Wednesday, December 27, 2006

Real test for 64bit CPU

Can your desktop chip handle 4GB? Some can't, some can.

Charlie at INQ says AMD frags at least 75% of Intel. Sounds familiar? But, Charlie is missing the big picture here. What we see in 2006 is that AMD is capacity constrained. With FAB36 starting its flood of 65nm chips, and with AMD's cost well below Intel's, AMD is going to deny Intel's oppurtunity for profit, not the other way around.

Microsoft is demoing Vista 64 on AMD notebooks.

AMD should make a high clock speed single core Athlon64. A 3GHZ single core is better than a 2GHZ dual core in most cases.

55 Comments:

Blogger Scientia from AMDZone said...

Well, your title is misleading because the problem is not the processor but the chipset. All Intel EM64T processors can handle at least 32GB's of memory and all AMD64 procesors can handle at least 512GB's.

I would disagree with a couple of Charlie's assertions. Intel is not having much effect in terms of taking revenue from AMD; 2006 will be AMD's highest grossing year in its history in terms of processors. Intel is having more success taking revenue from itself. 2006 will be the largest processor revenue drop for Intel since 2001. Intel's gross revenue for processors will be lower than 2005 and only about as high as 2004. If Intel has momentum other than replacing its own processors it isn't showing up. Intel still lost volume share in the 3rd quarter after releasing C2D.

6:14 PM, December 27, 2006  
Anonymous Anonymous said...

Booting or not with 4GB of RAM means nothing. Could it be the RAM itself is the issue, maybe the BIOS. Either way, i doubt ANY desktop AMD chip on the market with 4 GB out performs Kentsfield, or E6700 for that matter with 2 GB in any meaningful bench or real world application

6:59 PM, December 27, 2006  
Anonymous Anonymous said...

I read Charlies article also and all his conclusions were wrong.
Every chip intel has introduced in the last year has been tanking bad.
If the chip was not too bad the platforms are just power sucking junk.
It appears intel will loose its 13 year number 1 sales spot as AMD simply pulls more and more market share from intel.
It does not seem that intel has been able to alter there 6 year market free fall.
Intels stock is also tanking under the continuing pressure from AMD.
AMD will have another banner year with unlimited growth.
When the intels last quarter of 2006 sales figures are released they will be followed by ceo resignation.

7:05 PM, December 27, 2006  
Anonymous enumae said...

Sharikou said...

"Charlie at INQ says AMD frags at least 75% of Intel. Sounds familiar?"

One more time in another light, for every 100 processors made...

AMD = 25, and Intel = 75

Of the 75 Intel makes, 18.75 or 25% are Core 2 based

Of the 25 that AMD makes, 6.25 or 25% can beat Intel processors.

This leaves 50 out of every 75 processors Intel makes without competition from AMD.

Unless you know something I don't, how do you compete with out products?

"But, Charlie is missing the big picture here. What we see in 2006 is that AMD is capacity constrained. With FAB36 starting its flood of 65nm chips, and with AMD's cost well below Intel's, AMD is going to deny Intel's oppurtunity for profit, not the other way around."

AMD has said it themselves, they plan on having the ability to produce 30% of the market leaving 2007 and 40% leaving 2008.

Assuming that AMD gets to 30% capcaity and market share by the middle of 2007, and knowing that Intel got Core 2 to about 25% production in 6 months, we will assume that it continues on that path and we will see 50% by about the middle of 2007.

Look at the numbers again...

AMD = 30, and Intel = 70

Of the 70 Intel makes, 35 or 50% are Core 2 based, which will be enough to match all of AMD's capacity.

Simply looking at the numbers AMD is in trouble, and will not rebound no matter how much capcity they have unless they have a superior product which would be around the middle of next year, but if K8L dual core and quad core variants can not establish the performance lead or be competitive against Intel's Penryn processors, AMD's market share will stall and slowly reverse until they release there next generation processor.

8:19 PM, December 27, 2006  
Blogger Sharikou, Ph. D said...

One more time in another light, for every 100 processors made...

AMD = 25, and Intel = 75

Of the 75 Intel makes, 18.75 or 25% are Core 2 based

Of the 25 that AMD makes, 6.25 or 25% can beat Intel processors.


You need to read my analysis of compnay IA, II and AI to straighten your thinking.

8:44 PM, December 27, 2006  
Anonymous Anonymous said...

Please tell us more about this flood - the chips are only 33% smaller (not 50%!) - just out of curiosity, do you know why that is? And given the bulk of AMD's capcity continues to be 90nm (30KWSPW on 200mm is the equiv of ~15K on 300mm), where exactly is this 65nm flood flood? Even when AMD is at 15K 65nm in F36, it still will only be 50% converted to 65nm!!!

How about another good old fashion capacity analysis? Or did you take enough of a pounding just 1 week ago on this...

9:11 PM, December 27, 2006  
Anonymous Anonymous said...

That 75% analysis is terribly misleading. Som finer details:

Intel will be at 75% Core 2 on server by end of THIS year... this is the cash cow for AMD and other than 4P+, AMD will have serious competition in this area for the first time in a long while - notice Intel picked up significant server share in Q3... (and also keep in mind 4P+ segment represents <20% of overall server market).

So where is AMD "fragging" the old Intel products? Yup, on the high margin desktop market (which has been decreasing in size).

And Intel, due to the 20% Core 2 prodcuction, has caused AMD to cut FX prices by >20% on AM2 and do a buy one get 1 free deal for 4x4 chips. So AMD will be fragging away on chips that go for <$300 and low margins... While Intel will continue to own the 500-1000 CPU market.

