Dissappointing Year So Far for AMD

[duby229]

Do I really need to spell it out for you? Kentsfield has a 10% IPC advantage over Phenom, Yorkfield has a 15% advantage.

Now, you tell me, what you think AMD needs in order to compete with 2.5 / 2.66GHz 45nm C2Qs.

They certainly dont need the high end to compete well. As lond as they acn keep price/performance competitive on the main stream they'll be able to do reasonably well. They may not be able to take the crown, but they can compete on the main stream.

Do I really need to spell it out for you? The boat floats both ways.

 

[harpoon]

They certainly dont need the high end to compete well. As lond as they acn keep price/performance competitive on the main stream they'll be able to do reasonably well. They may not be able to take the crown, but they can compete on the main stream.

So Q9300 / Q9450 is 'high end' now? No, they are the budget/mainstream quads, that AMD is in direct competition with.

Also, I predicted price cuts come Q1 due to Penryn, did I not? Since AMD cannot compete at the high end, even with a 3GHz Phenom, they will have to play the price/performance card. The same card that has led them to lose over $2 billion in the past year...

 

[vista_blista]

Those staggered price cuts and dates were announced tentatively a few months ago. But your words tell us far more than you realize.

 

[duby229]

So Q9300 / Q9450 is 'high end' now? No, they are the budget/mainstream quads, that AMD is in direct competition with.

Also, I predicted price cuts come Q1 due to Penryn, did I not? Since AMD cannot compete at the high end, even with a 3GHz Phenom, they will have to play the price/performance card. The same card that has led them to lose over $2 billion in the past year...

Well we'll see, but lke I told someone else, come this time next year, it'll be interesting to revisit this thread, and I'll be willing ot bet that I'll be much closer to the mark then you are. Additionally I'll point you towards AMD's quarterly trends. If your unable to understand it, I'll be glad to give you my interpretation.

 

[Michaelius]

What they need to compete is cheap quad cores that can come at price ranges of intel dual core cpus (around 130-160$ range)

 

[Scali2]

What they need to compete is 45 nm and MCM quad cores that can come at price ranges of intel dual core cpus (around 130-160$ range)

65 nm native quadcore? Impressive technical achievement, but not what AMD needs right now.

 

[duby229]

65 nm native quadcore? Impressive technical achievement, but not what AMD needs right now.

I disagree. Dual die MCM's cost more to produce then AMD's current monolithic quad core. With AMD's current cash situation they indeed made the right choice. If they had cash to burn they may have been able to get a quad core out sooner, but given the performance problems of the current quad core, one can assume that an MCM quad core based on the same technology would be even worse. They definitely made the right choice with it's monolithic die.

In the end they would have had something that cost more, and performed worse....

 

[Scali2]

I disagree. Dual die MCM's cost more to produce then AMD's current monolithic quad core.

You're wrong.
With MCMs you have half the die-size, hence you have less 'critical area' per die, so you have higher yields.
You can easily mix-and-match two dies from anywhere in the wafer to get a proper quadcore at whatever speedbin you like.
So the effective yields are higher.

 

[duby229]

You're wrong.
With MCMs you have half the die-size, hence you have less 'critical area' per die, so you have higher yields.
You can easily mix-and-match two dies from anywhere in the wafer to get a proper quadcore at whatever speedbin you like.
So the effective yields are higher.

If you can show the numbers then I'll believe you, but in the mean time all of the calculations that I've done show MCM's as costing as much as 33% more. Also there is no way in hell that you can know what yields are. Neither company releases that information, and they both make very sure that they dont release enough supporting data so that it can be calculated. We can make educated guesses, but that is really the best that anybody can claim. Your claim assumes a very high defect ratio, which nobody knows what they actually are. So without that information all you can do in faith is assume the best case scenario. Anything else is being a pessimist.

 

[Scali2]

If you can show the numbers then I'll believe you, but in the meaqn time all of the calculations that Iiii've done show MCM's as costng as much as 33% more.

