Can AMD keep up with Intel at the factory level?

[imported_Questar]

Huh?? If you have a 144 mm2 die that's 90nm, and a 144mm2 die that's 65nm, why would one be less expensive than the other in terms of wafer usage?

So you're saying the only thing that costs money is the wafer. Nothing else in the production of a chip has costs that can be controlled or optimized.

Wow, you've left me speechless.

 

[Viditor]

Originally posted by: Questar

Huh?? If you have a 144 mm2 die that's 90nm, and a 144mm2 die that's 65nm, why would one be less expensive than the other in terms of wafer usage?

So you're saying the only thing that costs money is the wafer. Nothing else in the production of a chip has costs that can be controlled or optimized.

Wow, you've left me speechless.

Nothing that's effected by a process shrink, no.
Sorry to hear about your voice problem...

Edit: If you read a bit you'll notice that most of the Fab costs are fixed, making the only real variable wafer usage. A process shrink does cost money in cap-ex expenditures which is amortized over the life of the equipment, but again that's a fixed cost.

 

[Viditor]

Originally posted by: Dthom

Originally posted by: Viditor

Originally posted by: Dthom

And I've already made the point that intel can make money selling at a lower price than AMD.

I think this is where you're calculations go awry...
While it's true that a process shrink will reduce the amount of the wafer used, this is only the case if they manufacture the exact same design.

With a 2MB cache, the size of the cache is almost half of the die...can you imagine how big a 4MB cache makes the die?

Take a look at how much they get for the big cache chips? Its not like they're going to be selling 4MB chips for $100. They get their pound of flesh and then some for the bigger caches. Plus the bigger caches give then an even bigger advantage performance-wise, so they can charge even more now. Before, AMDs were still beating them with the bigger cache, so they couldn't price the parts over what AMD was charging. Now that AMD has nothing to compare, the moon is the limit on what they can get for huge cache chips, at least in the big-bucks server market.

You have to understand that before the new releases, the bigger caches only closed the performance gap. They were still restrained price-wise. When you have the fastest cpu available (see the price of high-end dual-core opterons), the pricing can go way up, since you have no competition.

I understand what your saying, but the issue is one of cost savings purely from a die shrink...
While they aren't getting a pittance for the chips, they aren't getting a price PREMIUM for them either. This was the whole point here...I was pointing out that a node shrink doesn't ALWAYS reduce the costs, and for the beginning of the ramp it actually increases them sometimes.

 

[imported_Questar]

Edit: If you read a bit you'll notice that most of the Fab costs are fixed, making the only real variable wafer usage.

Have you provided this insight to AMD, Intel, TI, UMC or any other chip manufacturer? I'm sure they would like to know that they don't need thousands of process and manufacturing engineers working to gain production efficiencies.

 

[Viditor]

Originally posted by: Questar

Edit: If you read a bit you'll notice that most of the Fab costs are fixed, making the only real variable wafer usage.

Have you provided this insight to AMD, Intel, TI, UMC or any other chip manufacturer? I'm sure they would like to know that they don't need thousands of process and manufacturing engineers working to gain production efficiencies.

No, but I'll be sure to contact the employees there and let them know that you've decided that their jobs are no longer a fixed cost and that they are subject to being fired on a moments notice...

 

[imported_Questar]

Originally posted by: Viditor

Originally posted by: Questar

Edit: If you read a bit you'll notice that most of the Fab costs are fixed, making the only real variable wafer usage.

Have you provided this insight to AMD, Intel, TI, UMC or any other chip manufacturer? I'm sure they would like to know that they don't need thousands of process and manufacturing engineers working to gain production efficiencies.

No, but I'll be sure to contact the employees there and let them know that you've decided that their jobs are no longer a fixed cost and that they are subject to being fired on a moments notice...

Nice try. I'm not the one who said nothing but a process shrink can reduce cost.

BTW, employee costs are variable - not fixed.

But I wouldn't expect any greater understanding from someone who said that the IPC on an Opteron increases with clock speed.

 

[the Chase]

Well if some speculation is correct AMD may be saving some die(wafer) space to increase their capacity without a die shrink. Notice the AT price cut charts had the new prices in red. But the 1MB cores had nothing in that slot? The FX CPU's had the same price in red in that slot but not anything for the 4000, 4400, and 4800 models. So could it be that these models are discontinued? That would save AMD some wafer space. Nothing bigger than that 147mm sqaured figure. (Except the 2 FX CPU's). http://www.dailytech.com/article.aspx?newsid=2800

 

[coldpower27]

Originally posted by: Viditor

Originally posted by: Questar

Regarding margin, AMDs margins are up because they currently have an advantage. When Intel cuts prices AND has a performance advantage, AMDs margins will disappear altogether.

I completetly agree with you. At 90nm it Cost Intel $40 to manufacture a chip. At 65nm it's considerably less. AMD's costs are nearly double this.