As for mobile, Intel's "old" gen (Core) is competitive with AMD's newest so conversion to Core 2 is less critical, then say server space.

By Q2'07 Core 2 will be >50% (aggregate) CPU production and probably close to 90+% in server space; what you fail to miss is that when AMD finally gets their magic "65nm" on line they will not be fragging 75% of Intel's production... I put 65nm in quotes as it is only providing ~30% die area scaling so it is not even clear if this really is a 65nm process...

Oh and according to you it takes a quarter to package the chips over at AMD, so add a quarter on to the timelines as AMD's capacity #'s are wafers out (not CPU's out)...

9:30 PM, December 27, 2006  
Anonymous Anonymous said...

"Can your desktop chip handle 4GB?"

Why does it need to?

Unless you are planning to throw away 1GB just to support a bloated MS OS which already has security holes reported, why would 95% of the desktop even need 4GB?

Any guesses on the average memory size today in desktops?

9:41 PM, December 27, 2006  
Anonymous enumae said...

Sharikou said...

"You need to read my analysis of compnay IA, II and AI to straighten your thinking."

Do you call your analysis un biassed?

I have read your post, and re-read it agin tonight, and I still believe the comments I made then.

It is your theory, nothing more, look down the page for Scientias post about servers and Sun.

The numbers are there, and OEM's went AMD for the performance and likely the future ability to supply them, all of which are (in my opinion) a way to get lower prices from Intel due to a competitive product.

Like I said that could all change if the next generation products (K8L Rev H) do not compete with Intels 45nm based processors let alone Core 2 based processors.

We will wait and see.

PS: You say your a journalist, yet almost every story that includes Intel is really just a way for you to bad mouth a large company that you do not like.

Is that journalism or a way to draw hits on your site?

9:41 PM, December 27, 2006  
Blogger Sharikou, Ph. D said...

Do you call your analysis un biassed?


Why can't you do some real analysis instead of firing empty allegations of bias? The exercise is simple, just put all Intel and AMD processors on the table, each with their unit share numbers. I made it easy for you by separating the CPUs in 3 groups, II, IA, and AI.

11:21 PM, December 27, 2006  
Anonymous Anonymous said...

"expect AMD's 65nm parts to see 40% performance improvement over 90nm."

posted by Sharikou, Ph. D @ 7/29/2006 06:32:00 PM

Well???

12:40 AM, December 28, 2006  
Anonymous Anonymous said...

"I reiterate my projection that Intel will post operating loss starting 3Q06, and BK in five to seven quarters."

posted by Sharikou, Ph. D @ 7/27/2006 05:36:00 PM

Well???

12:43 AM, December 28, 2006  
Anonymous Anonymous said...

"The fact that only AMD's FX CPU has 2x1MB cache indicates that AMD's yield is near 100%. Basically, AMD puts in a wafer and says let them be FX62s, and they all become FX62s. With APM3.0, AMD is super efficient."

posted by Sharikou, Ph. D @ 7/20/2006 06:21:00 PM

CLASSIC!!! And on the seventh day Ruiz rested...LOL!!!

12:50 AM, December 28, 2006  
Anonymous Anonymous said...

"Looking forward a few months, 4x4, 65nm, K*L, Bulldozer. Intel won't be able to walk out 2007 in one piece."

posted by Sharikou, Ph. D @ 7/18/2006 09:38:00 AM

WELL???

12:55 AM, December 28, 2006  
Anonymous Anonymous said...

"AMD64 4x4 ready to frag Conroe XE by 60%"

posted by Sharikou, Ph. D @ 7/17/2006 01:49:00 PM

Man oh man this gets better....

12:57 AM, December 28, 2006  
Anonymous Anonymous said...

"With 65nm parts, I expect AMD to achieve a clockspeed increase of 25%, sufficient to decisively defeat Conroe XE 6800 in all benchmarks by brute force. Coupled with architectural enhancements with K8L, AMD may achieve a whopping 60% performance/core advantage over Core2 on integer performance. On floating point, the K8L core will be at least 2x faster than Core2."

posted by Sharikou, Ph. D @ 8/18/2006 01:12:00 PM

You really don't have a clue do you doc.

1:08 AM, December 28, 2006  
Blogger Tom said...

Why can't you do some real analysis instead of firing empty allegations of bias? The exercise is simple, just put all Intel and AMD processors on the table, each with their unit share numbers. I made it easy for you by separating the CPUs in 3 groups, II, IA, and AI.

Look who is talking. You ”Analysis” were wrong then and still are. There are no II and IA company there is one Intel. Please try read the comments there were enough comments back then as to why you were wrong and still are.

1:49 AM, December 28, 2006  
Blogger Scientia from AMDZone said...

enunmae said

AMD = 25, and Intel = 75


Actually, AMD is at 23.5% and Intel is at 76.5%.

Of the 75 Intel makesi, 18.75 or 25% are Core 2 based

Yes, basically 19%.

Of the 25 that AMD makes, 6.25 or 25% can beat Intel processors.

Where did you get this number from? Why would you say that 25% of AMD's can beat Intel processors? Out of all of the P4 based processors that Intel makes only the Celeron, Mobile Celeron, and Tulsa MP processors are competitive with K8. Since Tulsa MP is low volume we'll just say it is 1 out of 76. I would guess that 1/3rd of Intel's P4's are Celerons so that would leave 2/3rds of 76% or 51. So, this would mean that K8's are better than 51% of Intel's processors. But, this doesn't matter much at the moment because AMD doesn't have the capacity to produce more K8's.