Why don't you show the numbers first?
I don't think anyone has to prove that MCM is cheaper. Intel itself mentioned this, and various sources also claim the same thing.
Eg http://en.wikipedia.org/wiki/Core_2

For example, the QX6700 consisted of two E6700 chips connected together by a 1066 MT/s FSB on one MCM, resulting in lower costs but less bandwidth to the northbridge.

Or http://www.maximumpc.com/article/Multicore-Madness

First answer: economy. The 65nm process shrink will dramatically reduce Intel's manufacturing cost for a single-core chip. MCMs are more expensive to produce, but Intel has made some advances in this technology since the Pentium Pro. Therefore, it will be cheaper for Intel to stuff two Presler-class dies into a single package than to make a dual-core Presler-class die, even with the additional cost of the MCM package.

This is widely regarded as truth, as it makes perfect sense, and all evidence points to its favour.

So you will have to disprove this common idea of MCM being a cheaper production method first.

 

[brucedeluxe169]

I don't think AMD has hit bottom yet.

the bottom for AMD will come when Nehalem is launched and it devastates the Penryn family.

 

[Digital Viper-X-]

Why don't you show the numbers first?
I don't think anyone has to prove that MCM is cheaper. Intel itself mentioned this, and various sources also claim the same thing.
Eg http://en.wikipedia.org/wiki/Core_2


Or http://www.maximumpc.com/article/Multicore-Madness


This is widely regarded as truth, as it makes perfect sense, and all evidence points to its favour.

So you will have to disprove this common idea of MCM being a cheaper production method first.

I think the whole reason MCMS are considered cheaper is due to the yields, which are probably much easier to get with 2 dual cores,

on 1 wafer, if a single dual core die is bad, oh well you lose 1 dual core chip, to make a quad, you need 2, there are less chances of 2 Dual cores going bad then 1 quad core going bad.

but since AMD is going to be pushing out tri and dual core varients, I don't see it as big problem, they just created another market for themselves.

 

[Deliximus]

there is no way B3 willl make Phenoms on par with C2D Conroe. I can see MAYBE a couple of percentage points but that is it.

Hampered by lower IPC (but much better than K8) and poor frequency scaling, AMD NEEDS B3 or 45nm to come on smoothly. Their 65nm is having troubles with k10 (k8 seems alright @ rev g2) they NEED to 45nm to come by nicely.

I will AMD still, Phenom still offers good performance, not the best, but more that average. My rigs will remain Intel free for a long long time. But I am not going to have false hope about k10 cuz AMD execs were spewing BS for more than 6 months about kicking Clovertown but this is a farce. AMD execs credibility with claims are worthless to me now.

All being said, I do think AMD has passed its lowest point. Considering how the 3800s are selling well, low/mid end ATI chipsets (690g and the upcoming 780, 790x) are gonna sell just fine. OEMs will gobble it up as it performs well enough (OEMs are still gonna use X2s) and good on power (think of how well the 2400/2600 radeons sold despite performance gaps against 8600).

 

[Digital Viper-X-]

btw, I think the limited NB/l3 cache speeds are seriously hindering the cpu =p

 

[Scali2]

but since AMD is going to be pushing out tri and dual core varients, I don't see it as big problem, they just created another market for themselves.

Not true ofcourse.
Given your own explanation above that yields are better with MCM, AMD's problem here is that they won't be able to get enough quadcores out, as many of them will have defects and will be tricores or even dualcores.
So they'd have cheaper quadcores simply because they'd be making more of them with MCM.

 

[duby229]

Why don't you show the numbers first?
I don't think anyone has to prove that MCM is cheaper. Intel itself mentioned this, and various sources also claim the same thing.
Eg http://en.wikipedia.org/wiki/Core_2


Or http://www.maximumpc.com/article/Multicore-Madness


This is widely regarded as truth, as it makes perfect sense, and all evidence points to its favour.

So you will have to disprove this common idea of MCM being a cheaper production method first.

Just look at the cost per die. Multiply that by one for AMD and multiply that by two for Intel. Additionally you also have the extra costs of the package itself.....

 

[duby229]

Not true ofcourse.
Given your own explanation above that yields are better with MCM, AMD's problem here is that they won't be able to get enough quadcores out, as many of them will have defects and will be tricores or even dualcores.
So they'd have cheaper quadcores simply because they'd be making more of them with MCM.