This is quite incorrect...at 65nm, Intel reduced the die on Presler by 21% but increased costs in lower percentage yields and the capital outlay of new equipment. Die size is all that counts for cost...and these are:

Manchester W/1MB = 147 mm2
Toledo W/2MB = 199 mm2
Presler W/2MB = 162 mm2
Conroe W/2MB = 144 mm2
Conroe W/4MB = ?

While it's nice that your quoting Manchester and Toledo. You might want to fix that charge up a bit.

Venice = 84mm2
Manchester w/2x512KB = 147mm2
Toledo w/2x1MB = 199mm2
Orleans = 103mm2
Windsor-512 w/2x512KB = 183mm2
Windsor w/2x1MB = 230mm2

Presler w/2x2MB = 2x81mm2
Cedar Mill = 81mm2

With Conroe there are no figures to suggest what the die size will be the current estimates place the Conroe core with it's full 4MB of cache in the 140mm2 range.

Allendale which is the same core with 2MB will undoubtly be smaller.

http://www.matbe.com/images/biblio/art_...4-x2-face-au-core-duo/000000038969.jpg

That's the 172mm2 measured by the site Taylor core vs the Yonah 90.3mm2 core. This gives you a sorta visual comparison ofhow big each core is in relation to each other.

Here is a pciture of the T7500 Merom die.

http://i83.photobucket.com/albums/j306/hicookie-ke3/DSC05213.jpg

Here is a picture of the T7400 Merom die.

http://coolaler.kj.idv.tw/merom/QLZT/1.JPG

You can see it is doubtful the Merom with it's 4MB of cache is anywhere close to the 172mm2 of Taylor.

Keep in mind these are the same size as the Conroe cores so Allendale I doubt is going to be as large as 144mm2

Your also forgetting to mention Intel reduce the die size by 21% but in the process added an additional 2MB of cache. It wasn't a straight die shrink.

In terms of processor costs the die counts, but you also have to add in packaging and such too. How complex a process is, AMD has SOI, DSL, and for 65nm looking to add SIGe, how many layers the process is.

 

[paperfist]

AMD has never been able to keep up with Intel at the factory level, yet AMD is still giving Intel a very hard time

 

[coldpower27]

Originally posted by: the Chase
Well if some speculation is correct AMD may be saving some die(wafer) space to increase their capacity without a die shrink. Notice the AT price cut charts had the new prices in red. But the 1MB cores had nothing in that slot? The FX CPU's had the same price in red in that slot but not anything for the 4000, 4400, and 4800 models. So could it be that these models are discontinued? That would save AMD some wafer space. Nothing bigger than that 147mm sqaured figure. (Except the 2 FX CPU's). http://www.dailytech.com/article.aspx?newsid=2800

Unfortuately, however you have to keep in mind 147mm2 only refers to the Manchester core Athlon 64x2, the new and going forward Windsor-512 is 183mm2.

I think AMD is just going to leave the 2x1MB core as they are and see if any suckers will buy them.

 

[coldpower27]

Originally posted by: Viditor

Originally posted by: Dthom

And I've already made the point that intel can make money selling at a lower price than AMD.

I think this is where you're calculations go awry...
While it's true that a process shrink will reduce the amount of the wafer used, this is only the case if they manufacture the exact same design.

With a 2MB cache, the size of the cache is almost half of the die...can you imagine how big a 4MB cache makes the die?

2MB of Cache is about 40-45% of a die for Cedar Mill

On Conroe from what you can see here:
http://images.anandtech.com/reviews/tra.../2006/IDF/Keynote1/fiveinnovations.jpg

4MB of cache represents about 40-45% of the die, since Conroe is a Dual Core processor. The cahce to core logic ratio has generally remained at aorund this level since Intel introduced Prescott-2M and then later Presler.

4MB of cache isn't that big of a deal for Intel with it's great cache densities combined with it's 65nm process on a Single Socket. It's actually quite a viable level of production.

 

[BasMSI]

Very nice arguments.....
But I have a few doubts about all of this.

We know the Conroe can't be bought, nor be tested in reviewers own labs.....what does this teach us?
Exactly, nothing.

We know that the AM2's are handicaped by the current DDR2 memory-modules.....what does this teach us?
Exactly, better low-latency memory modules need to be fabricated.

So the point is, all this balony is about a CPU that can't be bought or tested real good compared to a CPU that isn't able to perform at it's maximum level because of memory.

Ergo, there can't be said anything about Intel or AMD compared to performance and such.
Heck we all know that socket939 is faster because of low-latency memory access, why not compare them all?

Funny, just look here and see the S939 is faster: http://www.tomshardware.com/2006/05/23/amd_reinvents_itself/page29.html

 

[coldpower27]

Originally posted by: BasMSI
Very nice arguments.....
But I have a few doubts about all of this.

We know the Conroe can't be bought, nor be tested in reviewers own labs.....what does this teach us?
Exactly, nothing.