This leaves 50 out of every 75 processors Intel makes without competition from AMD.

No. This leaves 19 out of every 76 processors that Intel makes fairly well off. 19 out of 76 are about the same. 38 out of 76 are less competitive.

So, AMD's 24 have to compete with 19 about the same and 38 that are worse.

2:32 AM, December 28, 2006  
Blogger Scientia from AMDZone said...

Okay, now let's look at your second argument:

Look at the numbers again...

AMD = 30, and Intel = 70

Of the 70 Intel makes, 35 or 50% are Core 2 based, which will be enough to match all of AMD's capacity.


This again makes no sense. By the end of 2007 AMD will be at about 27% volume which would put Intel at 73%. By that point, almost nothing that Intel produces would be P4 based, probably just 4-way and up servers and perhaps some Celerons. However, also by that time AMD's server production should be mostly converted to K8L and at least the mid to upper desktop range.

If the Celeron's are still P4 by then then only AMD's lower K8 range above Sempron would not be competitive. This is perhaps 17% of AMD's chips. Of course, 17% would be less than half of what Intel has now so this doesn't seem too bad.

2:48 AM, December 28, 2006  
Blogger Scientia from AMDZone said...

the chips are only 33% smaller (not 50%!) - just out of curiosity, do you know why that is?

Yes; it is because AMD has not fully optimized the transistors for 65nm yet so it is only a partial shrink.

Even when AMD is at 15K 65nm in F36, it still will only be 50% converted to 65nm!!!

No. Mid 2007, when AMD is at 15K on 65nm, FAB 30 will already have starting dropping down and will only be about 1/3. So, AMD would be about 67% converted. I've seen estimates as high as 70% but this is close enough.

2:58 AM, December 28, 2006  
Anonymous enumae said...

Sharikou said...

"Why can't you do some real analysis instead of firing empty allegations of bias?"

The analysis is shown above.

Without bias.

Just simple numbers.

I understand that you don't like what is being said about AMD.

The simple, mathmatical fact still remains, AMD can only compete against 33% of Intels capacity.

Leaving 50% of the worlds processors without competition.

This is the area AMD would have had a chance to gain market share, but without capcaity to expand fast enough, and with Core 2 ramping there market share will stall somewhere around 27%.

Just my biased opinion...

6:23 AM, December 28, 2006  
Blogger TheKhalif said...

Well, your title is misleading because the problem is not the processor but the chipset. All Intel EM64T processors can handle at least 32GB's of memory and all AMD64 procesors can handle at least 512GB's.

I would disagree with a couple of Charlie's assertions. Intel is not having much effect in terms of taking revenue from AMD; 2006 will be AMD's highest grossing year in its history in terms of processors. Intel is having more success taking revenue from itself. 2006 will be the largest processor revenue drop for Intel since 2001. Intel's gross revenue for processors will be lower than 2005 and only about as high as 2004. If Intel has momentum other than replacing its own processors it isn't showing up. Intel still lost volume share in the 3rd quarter after releasing C2D.



True, true, true. I tried to tell the people on a forum that Intel is shooting themselves in the foot with this crazy "price war"

They can't stop AMD from growing. They can't grow. That's a problem with being the market leader.

A recent story shows that HP and Dell grew slower than the market in laptops, while the much smaller Fujitsu managed 49% growth.

Thsi is mirrored with Intel/AMD. And now they are saying they will destroy C2D with C2Q next year.

They are releasing a slower quad next month for less than the X6800. This will create confusion amongst OEMs as they try to get rid of NetBurst.

You can't even have Netburst next to Core 2 in a store as no one will buy Netburst after opening a program or two on each.


My saying for Intel in 2007 is "The Bleeding Continues."

7:49 AM, December 28, 2006  
Blogger TheKhalif said...

Of the 70 Intel makes, 35 or 50% are Core 2 based, which will be enough to match all of AMD's capacity.

Simply looking at the numbers AMD is in trouble, and will not rebound no matter how much capcity they have unless they have a superior product which would be around the middle of next year, but if K8L dual core and quad core variants can not establish the performance lead or be competitive against Intel's Penryn processors, AMD's market share will stall and slowly reverse until they release there next generation processor.



Though Sharikou gets a little excited, you are dead wrong with this statement. The market leader CAN'T match the growth of the smaller company. The smaller company can only grow especially in this case where lots of OEMs would stick it to Intel if they could.

7:56 AM, December 28, 2006  
Blogger TheKhalif said...

Intel will be at 75% Core 2 on server by end of THIS year... this is the cash cow for AMD and other than 4P+, AMD will have serious competition in this area for the first time in a long while - notice Intel picked up significant server share in Q3... (and also keep in mind 4P+ segment represents <20% of overall server market).

Anand's latest tests show that AMD STILL owns 2P in terms of power use (possibly because of FBDIMMS) and unless you saturate the CPU, 5160 just barely wins. A 3GHz Opteron (the same clock as 5160) would make a difference.

Even Clovertown x 2 can't take the top TPC-H slot from Opteron 4P (same number of cores). Tulsa couldn't do it either. AMDs server market is safe.

If Barcelona does get the 70% they are reporting, no Intel 4P will even come close.

I actually expect 80% for quad and 25% for dual "Cities."

8:09 AM, December 28, 2006  
Blogger Sharikou, Ph. D said...