Your assuming ver high defect ratio's which cant be known. The fact if the matter is that Intel is manufacturing it's dual core chips that cost less then AMD's quad cores, but countingh two of them together, and then adding the additional packaging costs you end up with a about a 33% higher total cost.

 

[RiverRunsRed]

Duby:

Please do some researches before posting misleading message. Being in chip supply-chain management industry for almost ten years, MCM is cheaper to produce. The current project of mine is to provide planning solution for Intel's next 32-nm chip: Nehalem. Since the process steps are so different, planning system needs to do a lot of modification.

If you can show the numbers then I'll believe you, but in the mean time all of the calculations that I've done show MCM's as costing as much as 33% more. Also there is no way in hell that you can know what yields are. Neither company releases that information, and they both make very sure that they dont release enough supporting data so that it can be calculated. We can make educated guesses, but that is really the best that anybody can claim. Your claim assumes a very high defect ratio, which nobody knows what they actually are. So without that information all you can do in faith is assume the best case scenario. Anything else is being a pessimist.

 

[jvrobert]

Your assuming ver high defect ratio's which cant be known. The fact if the matter is that Intel is manufacturing it's dual core chips that cost less then AMD's quad cores, but countingh two of them together, and then adding the additional packaging costs you end up with a about a 33% higher total cost.

Wow, you like to just totally make up numbers. Sharikoo, is that you? You don't seem to grasp basic semiconductor economics. Seriously, give me a break here.

Phenom: 285 square mm.
Each of the two chips in a Penryn quad core: 107mm.

You do the math, genius. Take even the Q6600. Die size of each one is 143mm^2. So roughly equivelent. Now, one defect for a Q6600 takes out 1/2 of a Q6600. One defect on a Phenom takes out the whole 285mm^2.

Edit: Ooh, sorry. Forgot "tri-core". Haha. So it's not quite as bad as losing the whole chip, but still not good for AMD.

 

[duby229]

Wow, you like to just totally make up numbers. Sharikoo, is that you? You don't seem to grasp basic semiconductor economics. Seriously, give me a break here.

Phenom: 285 square mm.
Each of the two chips in a Penryn quad core: 107mm.

You do the math, genius. Take even the Q6600. Die size of each one is 143mm^2. So roughly equivelent. Now, one defect for a Q6600 takes out 1/2 of a Q6600. One defect on a Phenom takes out the whole 285mm^2.

Edit: Ooh, sorry. Forgot "tri-core". Haha. So it's not quite as bad as losing the whole chip, but still not good for AMD.

Which can then be sold as an X3 or or an X2... Like I sad, your assuming a very high defect ratio that cannot be known.

And of course it's made up... Everybodies numbers are becouse neither Intel or AMD makes this information available. The best that anybody can do is make an educated guess. And while you choose to compare die sizes with a chip that doesnt exist the reality is that right now Intels die is about 143^2mm, not 107. When that day cmes we can then recalculate the numbers and make a new guess at what the values might be. Until then it seems that provided a half decent defect ratio, AMD's monolithic quad core costs less to manufacture.

 

[duby229]

Duby:

Please do some researches before posting misleading message. Being in chip supply-chain management industry for almost ten years, MCM is cheaper to produce. The current project of mine is to provide planning solution for Intel's next 32-nm chip: Nehalem. Since the process steps are so different, planning system needs to do a lot of modification.

If you have them share them. I'd be interested in seeing some hard numbers for once instead of relying on third and forth party speculation.

 

[Scali2]

Your assuming ver high defect ratio's which cant be known.

No I'm not. I'm just assuming that AMD has higher defect ratios than Intel because AMD uses dies that are much larger, and defects are proportional to die area.
So, assuming AMD and Intel have equal defect ratios on their wafers, AMD will have higher defect ratios per die.

