -- a 2P server with quads just smashes anything 2P with dualies. The question was.. could Intel bring that in at reasonable costs such that the price/performance is compelling... indeed they did... Intel's Xeon 53XX series were not priced 2x above the dualies, rather some 10-20% overall... this is just killing AMD, if AMD had the resources and the design features capable they would have MCMed K8 to get quads out there asap.

http://www.fool.com/community/pod/2007/070823.htm

Quote:

---

Rivas was quoted as saying, "If I could do something different, I wish we would have immediately done a MCM - two dual cores and call it a quad-core."

---

Quote:

---

Originally Posted by largon

I wonder how revolutionarily ingenius does SOLDNER-MOFO64 declare the MCM approach when AMD introduces Montreal octacore...
You know, Montreal, the dual die Shanghai MCM.

Oh the irony... :p:

---

By the time Shanghai comes out, Nehalem will be out.

Native quad vs. native quad.
CSI vs. HT
IMC vs. IMC


And they way Intel is humming along, I've got $5 that they get there first and do a better job. The IPC w/ Conroe/Penryn looks like it is already better than K10. Goodness knows what kind of IPC increases are in store for us with Nehalem.

Quote:

---

Originally Posted by savantu

Intel's decision to go with an MCM aproach was based on the following factors :

-costs
-performance
-time to market

Barcelona falls flat on all 3 :

-huge die , expensive to make
-looks like it will perform worse than C2Q
-no comment

---

1- Barcelona core is smaller then Intel core. And a native core is much better then gluing cores.
2- That comment about performance is nothing but your personal believe. It´s just Fud what you are saying about performance.
3- Late.

Quote:

---

Originally Posted by SOLDNER-MOFO64

Much the same as when it out-performs your pentium 6 (2xPENTIUM-3's) you'll moan "it's supposed to, ours are only pentium III's glued together in a fake quad-setup :hm: "

So what's the term for someone who buys a product which they are lead to believe is something which it is not? DUPED?

What's the term for someone who cries into they're cornflakes when BARCELONA is released? INTEL-USER :ROTF:

---

<Rant>

When facts are few , experts are many.

One of the downsides internet forums have , is that it forces normal , reasonably intelligent people to see the ramblings of others from a different category.

And the funny part is that you can be warned or banned for insulting someone , although your comment is as valid as a medical diagnosis put by an expert.

Trying to be "politically correct" versus other people , you're forced to deny yourself the option of telling them what they really are.

Can't there be a warning or a ban for "plain stupidity" ?

<end rant>

Quote:

---

Originally Posted by savantu

Intel's decision to go with an MCM aproach was based on the following factors :

-costs
-performance
-time to market

Barcelona falls flat on all 3 :

-huge die , expensive to make
-looks like it will perform worse than C2Q
-no comment

Intel executives said that single die QC is too much for 65nm , the sweet spot is 80-150mm^2.Barcelona is 2x that and this will bite AMD dearly.

Secondly , AMD looks more and more like 3DFX preparing to release the 5000 and 6000 series ; huge monster that ultimately were the last nails in the company's coffin.

---

I don't know if it will bite AMD badly. They'll already have laid the groundwork for the move to 45nm. Good prove to be helpful.

Quote:

---

Originally Posted by v_rr

1- Barcelona core is smaller then Intel core. And a native core is much better then gluing cores.
2- That comment about performance is nothing but your personal believe. It´s just Fud what you are saying about performance.
3- Late.

---

1.Probably that's why it is less performing j/k.Did you do a die comparison ?
Secondly , what's a native core ? Please enlighten us.
2.The evidence is all around the web ; turning a blind eye to reality is a personal problem
3.Thanks for agreeing at least on one point.

Quote:

---

Originally Posted by SOLDNER-MOFO64

As it stands, the 5&#37;-6% market-share they may have recovered is gonna all but dissapear overnight when AMD launch they're OPTY-QUADS.

So now INTEL have an even bigger fight on their hands than they would've had they gone the native route. IMO they'd have got there before AMD with all they're $$$$$ and know-how....and would have left AMD in a very precarious position. Alas they didn't and decided to rush out C2Q's still clinging onto the FSB like nobody's buisness and raking in as much $$$$ as possible....and as a result now find themselves in this position, by which I mean staring at BARCELONA.