If Barcelona does get the 70% they are reporting, no Intel 4P will even come close.


Clovertown will be Intel's last sign of life, even with 2 Clovertowns, Intel is slower on FP performance at 2P level. Yes, 4 Opteron cores beat 8 Clovertown cores in FP performance. Once K8L is out, game over for Intel.

10:12 AM, December 28, 2006  
Anonymous enumae said...

Scientia said...

"Where did you get this number from? Why would you say that 25% of AMD's can beat Intel processors?"

I am not understanding what is so difficult about this, maybe I am not making this very clear, but I will try again...

If the total market for semiconductors was 100, of that 100 (using current marketshares) Intel would be selling 75, and AMD would be selling 25.

If Core 2 is 25% of Intels total production, that would equate to 18.75 processors.

If AMD is only making 25 processors they can only compete against 25 of Intels processors.

This means AMD can only beat (AMD 25 - Core 2 18.75 = 6.25) 6.25 processors made by Intel, again leaving 50 processors with out competition from AMD.

AMD can not beat 75% of Intels products if they themselves do not have a product to compete against Intel with.

10:32 AM, December 28, 2006  
Anonymous enumae said...

The Khalif said...

"...The market leader CAN'T match the growth of the smaller company..."

Intel can ramp production of Core 2 to match AMD's total capacity faster than AMD can increase its own capacity.

10:44 AM, December 28, 2006  
Blogger Dr. Yield, PhD, MBA said...

Some anonymous feller wrote:
the chips are only 33% smaller (not 50%!) - just out of curiosity, do you know why that is?

To which Scientia opined:
Yes; it is because AMD has not fully optimized the transistors for 65nm yet so it is only a partial shrink.

Scientia, you are confusing two different issues in a node shrink: density and transistor dimension. In logic, 65nm node is an indicator not of transistor size, but contacted Metal 1 half-pitch (see the ITRS roadmap). Associated with this node is a targeted gate CD (critical dimension) of 25nm (42nm in resist).

Everyone here seems to agree that AMD's 1st revision of the 65 node process did not scale gate CD down to the target CD- this would explain the lack of improved performance (larger gates switch slower). That said, it does not explain the lack of density improvement. True 65nm node means that the metal half pitch has shrunk, and die size scale accordingly.

Why does AMD not have a true 50% node shrink benefit? A couple of possibilities:
1. They aren't running a true 65nm process, but rather something larger.
2. The shrink of the layout introduced speed path issues that could not be worked out without rerouting in a fashion that consumed more area. Not likely for such a big delta, but maybe part of the issue.
3. Designers added more circuitry/functionality and aren't telling you. This is possible, for example Intel shipped P4s with EMT64 disabled- but it was there, taking up space.

I don't know why AMD didn't get better scaling, but I can say that the increased area certainly consumes some of the benefit of a node shrink, particulary due to the substantial increase in processed wafer cost for 65nm (200->300mm, plus I believe extra metal layer(s)).

11:12 AM, December 28, 2006  
Blogger Sharikou, Ph. D said...

Designers added more circuitry/functionality and aren't telling you. This is possible, for example Intel shipped P4s with EMT64 disabled- but it was there

Intel shipped P4 with broken EM64T... It took Intel another year to make it marginally compatible with AMD64.

11:17 AM, December 28, 2006  
Blogger Dr. Yield, PhD, MBA said...

Scientia opined:
However, also by that time AMD's server production should be mostly converted to K8L and at least the mid to upper desktop range.


and our humble host weighed in with:
Once K8L is out, game over for Intel.

To wit gentlemen- as I noted in the last treatment of AMD's capacity, K8L simply can't address that much of the server market space until AMD has drastically ramped their 65nm capacity. If you will be so kind as to refer back to my earlier postings, I noted that due to the increased die size of K8L and the accordant yield loss due to said size, AMD consumes roughly a minimum of 3 2core 65nm die capacity to produce a single K8L die.

Implications? Well, Intel is indicating that by the time K8L ships, it will be shipping >40% of its server CPUs as 4core. Can AMD produce that substantial a volume of K8Ls? Not without cratering the rest of their 65nm production. Additionally, if the ASP of a 65nm X2 is $250 (disclaimer, number pulled out of thin air), the every K8L will need to sell for >$750 to make it worthwhile. Should be an easy choice then, no? Yes- unless it means that AMD has to tie up enough capacity that it has issues supplying other market segments with pre-existing commitments. As I've said before- Intel is pushing hard on quadcore more to put the screws to AMDs capacity than for any other reason. Any way you cut it, Intel will still be feeding at the profit trough in servers, even after K8L comes out.

11:33 AM, December 28, 2006  
Blogger Dr. Yield, PhD, MBA said...

Our humble host spewed:
Intel shipped P4 with broken EM64T... It took Intel another year to make it marginally compatible with AMD64.

Completely irrelevant to the point. Intel consumed space that reduced the value of the shrink. No one (officially) knew it was there. Has AMD done something similar or not? If so, less concern about the lack of die scaling. If not, what are they doing wrong? Your comment is just a bunch of anti-Intel spew that does nothing to address the question at hand. Have you no desire to learn something? Not very PhD-like...

11:36 AM, December 28, 2006  
Blogger Sharikou, Ph. D said...

To wit gentlemen- as I noted in the last treatment of AMD's capacity, K8L simply can't address that much of the server market space until AMD has drastically ramped their 65nm capacity.

Chartered FAB7 will start 65nm production by the time K8L is out. I estimate that AMD will take 90% of the server market with K8L, and Chartered capacity is just enough for that.