Funny thing is that you seem to *ignore* defect ratios and die size in your 'calculations', while they are directly responsible for the die cost, which you base your 'calculations' on.
Yes, we can multiply Intels die cost by 2 for an MCM package. But the genius of MCM is that the cost of a single die is LESS than half of a native quadcode die, because the size is smaller (pretty much exactly half) and the defect ratio is lower. And by such a margin that packaging costs can be compensated.

 

[duby229]

No I'm not. I'm just assuming that AMD has higher defect ratios than Intel because AMD uses dies that are much larger, and defects are proportional to die area.
So, assuming AMD and Intel have equal defect ratios on their wafers, AMD will have higher defect ratios per die.

Funny thing is that you seem to *ignore* defect ratios and die size in your 'calculations', while they are directly responsible for the die cost, which you base your 'calculations' on.
Yes, we can multiply Intels die cost by 2 for an MCM package. But the genius of MCM is that the cost of a single die is LESS than half of a native quadcode die, because the size is smaller (pretty much exactly half) and the defect ratio is lower. And by such a margin that packaging costs can be compensated.

It all comes down to yields, and maybe on these first few batches, yields probably are fairly low, and these first few batches are probably costing an arm and a leg. However it is reasonable assume that yields will improve. Nobody kows what yields are or will be. AMD does not release enough data top make that information available, neither does Intel.

Your assuming great yields for Intel, and horrible yields for AMD... You cant do that. It's not fair, and assumes too much.

 

[Scali2]

Your assuming great yields for Intel, and horrible yields for AMD... You cant do that. It's not fair, and assumes too much.

No I'm not, I'm assuming equal yields (well defects per surface area, not per die, because die-sizes are different) for both (which probably isn't even true, because Intels 65 nm process is more mature).
Read what I said again. Larger die size -> more defects per die, given the same process.
Since you only need 1 defect to get a bad die, smaller dies give higher yields. Everyone in this thread seems to know this, except you.
Basically Intel needs two bad dies to get a bad MCM, so it can have twice as many defect cores to break even.
So yes, yields in terms of dies will be much higher for Intel, because the dies are much smaller, given the same process with the same defect ratio per wafer.

 

[duby229]

No I'm not, I'm assuming equal yields for both (which probably isn't even true, because Intels 65 nm process is more mature).
Read what I said again. Larger die size -> more defects per die, given the same process.
Since you only need 1 defect to get a bad die, smaller dies give higher yields. Everyone in this thread seems to know this, except you.
Basically Intel needs two bad dies to get a bad MCM, so it can have twice as many defect cores to break even.

And what I'm saying is that most of those will still yield as an X2, or an X3. Now Intel has the benefit being able to toss out that bad die, and matching the one next to it with another good die.. So given this it is possible for Intel to yield more quad cores then AMD provided that the defect ratio is the same for both companies.

However you and I both know that is assuming way too much. And that is the point that I'm making. To say that Intel's MCM costs less then a monolithic die is assuming more then what is possible. The only thing that you can do is take the best case scenario, and say that if Intel has a great yield, and AMD has a great yield, then AMD's product will in fact cost less. There simply isnt enough data available to say anything else.

 

[Darakian]

I believe what duby is trying to say is that the production of an MCM with the k8 dual core would have cost more than the phenom quad core for AMD.
Although with a die size of 126 mm^2 (IIRC) for the 65nm x2 and if phenom is indeed 285 mm^2 that argument would rely on unknown packaging costs.

 

[Scali2]

And what I'm saying is that most of those will still yield as an X2, or an X3. Now Intel has the benefit being able to toss out that bad die, and matching the one next to it with another good die.. So given this it is possible for Intel to yield more quad cores then AMD provided that the defect ratio is the same for both companies.

Exactly, we were talking about yields for quadcores. X2 and X3 won't be competitive against a quadcore, and have to be priced lower, bringing in less profit for the same production costs. This is the crux of the matter with native quadcore.

However you and I both know that is assuming way too much. And that is the point that I'm making. To say that Intel's MCM costs less then a monolithic die is assuming more then what is possible.

No it isn't. Everyone in this thread seems to agree on that, as do various sources of which I have already linked a few.

The only thing that you can do is take the best case scenario, and say that if Intel has a great yield, and AMD has a great yield, then AMD's product will in fact cost less. There simply isnt enough data available to say anything else.