---

So I will follow up with a second post on this topic.... perhaps AMD will start to regain lost share, it is not certain at this point, because Barcey is not launching at a clock speed that will dominate any benchmarks over Intel excpet for FPU, and the majority of the market is not FPU driven. AMD will solidify/gain in the scientific/engineering applications where they already have the largest install base, but right now Barcey's top bin will not match or beat Intel's top bin in server or desktop, this is simply not going to happen and the initial data (we are finally getting some) is starting to show that.

This is the second problem inherent with a monolithic design. The clock speed you are capable of achieving is dependent upon the slowest clockable component of the die. During manufacturing there is spatial variation from one die to the next that alters the Fmax (max clocks) that the die can achieve. Intel's MCM approach means they can bin out and match die of like speeds more effectively, where as a monolithic die, say 3 core are capable of 2.8 GHz but the 4th is screwed up and only clocks at 2.0 Ghz before giving out... the overal processor will only give you 2.0 Ghz.... what I am talking about here is die to die and within die variation.

http://eda.ee.ucla.edu/EE201A-04Spring/GIT-PV.pdf

Statistically speaking, a MCM approach will not only drive yields higher but also allows better binning for higher clocking parts.

Jack

Quote:

---

Originally Posted by JumpingJack

This is where I think people need to really take some time and study the economics of this industry and think through the logic of what it takes to design a CPU.

The cost to manufacture a CPU is wrapped around a whole bucket of details, but ultimately it fits within die size and yield... if the die size is large yield is smaller, just by nature of the way CPUs are made. Example say you have 3 random particle defects that kill 3 random die on a wafer. Just as an example, say one die size puts 10 die to a wafer, but a different approach (smaller die) puts 20 die to a wafer, in one case you can kill 3 of 10 die, and yield is 70%... in the second case you kill 3 die or get 17 of 20 die working, yield is 85%. So to get the same through put and on larger die, a company (in this example) needs to increase capacity (tooling) by 15%... this runs into the hundreds of millions of dollars.

Making a monolithic die at 65 nm is relatively large, both AMD and Intel have the quad die of about 280 mm^2. But Intel will get higher yields as a killer only kills 1/2 the die.

Make no mistake, Intel will undoubtedly continue this trend... their monolithic quad will be at 45 nm, to go 8 core they will MCM it.... the monolithic 8 core will be at 32 nm and the 16 core will be MCM.... Intel takes the cost and design complexity and balances that approach to optimize yield and lower costs. Hence, we have a 270-300 buck quad core available to us.

Now, AMD will get a 250-300 buck quad to you as well... but they will not make nearly the money... this is good for you and me, but bad for AMD as their cost will be higher overall.

The compromise to time to market and costs is slightly lower performance scaling across threads.... in servers Intel is paying a larger price for this though as the BW of a shared bus does not compound with core count, this is the uniqueness of AMDs platform design/approach and has enabled AMD to get throughput performance way way up... just smart if you ask me, I think Intel is making a big mistake sticking with the parallel UMA approach this long.

In terms of how much better would they have done... it doesn't matter, pre-Barcelona -- a 2P server with quads just smashes anything 2P with dualies. The question was.. could Intel bring that in at reasonable costs such that the price/performance is compelling... indeed they did... Intel's Xeon 53XX series were not priced 2x above the dualies, rather some 10-20% overall... this is just killing AMD, if AMD had the resources and the design features capable they would have MCMed K8 to get quads out there asap.

http://www.fool.com/community/pod/2007/070823.htm

---

Interesting read. Good argument.

The fact that someone from AMD admits they wish they'd gone MCM is a revelation to me. IMO though that could be more as a result in hindsight (due to the probs they've encountered) as opposed to the end result performing any greater. They probably do wish they'd gone the MCM route now, since it would've been much quicker than BARCELONA has ended up being.....but on the up-side.....it may turn out for the better. Who knows?

edit & deleted:
It's gotta stop somewhere...
...better stop here.

For me.

Heh...native vs true and blabla. Core 2 is a true dualcore over K8 then. And so what? It means nothing.

All that matter is performance. And K10 seems to lack that as predicted. Hell it could be a :banana::banana::banana::banana:ing pink doll in the CPU socket for what I care. As long as the performance is there.