11:37 AM, December 28, 2006  
Anonymous netrama said...

enumae said .....blah blah
AMD can not beat 75% of Intels products if they themselves do not have a product to compete against Intel with.


Looks like your logic train is a bit late. Now, we know that AMD products are far better then Intel's Netburst Products. In the terms of the numbers this line can be described as AMD's Products are better than 75% (Aka Netburst) of Intel's production. why do you find this so difficut to grasp ??

For anybody shopping around in that 75% crappy intel space , AMD is the clear choice. And the bleeding for Intel will continue through 2007 as 'the khalif' has said ..

11:48 AM, December 28, 2006  
Anonymous Anonymous said...

"However, also by that time AMD's server production should be mostly converted to K8L"

Scientia, in general I like your analysis. But this projection is a bit off - if nothing else there will be a sizeable amount of production for customers who want to maintain K8 and do not want to change existing infrastructure.

Also the desktop designs are not expected until late 07, so the penetration of K8 in this space in 2007 will be minimal. Keep in mind AMD "released" 65nm product in Dec (and on the other side of the coin Intel "released" COnroe/Merom in July). AMD EE processors? Mid-07 AMD release refers only to server (AMD has continued to say this although at time some folks have misinterperted this to mean desktop too)

Unfortunately these days for both Intel and AMD release means small availability and a min of 1-2 quarters to get appreciable volumes.

K8 will have decent penetration on server (assuming no issues/delays) by end of year, but desktop volume will be near or less than 10% in Desktop space in '07. It will be a PR maneuver to say look we can compete (kind of like 4x4 "platformance", except this product will likely actually deliver!)

12:23 PM, December 28, 2006  
Anonymous Anonymous said...

" I estimate that AMD will take 90% of the server market with K8L"

Wow, another crazy prediction.

12:29 PM, December 28, 2006  
Anonymous Anonymous said...

"No. Mid 2007, when AMD is at 15K on 65nm, FAB 30 will already have starting dropping down and will only be about 1/3. So, AMD would be about 67% converted. I've seen estimates as high as 70% but this is close enough."

Actually as I look at the graph (page 5)
http://www.amd.com/us-en/assets/content_type/DownloadableAssets/DarylOstranderAMDAnalystDay.pdf

I think AMD has made this into equivalent wafer outs correcting for 200mm/300mm size delta.

The reason I say this is if you look at the Q2'07 data point it show equivalent 200mm and 300mm wafer outs. As the Q2'07 200mm point is flat (keep in mind if you look closely at the graph, there is only 1 data point/qtr and it is in between the ticks) - thus ~25-30K WSPM.

This would mean F36 would also be at 25-30K WPM (keep in mind also this is outs, not starts) which is beyond the fabs capacity.

If you look out into the future - in Q2'08 300mm WPM is roughly double the mature 200mm level which would mean ~50-60K WPM - which is beyond F36+38 combined.

Notice it is also still increasing beyond this point which is consistent with the fact that while 200mm F30 production will be ramped down, the 300mm production will still be ramping. Also at this point ther may be some 45nm ramping - of course wafer outs won't change, but if AMD is doing some sort of equivalency garbage than it would still increase.

So unless F30+F36 is going to be >3X (look at Q1'09) the 200mm wafer outs capability (meaning 75-90K); this data is not wafer outs but some corrected version for BOTH wafer size AND technology node.

The whole point of this is the Q2 wafer outs are roughly equivalent in between 300mm (90nm/65nm mix) and 200mm (only 90nm), so even if AMD was 100% converted to 65nm in F36 (which is unlikely as this is OUTS, not starts), best case is ~50% of their production at end of Q2 is 65nm. Again this is likely much lower as some of the 300mm outs will stull be 90nm and there is some lag for packaging.

I would expect "true" 65nm/90nm crossover for AMD to occur in ~Q3'07 (maybe early Q4). Either way 67% mid 07 is too high.

12:54 PM, December 28, 2006  
Anonymous Anonymous said...

"Anand's latest tests show that AMD STILL owns 2P in terms of power use (possibly because of FBDIMMS) and unless you saturate the CPU, 5160 just barely wins. A 3GHz Opteron (the same clock as 5160) would make a differenc"

My point was the Server Core2 is at a minimum on par/competitive with the opteron in 1P and 2P space. This is an area where AMD has expanded their market share the most (isn't server now >40% AMD?) - my point is Intel focused their Core 2 conversion in an area where AMD was having the greatest impact (and getting the most profit). This will hurt AMD short term (as it already has shown in Q3'06 market share #'s in server space).

Intel's been smart on the Core 2 conversion - convert the high end/higher margin first and let AMD dominate in low end desktop space (at price points below an E6600).

So while AMD is "fragging" 75% of Intel's products (the use of this term is liberal, especially if you consider Intel's "old" mobile architectures), but the vast majority of this "fragging" is low end/low margin desktop.

The only real good space where AMD can dominate is 4P, what will be interesting is how much this grows (if at all) - it is still a fairly small volume of the overall server market and it's unclear if 2P quad core systems will start cannibalizing this space... anyone have thoughts on this?

As a note - AMD will do the exact same thing with K8l - the conversion will focus on server and desktop (with exception of maybe 1 or 2 high end parts) will lag 6 months behind.

This is very clos eto what Intel did this year. They will lauch new architecture "mid-07" (whatever the hell that means - let's say June). At end of the year will the ENTIRE AMD line be more than 25% converted to K8l? Will Sharikou be pointing this out, or perhaps will he be focusing on a single segment (like say server) and say look at how fast they converted it?