Erm, no you can't. You just said so yourself, you end up with more X2s and X3s.

 

[M1dknight]

You seem to be presuming Intel and AMD spend the same amount to make a die with the same area ignoring defects. But if I am not mistaken AMD is still using SOI which according to several sources including Intel is significantly more expensive.

 

[Scali2]

I believe what duby is trying to say is that the production of an MCM with the k8 dual core would have cost more than the phenom quad core for AMD.
Although with a die size of 126 mm^2 (IIRC) for the 65nm x2 and if phenom is indeed 285 mm^2 that argument would rely on unknown packaging costs.

Phenom has only one memory controller. Only one of the two X2 dies would need a memory controller (or both one channel, however you design it, doesn't matter).
So the overall die area would be less than twice that of a X2...
And packaging costs aren't such a big deal. Processor prices, especially in the high end, depend almost solely on yields, and not on production costs.
MCM packaging may add a handful of dollars to each unit at most. Negligible when you're talking about CPUs of hundreds of dollars. I think we can all agree that an MCM package won't add $50 or more to a CPU. That's just complete nonsense.

 

[duby229]

Erm, no you can't. You just said so yourself, you end up with more X2s and X3s.

Well, you cant give Intel the best case scenaro, and give AMD the worst.... As this is exactly what your trying do by claiming that the defect ratio is the same across both companies. If AMD and Intel had the same defect ratio AMD would end up with something like a 50% yeild, which you and I both know is BS.

By the way guys, happy thanksgiving... Enjoy the turkey feast.

 

[Scali2]

Well, you cant give Intel the best case scenaro, and give AMD the worst.... As this is exactly what your trying do by claiming that the defect ratio is the same across both companies. If AMD and Intel had the same defect ratio AMD would end up with something like a 50% yeild, which you and I both know is BS.

What are you trying to say here?
You say that Intel must have worse defect ratios than AMD, because else AMD will come off bad?
Reality doesn't work that way.
With assuming equal defect ratios for Intel and AMD I'm actually *favouring* AMD, because we all know (well, you're still in denial) that Intel's 65 nm process is more mature, and it is highly unlikely that AMD has better defect ratios than Intel at this point (especially considering AMD's product line, with 65 nm not even being used for the faster X2 models, and now Barcelona and Phenom at such lackluster speeds).

 

[duby229]

You seem to be presuming Intel and AMD spend the same amount to make a die with the same area ignoring defects. But if I am not mistaken AMD is still using SOI which according to several sources including Intel is significantly more expensive.

The wafer costs matter a whole lot on the low end, but can largely be compensated --iff- yields are high enough.. In the end when were talking ablout mainstream desktop chips the added cost of SOI is largely taken care of by the shear volume of products that are being sold.

 

[duby229]

What are you trying to say here?
You say that Intel must have worse yields than AMD, because else AMD will come off bad?
Reality doesn't work that way.
With assuming equal yields for Intel and AMD I'm actually *favouring* AMD, because we all know (well, you're still in denial) that Intel's 65 nm process is more mature, and it is highly unlikely that AMD has better yields than Intel at this point (especially considering AMD's product line, with 65 nm not even being used for the faster X2 models, and now Barcelona and Phenom at such lackluster speeds).

The same defect ratio would provide Intel with a much higher yield, because their dies are smaller. You cant assume that. Saying that they have the same defect ratio automatically gives Intel a much higher yield, and you cant do that. You cant say that they have the same defect ratio. If you do that gives AMD a much lower yield then Intel, and that is simply not fair.

 

[Scali2]

The same defect ratio would provide Intel with a much higher yield, because their dies are smaller. You cant assume that. Saying that they have the same defect ratio automatically gives Intel a much higher yield, and you cant do that.

Erm, ofcourse you can!
That's the whole point of MCM!
Yes, Intel has much higher yields. They just need to use twice as many dies for a quadcore, so this negates the much higher yields somewhat.
Still they are above breaking equal, else they wouldn't have stuck with the MCM approach in the first place.