We know that the AM2's are handicaped by the current DDR2 memory-modules.....what does this teach us?
Exactly, better low-latency memory modules need to be fabricated.

So the point is, all this balony is about a CPU that can't be bought or tested real good compared to a CPU that isn't able to perform at it's maximum level because of memory.

Ergo, there can't be said anything about Intel or AMD compared to performance and such.
Heck we all know that socket939 is faster because of low-latency memory access, why not compare them all?

Funny, just look here and see the S939 is faster: http://www.tomshardware.com/2006/05/23/amd_reinvents_itself/page29.html

The thing is Lower Latency modules will help both parties, Conroe and Windsor, so it's likely whatever performance Windsor gains from it will be negated by Conroe gains from the same memory.

 

[daballard]

Originally posted by: BasMSI
Very nice arguments.....
But I have a few doubts about all of this.

We know the Conroe can't be bought, nor be tested in reviewers own labs.....what does this teach us?
Exactly, nothing.

We know that the AM2's are handicaped by the current DDR2 memory-modules.....what does this teach us?
Exactly, better low-latency memory modules need to be fabricated.

So the point is, all this balony is about a CPU that can't be bought or tested real good compared to a CPU that isn't able to perform at it's maximum level because of memory.

Ergo, there can't be said anything about Intel or AMD compared to performance and such.
Heck we all know that socket939 is faster because of low-latency memory access, why not compare them all?

Funny, just look here and see the S939 is faster: http://www.tomshardware.com/2006/05/23/amd_reinvents_itself/page29.html

Very nice, but this thread was not comparing the AM2 to Conroe. The discussion on hand is manufacturing capabilities.

 

[Fox5]

Originally posted by: coldpower27

Originally posted by: the Chase
Well if some speculation is correct AMD may be saving some die(wafer) space to increase their capacity without a die shrink. Notice the AT price cut charts had the new prices in red. But the 1MB cores had nothing in that slot? The FX CPU's had the same price in red in that slot but not anything for the 4000, 4400, and 4800 models. So could it be that these models are discontinued? That would save AMD some wafer space. Nothing bigger than that 147mm sqaured figure. (Except the 2 FX CPU's). http://www.dailytech.com/article.aspx?newsid=2800

Unfortuately, however you have to keep in mind 147mm2 only refers to the Manchester core Athlon 64x2, the new and going forward Windsor-512 is 183mm2.

I think AMD is just going to leave the 2x1MB core as they are and see if any suckers will buy them.

Why did Windsor get bigger?

 

[the Chase]

Originally posted by: coldpower27

Originally posted by: the Chase
Well if some speculation is correct AMD may be saving some die(wafer) space to increase their capacity without a die shrink. Notice the AT price cut charts had the new prices in red. But the 1MB cores had nothing in that slot? The FX CPU's had the same price in red in that slot but not anything for the 4000, 4400, and 4800 models. So could it be that these models are discontinued? That would save AMD some wafer space. Nothing bigger than that 147mm sqaured figure. (Except the 2 FX CPU's). http://www.dailytech.com/article.aspx?newsid=2800

Unfortuately, however you have to keep in mind 147mm2 only refers to the Manchester core Athlon 64x2, the new and going forward Windsor-512 is 183mm2.

I think AMD is just going to leave the 2x1MB core as they are and see if any suckers will buy them.

http://www.dailytech.com/article.aspx?newsid=2858 -well it will still spare some wafer space by not going with the 1MB per core chips. But yeah- what got bigger?

 

[coldpower27]

Originally posted by: the Chase

Originally posted by: coldpower27

Originally posted by: the Chase
Well if some speculation is correct AMD may be saving some die(wafer) space to increase their capacity without a die shrink. Notice the AT price cut charts had the new prices in red. But the 1MB cores had nothing in that slot? The FX CPU's had the same price in red in that slot but not anything for the 4000, 4400, and 4800 models. So could it be that these models are discontinued? That would save AMD some wafer space. Nothing bigger than that 147mm sqaured figure. (Except the 2 FX CPU's). http://www.dailytech.com/article.aspx?newsid=2800

Unfortuately, however you have to keep in mind 147mm2 only refers to the Manchester core Athlon 64x2, the new and going forward Windsor-512 is 183mm2.

I think AMD is just going to leave the 2x1MB core as they are and see if any suckers will buy them.

http://www.dailytech.com/article.aspx?newsid=2858 -well it will still spare some wafer space by not going with the 1MB per core chips. But yeah- what got bigger?

Assuming hey do terminate production on the "full" Windsor cores entirely.

The new Socket AM2 cores increase in die size, but have not been explain throughly why, my guess is for reduced transistor density, so clockspeed can now be ramped a tad higher before 65nm, and also the new memory controller is more complex in addition to the Pacifica Viritualization technology.

Well we do know something did get bigger.