Intel went from 0%-75% Core 2 production in server space in ~6months. This is a pretty fast ramp and is clearly focused in the right area from a business perspective.

1:13 PM, December 28, 2006  
Anonymous Anonymous said...

"65nm node is an indicator not of transistor size, but contacted Metal 1 half-pitch (see the ITRS roadmap). Associated with this node is a targeted gate CD (critical dimension) of 25nm (42nm in resist)."

Dr yield - thank you for pointing this out to the ignorant. Usually people also look at SRAM to benchmark this (logic is a bit difficult). It'll be interesting to see when someone finally cuts one of these open and analyzes this as well as the process technology - example how many of the 65nm stressors actually made it into the first rev? (there are several companies who do this and then charge for a report - I would expect this to come out soon as AMD for some reason has decided not to comment on it themselves)

And by the way my money is on #2 - unless you re-layout the design, a dumb shrink will likely introduce some speed path issues, one way around this (in a quick fashion) would be to shrink less aggressively - this is likely what AMD must have done so their PR people could say "we're only 1 year behind Intel" (just don't look at actual die area scaling and performance please!). It's questionable whether AMD would have been better off doing a new layout and getting better die size scaling (tradeoff of time to market vs capacity). AMD in this case took the quick fix (which might have been the right decision considering K-late is on the horizon)

The funny thing is if the 65nm process was indeed faster than the 90nm counterpart you would likely be able to offest the speed path issues somewhat with the faster process technology and get a better shrink. Unfortunately AMD's 65nm process is not any faster than 90nm right now and it will be at least another year until it hits maturity for speed. By then hopefully AMD will do a smarter shrink...

BTW this brings up one potential problem with CTI, as AMD improves the process, AMD will not be getting the maximum out of each process rev without re-laying out the design each time (thus they are leaving some performance on the table). As this is both time consuming and costly, I wonder how much performance is being left on the table between these CTI transitions.

One of the benefits of the Intel process transition philosophy is the near final 65nm process performance is known (and exists) at the start of the node ramp so the designs can be done accordingly.

1:36 PM, December 28, 2006  
Anonymous Anonymous said...

"I can say that the increased area certainly consumes some of the benefit of a node shrink, particulary due to the substantial increase in processed wafer cost for 65nm (200->300mm, plus I believe extra metal layer"

You're also forgetting:
- The extra process steps for the new stressors (if they are even in this rev of the chips)
- Yield delta between a 90nm process which has been running on 200mm forever vs a new 300mm and new 65nm process.
- Wafer cost between 200mm and 300mm, especially with SOI
- The increased mix of high end litho (more 20mil/tool litho layers)

1:50 PM, December 28, 2006  
Blogger TheKhalif said...


To wit gentlemen- as I noted in the last treatment of AMD's capacity, K8L simply can't address that much of the server market space until AMD has drastically ramped their 65nm capacity. If you will be so kind as to refer back to my earlier postings, I noted that due to the increased die size of K8L and the accordant yield loss due to said size, AMD consumes roughly a minimum of 3 2core 65nm die capacity to produce a single K8L die.

Implications? Well, Intel is indicating that by the time K8L ships, it will be shipping >40% of its server CPUs as 4core. Can AMD produce that substantial a volume of K8Ls? Not without cratering the rest of their 65nm production. Additionally, if the ASP of a 65nm X2 is $250 (disclaimer, number pulled out of thin air), the every K8L will need to sell for >$750 to make it worthwhile. Should be an easy choice then, no? Yes- unless it means that AMD has to tie up enough capacity that it has issues supplying other market segments with pre-existing commitments. As I've said before- Intel is pushing hard on quadcore more to put the screws to AMDs capacity than for any other reason. Any way you cut it, Intel will still be feeding at the profit trough in servers, even after K8L comes out.


None of you seem to understand how the global CPU market works. If Intel continues adding new SKUs, they are competing more with themselves than with AMD.

OEMs consider product differentiation very important. If there are 4 different SKUs at or near the same price, how do you choose which to market more?

The server market will continue to follow AMD because they are promising drop in upgrades for 1207. Some companies like IWill have even gone a step further and created hot-swap processor boards that can have either socket.

With a maintained power envelope and the current 70% projected increase, Barcelona will be formidable indeed.

My assumption is that Brisbane is more like stop gap to improve 65nm without affecting Barcelona. They have achieved at least parity in power use and this is their first 65nm batch. The Apr chips should be MUCH better. The Jun chips should be the Barcelona final revs.

I can assume also that Chartered will switch to 65nm by Feb as the added output from Fab36 will allow them to start retooling FAB30 while still keeping up with demand.

AMD seems to be used to dealing with Intel now and will probably end up getting another billion or two from the latest lawsuit.

They've already been convicted in Korea.

Anyway, I see 2007 being a banner year for AMD, while Intel will be bleeding profusely trying to maintain their operation with their "incredible shrinking margins and revenue."

I am totally confused as to why they are still continuing to lower prices. Core 2 is worth more than they charged for NetBurst, so they immediately take a hit as they ramp Core 2 up.

They are talking about charging $531 for a quad core next year.

That's psychotic.

2:00 PM, December 28, 2006  
Blogger Dr. Yield, PhD, MBA said...

Our humble host postulated:
I estimate that AMD will take 90% of the server market with K8L, and Chartered capacity is just enough for that.