You cant say that they have the same defect ratio. If you do that gives AMD a much lower yield then Intel, and that is simply not fair.

This is about technology, not about being fair.
Was it fair that Athlon64 outperformed Pentium 4, even though it had a smaller die, and higher yields, lower price etc?
Well no, but it happened anyway (and I didn't hear you complain).
The shoe is on the other foot now.

Geez, I knew you were more interested in the name AMD than in the technology itself, but if your best argument is "no fair", you have just reached an all-time low in fanboyism, even for you!

 

[RiverRunsRed]

I don't think I can share the numbers (I don't want to have trouble). Since Nehalem is a MCP product (that is what we call it, MCP project internally), planning needs to be very careful. Cost reduction is the key. That is all I can tell you.

If you have them share them. I'd be interested in seeing some hard numbers for once instead of relying on third and forth party speculation.

 

[duby229]

I don't think I can share the numbers (I don't want to have trouble). Since Nehalem is a MCP product (that is what we call it, MCP project internally), planning needs to be very careful. Cost reduction is the key. That is all I can tell you.

Not a problem I fully appreciate your info though.

 

[duby229]

Not that bad though, what your saying is that AMD has a yield right around 50%. You and I both know that is BS. If it was that low they would have scrapped it, and wouldnt have even bothered trying. Yields that low are simply not sustainable. Shit anything less the 70% is hard to sustain. The truth is that it's probably alot closer to 85%. But nobody knows for sure. The best we can do is put both companies on equal footing and say they have the same yield. In this case AMD's monolithic core costs less.

 

[rantanamo]

How about AMD just go and die and no one use them. That way the Intel fanboys can have sky high prices and probably not even be able to afford their current rigs.

These arguments are crazy. Unless you're playing Crysis or are editing 1080p resolution video can you honestly tell that much of a difference in these top end speeds? Let alone J6P.

 

[Scali2]

Not that bad though, what your saying is that AMD has a yield right around 50%. You and I both know that is BS. The truth is that it's probably alot closer to 85%. But nobody knows for sure. The best we can do is put both companies on equal footing and say they have the same yield. In this case AMD's monolithic core costs less.

What are you talking about with yields?
Percentage of good dies per wafer?
In which case I'm not saying anything like that, you seem to forget that AMD gets a lot less dies out of a wafer, because they're so much larger, hence they do cost more.
If both Intel and AMD would get 85% good dies out of a wafer, that means twice as many good dies for Intel. Which means just as many good quadcores out of a wafer.
However, because AMDs dies are much larger, the yields will be lower by default, as already explained above. With AMD a single defect can spoil an entire quadcore. With Intel you need two defects on two individial dies before an entire quadcore is spoilt. The chances of that happening are simply smaller. Laws of distribution and all that.

See, if we would assume two good quadcores, both the same total size, one MCM and one single-die, then yes, the single-die would cost less, because you have slightly cheaper packaging.
However, the price is mostly determined by the BAD quadcores, not the good ones. And that's where MCM comes out a winner. You throw less bad dies away, and end up with more good quadcores. And this is something you're ignoring. Not that it matters. It's AMD's problem, they're losing money on the Phenom because they're not using MCM, no matter what you say.

 

[duby229]

What are you talking about with yields?
Percentage of good dies per wafer?
In which case I'm not saying anything like that, you seem to forget that AMD gets a lot less dies out of a wafer, because they're so much larger, hence they do cost more.
If both Intel and AMD would get 85% good dies out of a wafer, that means twice as many good dies for Intel. Which means just as many good quadcores out of a wafer.
However, because AMDs dies are much larger, the yields will be lower by default, as already explained above. With AMD a single defect can spoil an entire quadcore. With Intel you need two defects on two individial dies before an entire quadcore is spoilt. The chances of that happening are simply smaller. Laws of distribution and all that.

Not so, becouse all you have to do is look at die size. 143^2mm*2 compared with 284^2mm Provided that both companies have the same yeild, they will both be able to produce the same number of quad cores from a wafer or pretty close to it, plus AMD has the added benefit of binning some defects as X3''s.

In the situation where yields are the same AMD's products do in fact cost less.