Be nice to SOLDNER-MOFO64, its gonna be a very rough day for him tomorrow. He will need support to get through the day when his favoured company fails again. And yet another hype is busted and broken to pieces.

why don't we skip the bullship and judge on performance alone?

Quote:

---

Originally Posted by JumpingJack

So I will follow up with a second post on this topic.... perhaps AMD will start to regain lost share, it is not certain at this point, because Barcey is not launching at a clock speed that will dominate any benchmarks over Intel excpet for FPU, and the majority of the market is not FPU driven. AMD will solidify/gain in the scientific/engineering applications where they already have the largest install base, but right now Barcey's top bin will not match or beat Intel's top bin in server or desktop, this is simply not going to happen and the initial data (we are finally getting some) is starting to show that.

This is the second problem inherent with a monolithic design. The clock speed you are capable of achieving is dependent upon the slowest clockable component of the die. During manufacturing there is spatial variation from one die to the next that alters the Fmax (max clocks) that the die can achieve. Intel's MCM approach means they can bin out and match die of like speeds more effectively, where as a monolithic die, say 3 core are capable of 2.8 GHz but the 4th is screwed up and only clocks at 2.0 Ghz before giving out... the overal processor will only give you 2.0 Ghz.... what I am talking about here is die to die and within die variation.

http://eda.ee.ucla.edu/EE201A-04Spring/GIT-PV.pdf

Statistically speaking, a MCM approach will not only drive yields higher but also allows better binning for higher clocking parts.

Jack

---

Jeeeezy peeps your hard to bloody argue with.....can't you just forget you know all that stuff and pretend I'm right? :up:

Just kiddin'....:)

Quote:

---

Originally Posted by Shintai

Heh...native vs true and blabla. Core 2 is a true dualcore over K8 then. And so what? It means nothing.

All that matter is performance. And K10 seems to lack that as predicted. Hell it could be a :banana::banana::banana::banana:ing pink doll in the CPU socket for what I care. As long as the performance is there.

Be nice to SOLDNER-MOFO64, its gonna be a very rough day for him tomorrow. He will need support to get through the day when his favoured company fails again. And yet another hype is busted and broken to pieces.

---

You never know your luck Shintai.......(better watch my spelling huh? ;))

http://i65.photobucket.com/albums/h2...ys/popcorn.gif

Quote:

---

Originally Posted by savantu

1.Probably that's why it is less performing j/k.Did you do a die comparison ?
Secondly , what's a native core ? Please enlighten us.
2.The evidence is all around the web ; turning a blind eye to reality is a personal problem
3.Thanks for agreeing at least on one point.

---

1- Transistor count.
2- Those evidences are nothing just FUD. Arround the web you have nothing but 2 spec rates in Int and Float. That´s nothing to say something about performance.

Quote:

---

Originally Posted by SOLDNER-MOFO64

Jeeeezy peeps your hard to bloody argue with.....can't you just forget you know all that stuff and pretend I'm right? :up:

Just kiddin'....:)



You never know your luck Shintai.......(better watch my spelling huh? ;))

---

This is what makes arguing fun, if it can stay civil ... and, so long as people can concede points where good points are made.

Here's the ticket.... Barcelona is gonna be a good product, AMD has some of the top designers in the world working on it. A quad product (let's pretend it was not even reworked but just a K8) is going to give a massive performance boost in many situations, so people who are AMD devoted for what ever reason will enjoy this CPU no doubt.

The personal preference of a CPU is an odd one, and can come wrapped up in an emotional attachment -- which I have always found odd. Look at it this way, there is nothing I can do on an Intel CPU that I cannot do on an AMD CPU, the only real differential is price and performance, at any given time I can accomplish a computational task faster on one vendor's product than I can on the other... but there is little else to differentiate the two products other than silly details that us 'geeks' like to argue about....

In the interim, if someone makes an point that I do not think is true I typically will research the point and in doing so, learn something new... sometimes I am right, sometimes I am wrong (the best learning experience), in either case I still find it fun talking about this stuff :)

More on point -- the MCM vs monolithic debate will rage on forever I suspect, but almost all arguments I see on this topic are from two very different points of view and thus will never be settled. Monolithic, expecially AMD's design, is unquestionable more sound in terms of extracting all that a multicore processor could do in a threaded environment. People who look at the MCM and take this line of argument focus mostly on output, but when thinking through which is 'better' or 'worse' than the other it is best to think of both input and output and understand it is a give and take scenario.