Let's take an educated approach to that estimate.

Givens:
+ 188 potential good dpw for K8L (16.8mm x 16.8mm, 3mm EE, .15mm scribe, bulk of edge die zero yield)
+ 15000 wsp quarter at Chartered (5k wspm)

Assumptions:
+ 40-50% yield for a large die such as K8L (a conservative, not unreasonable estimate)
+ 96% wafer yield (scrap, test wafers, lost production time)
+ 95% sort/packaging yield (swag, probably a little high)

Result:
K8L die per quarter from Chartered at: 188*15000*(.4to .5)*.96*.95

or 1-1.3 million.

So, given your numberless "estimate", the x64 server chip market size is roughly 1.1-1.4M chips/quarter. Any evidence to support? Seems low to me, especially when you start applying bin splits to the numbers, and find that you are now in the range of only a few 100k of top bin parts/quarter.

Also, this presumes that AMD is using their full allotment of 5kwspm at Chartered for K8L, and the rest of their 65nm capacity for X2 parts.

Please present your "analysis" to support your assertions. Make sure you document your givens and assumptions.

2:09 PM, December 28, 2006  
Anonymous Anonymous said...

"I estimate that AMD will take 90% of the server market with K8L"

As Intel will BK by Q208, when will AMD have 10% of server market using K8L's? When will it be 30/50/90%? Any estimates so we could start counting days?

3:01 PM, December 28, 2006  
Anonymous Anonymous said...

Anónimo dijo...

Booting or not with 4GB of RAM means nothing. Could it be the RAM itself is the issue, maybe the BIOS. Either way, i doubt ANY desktop AMD chip on the market with 4 GB out performs Kentsfield, or E6700 for that matter with 2 GB in any meaningful bench or real world application

6:59 PM, December 27, 2006

actually, its intel's chipsets faults
funny that most CROSSFIRE actual crossfiremobos can get more than 4 GB of ram...
same with some Nvidia...
funny that at same time, some nvidia chipsets for intel do not support over 4 GB despite saying it does..

most of them because they dont support more than 1 GB for each dimm.

3:42 PM, December 28, 2006  
Anonymous Anonymous said...


Anónimo dijo...

Please tell us more about this flood - the chips are only 33% smaller (not 50%!) - just out of curiosity, do you know why that is? And given the bulk of AMD's capcity continues to be 90nm (30KWSPW on 200mm is the equiv of ~15K on 300mm), where exactly is this 65nm flood flood? Even when AMD is at 15K 65nm in F36, it still will only be 50% converted to 65nm!!!

How about another good old fashion capacity analysis? Or did you take enough of a pounding just 1 week ago on this...

9:11 PM, December 27, 2006

its the first revision, dont expect miracles, intel didnt do miracles as core2duo first versions had quite a bit of errors, specially in power consumition and random overheating..
intel so far as done 3 revissions to core2duo ( 6th in 64nm ).
AMD its on its first 65nm ...

3:44 PM, December 28, 2006  
Anonymous Casual Reader said...

Not to detract from the the subject, but this has been one of the most intellignet reads in a long time. The publics input in this subject is outstanding. Bravo Gentlemen, and ladies if there out there as anonymous.

7:35 PM, December 28, 2006  
Anonymous enumae said...

netrama said...

"...In the terms of the numbers this line can be described as AMD's Products are better than 75% (Aka Netburst) of Intel's production."

If you are trully going to look at it that way, then you could say the same thing about AMD getting fragged by Intel.

In a simple core vs core (not platform) comparison Intel's Core 2 frags 100% of AMD's K8 production...

Do you see how stupid that sounds, and then how stupid the 75% calim is as well?

The comment was originally from a excedingly biased journalist, and then from the Inq.. If you see that comment coming from a respectable source, please point it out.

"And the bleeding for Intel will continue through 2007 as 'the khalif' has said..."

Until the Q2 07 reports are released we will not know, but the numbers as far as capacity and ramping are in Intels favor.

9:10 PM, December 28, 2006  
Blogger Dr. Yield, PhD, MBA said...

an anonymous bloke 'dijo':

intel so far as done 3 revissions to core2duo ( 6th in 64nm ).
AMD its on its first 65nm ...


another confused interpretation. You are confusing steppings/mask revisions with process changes. When people here and elsewhere refer to AMD's "first" 65nm process, they are referring to major revisions in the process itself- namely gate dimensions, gate oxide thickness, metal pitch and thickness, dielectric type, etc. These all have a significant impact on device performance. The changes you refer to Intel making are design tweaks. A-step to B-step may address "errata" or what us mere mortals call bugs, and they also often address speed paths and hot spots that cause excess power consumption. "Dash steppings", such as B2 to B3, are usually minor microcode or errata fixes handled in just a few metal and contact layers. Again, not a major revision to the process, but rather the design and therefore masking layers. Don't confuse the two things- there is a world of difference

9:55 PM, December 28, 2006  
Anonymous Anonymous said...

There are no companies that make money by running benchmarks 24/7.
C2D consumes 22 watts of power at idle while AM2 draws only 7 watts at idle.
Kentsfield sucks way too much power at 60% server loading compared to opterons.
65 nm does not mean anything if your chips are not selling.
Intel screwed up there pentium 3 redesign and hence the 2 core and 4 cores are not selling.
AMD is in a sales boom and is selling everything they make.
It is sales that keeps a company running strong and AMD sure has lots of sales and profits.
I fully expect that AMD will continue to gain market share and intel to lose market share.
Intels platforms are outdated antiques.
AMD platforms are many, new, state of the art, and cheap to buy.
AMD always asks the customers what they need and want and they are always on target with the right product at the right time.

12:31 AM, December 29, 2006  
Anonymous Anonymous said...

Intels platforms are outdated antiques.
AMD platforms are many, new, state of the art, and cheap to buy.
AMD always asks the customers what they need and want and they are always on target with the right product at the right time.


Thank you for that insightful, well-supported point. Your commentary has elevated the discussion to a higher level, and for this, we will be forever in your debt. Have a nice day.

10:13 AM, December 29, 2006  
Blogger Jeach! said...


Dr. Yield, PhD, MBA said...

Completely irrelevant to the point. Intel consumed space that reduced the value of the shrink. No one (officially) knew it was there.


I wouldn't agree that it is completely irrelevant. If you hide /disable (very expensive) space on a processor, you do it for strategic purposes. Since Intel had all to gain to release this feature early, the fact that they did not only means that it wasn't properly working!

Anyway, your other post said:


Why does AMD not have a true 50% node shrink benefit? A couple of possibilities:

1. They aren't running a true 65nm process, but rather something larger.

2. The shrink of the layout introduced speed path issues that could not be worked out without rerouting in a fashion that consumed more area. Not likely for such a big delta, but maybe part of the issue.

3. Designers added more circuitry/functionality and aren't telling you. This is possible, for example Intel shipped P4s with EMT64 disabled- but it was there, taking up space.


Interesting overall, but it's all Chineese to me, please be a little more 'blog-friendly'.

But your 3 options are interesting.

What I would like to know, for those who know, is has this happened in the past when AMD shrunk to 130nm and 90nm? History tends to repeat itself.

But regardless, I believe it could be all three! And to this day, I still have Dr. Ruiz' words in my mind:

We will be releasing features that will be jaw-dropping!

He would not have said something like that on national television if they were not working on something EXCEPTIONAL!

And the fact that they seem to not be releasing anything substantial lately, maybe is proof that they are working on something BIG!

I've been detecting some very strange behaviour out of AMD lateley and it's not normal.

They are quiet, somewhat late, shipping ordinary products... simply put, they are distracted!

So I'm hoping that your 3rd point is what is happening here!

Jeach!

10:42 AM, December 29, 2006  
Blogger Dr. Yield, PhD, MBA said...

They are quiet, somewhat late, shipping ordinary products... simply put, they are distracted!

Unfortunately, I think this is exactly the problem. And I've said it before- the ATI acquisition has been a big drain on both the company's financial resources and management's attention. Given the size of the acquisition, it HAS to be successful. And that means putting focus there instead of the core business. Great if it works, but extraordinarily high, bet-the-farm risk levels.

Interesting overall, but it's all Chineese to me, please be a little more 'blog-friendly'.

Sorry, but I'm only willing to spend so much time being educational here, particularly because a lot of responses I get are of the "out-your-ass, who needs data anyway,doesn't fit with my world view, you probably don't have a PhD anyway" variety. I'd be happy to recommend some references, and I will try to keep it simple, but I have to assume some basic industry knowledge here. Not a safe assumption, but such is life.

He would not have said something like that on national television if they were not working on something EXCEPTIONAL!


Sure he would! His job, in addition to CEO, is pumper-in-chief. If the CEO can't evangelize the business, who can? Seriously- think back to Andy Grove- that guy was an evangelist. So was Sanders- to the point where he neglected running the business well. Barrett? No way. Part of the reason he wasn't a great CEO- he didn't have a lot of vision.

11:02 AM, December 29, 2006  
Blogger Sharikou, Ph. D said...

Barrett? No way. Part of the reason he wasn't a great CEO- he didn't have a lot of vision.


Not true. Barret touted 20GHZ CPUs. Paul O touted 100 core CPUs. AMD's promises a step-function increase in performance. Just wait a bit longer..

1:38 PM, December 29, 2006  
Anonymous Anonymous said...

"AMD's promises a step-function increase in performance. Just wait a bit longer.."

..and in the meantime by our "visionary" 4x4 system which will double power consumption with minimal performance benefit over our competitor (and at a higher overall system cost)

...but it will be future proof...oh, except the 1 board that you can buy today (hey, we encourage competition and "build the ecosystem") probably is not going to be able to use the more advanced features of our next gen chip - but hey it still fits!

...never mind, use the AM2, it is far superior, uh, I mean more future proof/platfomancable, to the 939 - it's not like we are using the same architecture and just changing memory compatibility...on second thought you might want to wait for am2+...or was it am3? or 4x4+? or 4x4++? (all actually seen on an AMD roadmap)

Yup step function - means AMD is going to step on consumers who are buying there products today...

Oh by the way, I'm going to quadruple the speed of the internet, not really going to tell you how, just wait and see...after all it's all about platformance!

2:53 PM, December 29, 2006  
Blogger Dr. Yield, PhD, MBA said...

So I take it you don't dispute my assertions on the server market size, and Chartered's ability to service that entirely? I haven't heard the courtesy of a response, particularly since you told you readers to do the math. I did do the math- and I'm assuming your silence is tacit agreement.

8:35 AM, January 02, 2007  
Anonymous Anonymous said...

>>> A 3GHZ single core is better than a 2GHZ dual core in most cases.

future to the back: opteron 144, 1.5V anybody?

4:16 PM, January 14, 2007  

Post a Comment

Links to this post:

Create a Link

<< Home