Jack

v_rr...how big is k10..283mm2? A kentsfield/Clovertown is 2x143mm2.

Now what do you think got the best yield? Get 2 cores per chip to work and at a certain speed. Or 4? Also with a single die 283mm2 with relatively little cache. Defects are a bad and serious business. on the 143mm2 C2Ds its alot easier. Sicne half the chip is cache roughly and cache defects are relatively easy to fix.

Making a Core2Quad is ALOT cheaper than making a 65nm K10 Quad ever will be. Its about the same size. But defects and bins kills AMD flat out in cost and speed.

All I hope for is that Barcelona is close to performance versus Intel clock for clock. I don't care if core is slightly faster. I want 8 cores with 4x4 or 8x8 if its going to be called that without having to buy registered ddr2 memory. If I hear of some plans for an Intel motherboard to support dual sockets with regular ddr2 then I may change my mind.

Quote:

---

Originally Posted by Shintai

v_rr...how big is k10..283mm2? A kentsfield/Clovertown is 2x143mm2.

Now what do you think got the best yield? Get 2 cores per chip to work and at a certain speed. Or 4? Also with a single die 283mm2 with relatively little cache. Defects are a bad and serious business. on the 143mm2 C2Ds its alot easier. Sicne half the chip is cache roughly and cache defects are relatively easy to fix.

Making a Core2Quad is ALOT cheaper than making a 65nm K10 Quad ever will be. Its about the same size. But defects and bins kills AMD flat out in cost and speed.

---

V_rr, read what Shintai has written. Barcelona is 283 mm^2, monolithic. Kentsfield (2 Conroe glued so to speak) is 143 mm^2 x 2 or 286 mm^2.... the total Si realestate is the same between the two.... but Intel will enjoy higher yields and lower costs overall.

The die sizes are readily available all over the net if you need I can provide you links.

jack

Quote:

---

Originally Posted by Shintai

v_rr...how big is k10..283mm2? A kentsfield/Clovertown is 2x143mm2.

---

Transistor count ;)
Also look at the transistor count of Penryn.

Quote:

---

Now what do you think got the best yield? Get 2 cores per chip to work and at a certain speed. Or 4? Also with a single die 283mm2 with relatively little cache. Defects are a bad and serious business. on the 143mm2 C2Ds its alot easier. Sicne half the chip is cache roughly and cache defects are relatively easy to fix.

Making a Core2Quad is ALOT cheaper than making a 65nm K10 Quad ever will be. Its about the same size. But defects and bins kills AMD flat out in cost and speed.

---

Well that&#180;s what you say. But anyone here know how in practice things are going in AMD and Intel fabs.
What you are saying is mixing logic, but in pratice we don&#180;t know how things truly are.

Nvidia 8600/8400 -> 80nm
ATI HD 2600/2400 -> 65nm

Nvidia 8800 -> 90nm
ATI HD 2900 -> 80nm

Even HD 2600 have a smaller core then 8400.
Nvidia have a lot bigger core in every segments and is doing well with that.... and don&#180;t have it&#180;s own factorys like AMD and Intel have.

The problem is not in core sizes.... ATI/Nvidia example just shows it.

Quote:

---

Originally Posted by v_rr

The problem is not in core sizes....

---

Yield is always related to die size... this is the nature of the beast.

Also, your 80 / 90 nm are not die sizes, these are litho nodes.... depending on how companies spend tranistor budget does not always mean smaller die relative to the competition.

remember yields are quadratic to die size

Quote:

---

Originally Posted by JumpingJack

Also, your 80 / 90 nm are not die sizes, these are litho nodes.... depending on how companies spend tranistor budget does not always mean smaller die relative to the competition.

---

HD 2600 have a smaller core then 8400.

Is that enought for you ;)

Quote:

---

Originally Posted by v_rr

HD 2600 have a smaller core then 8400.

Is that enought for you ;)

---

Perhaps it is smaller, what is the die area in mm^2.... you posted litho nodes and stipulated that these relate to die size, this is only true within the same class of chips... but between manufacturers, die sizes vary based on design and litho node != die size as you alluded to..... what you posted was an incorrect conveyance of what were trying to argue... i.e. you did not prove a thing.