Will 'Barcelona' be AMD's savior?

Linc

Still blaming their latest earning report on a supply shortage to channel partners, AMD claims to be ready to meet all demand for their new Barcelona core. Current expectations mark July as the anticipated release date.

While no one expects a release on the scale of Opteron, many are looking to the event as a gauge of the viability of the company.

Read more at <a href='http://www.theregister.com/2007/03/17/amd_rivas_barcelona/'>The Register</a>.



<em>Suggested by Qeldroma</em>

mirage

Normally, AMD would not publicly admit the real reasons. Their current processors do not perform as good as Intel's latest architecture and their cost/chip is higher than Intel. Unfortunately, market prices are set by Intel due to their advantage in performance and process technology. So AMD needs to sell their under performing processors, probably, at a loss.

I am afraid Barcelona will not be sufficient to turn around the tide for AMD. They need to beat Intel's product performance by a wide margin to take over the control of market prices and charge the premium to offset their high process costs. But their competition seems extremely tough, Intel's Netburst fiasco looks like a forgotten past already and their new products and process technologies are coming very strong.

These are not good times for AMD indeed, maybe the worst of all time, ironically, just after their most glorious period during Intel's mishaps (i.e. Netburst and Itanium). AMD gained a considerable market share during last 3-4 years and initiated huge investments to expand its production capabilities accordingly, mainly at Germany and New York, US. Because of this, they are short in capital and probably in debt. It is not a good time to loose money and market share when profits are crucial.

This might look far fetched but I will not be surprised too see AMD selling part of the company (ATI?) or even be acquired completely by another company in 2008. But I hope not.

mmonnin

Where have you seen the cost per wafer out? That is a HIGHLY regarded secret in the semicondoctor industry that no manufacture likes to tell how much a wafer costs to produce.

mirage

I worked several years in a process design team in industry and still doing research in academia. Actually, the cost is not a secret, you can easily guestimate. You might want to read this thread as well.
http://www.short-media.com/forum/showthread.php?t=53885

mmonnin

There is no info in that thread about PURE, RAW cost per wafer out...How much did the fab spend this week for a weeks worth of wafers? I work at Micron, cost per wafer and yield is VERY confidential information.

There are multiple ways to lower cost per die during the course of a manufacturing process. Yields improve, new revisions, reducing cycle time from wafer start to probe.

There is a good change that Intel does have a lower wafer cost than AMD, maybe even with Intels large overhead. But I still have never seen that the average cost of a wafer from a semiconductor manufacture, besides the one I work for.

mirage

One does not need to know the PURE, RAW costs to make a relative cost comparison.

Intel is using 65nm exclusively and AMD is still using 90nm process exclusively (small portion of the production in 65 nm). 30% shrinkage in the feature size translates to (0.7x0.7=0.49) half die size and two times more chips per wafer. 1/3 cost of the processor is process (+ 1/3 testing + 1/3 packaging) on average. So, the back-of-the-envelope estimate makes Intel process 17% cheaper (33% /2). But AMD process is SOI which is inherently more expensive than Intel's non-SOI, strained-silicon process (more process steps). And yield of Intel is actually higher than AMD (evidence is that AMD still can not shift to 100% 65nm). Also Intel has completed their 300mm transition before every other company that puts even more chips on a wafer compared to 200mm wafers. Based on these, I would estimate Intel's cost/chip is at least 20% less than AMD. Here you go, an estimate without insider information
And a link from 2005: http://www.xbitlabs.com/news/cpu/display/20050913222050.html

Sledgehammer70

mirage wrote:

One does not need to know the PURE, RAW costs to make a relative cost comparison.

Intel is using 65nm exclusively and AMD is still using 90nm process exclusively (small portion of the production in 65 nm). 30% shrinkage in the feature size translates to (0.7x0.7=0.49) half die size and two times more chips per wafer. 1/3 cost of the processor is process (+ 1/3 testing + 1/3 packaging) on average. So, the back-of-the-envelope estimate makes Intel process 17% cheaper (33% /2). But AMD process is SOI which is inherently more expensive than Intel's non-SOI, strained-silicon process (more process steps). And yield of Intel is actually higher than AMD (evidence is that AMD still can not shift to 100% 65nm). Also Intel has completed their 300mm transition before every other company that puts even more chips on a wafer compared to 200mm wafers. Based on these, I would estimate Intel's cost/chip is at least 20% less than AMD. Here you go, an estimate without insider information
And a link from 2005: http://www.xbitlabs.com/news/cpu/display/20050913222050.html

Actually AMD is at 60-65% on 65nm not 30%

The performance isn't to obvious to the online community as AMD is keeping pretty tight lipped on the subject, but I have seen a Barcelona chip in action next to a Intel setup and it is indeed a faster chip...much faster in the application that was running at the time. You will see the results in due time.

a2jfreak

mirage wrote:

Intel is using 65nm exclusively and AMD is still using 90nm process exclusively (small portion of the production in 65 nm).

What? Either AMD is exclusively using a 90nm process or they're not. It can't be exclusively with a small portion of 65nm. I understand what you meant, but when you're wanting to debate someone, don't make such mutually exclusive statements because it makes taking seriously the rest of your argument more difficult.

mirage wrote:

30% shrinkage in the feature size translates to (0.7x0.7=0.49) half die size and two times more chips per wafer. 1/3 cost of the processor is process (+ 1/3 testing + 1/3 packaging) on average. So, the back-of-the-envelope estimate makes Intel process 17% cheaper (33% /2).

Can you provide some evidence to this? I don't recall having seen estimates that show process, testing and packaging are each 1/3 the cost of the product.

mirage wrote:

But AMD process is SOI which is inherently more expensive than Intel's non-SOI, strained-silicon process (more process steps).

Perhaps this is true, but I don't recall having read that SOI is more costly than strained silicon.

mirage wrote:

And yield of Intel is actually higher than AMD (evidence is that AMD still can not shift to 100% 65nm).

Higher yield? Where do you find this information? Besides, doesn't AMD's usage of SOI actually increase yield because it allows for higher clockspeeds and lower TDP, thus allowing the sale of CPUs that might otherwise not have "made the cut?"

mirage wrote:

Also Intel has completed their 300mm transition before every other company that puts even more chips on a wafer compared to 200mm wafers.

I thought that AMD transitioned to 300MM wafers in tandem with the 65nm process.

mirage wrote:

Based on these, I would estimate Intel's cost/chip is at least 20% less than AMD. Here you go, an estimate without insider information

Based on these unsubstantiated claims you might be correct. In fact, I'm not even saying you're not correct, I'm just saying please give some information to help support your arguments.

mirage

a2jfreak wrote:

Can you provide some evidence to this? I don't recall having seen estimates that show process, testing and packaging are each 1/3 the cost of the product.

Oh boy! I guess you think package is the paper-plastic package that they put the processors in the stores and testing and sorting are done with hammer and screwdriver . This is the first link for the Google search on chip packaging

http://www.amkor.com/enablingtechnologies/FlipChip/index.cfm

You do the search for testing/sorting part if you care.

a2jfreak wrote:

Perhaps this is true, but I don't recall having read that SOI is more costly than strained silicon.

So I need to explain you what is SOI, and why it is more expensive? It is a known fact since 6-7 years, learn yourself. Have a look at strained-silicon as well.

You made my day thanks, I am still laughing ....

mmonnin

AMD is doing the cardinal sin of manufacturing according to some in the business...a manufacturing process change and a new part design. What Intel does is better....new part then change that developed part to a new process. Although this does create quite a bit of extra work for those Actually in the Fab..ME.

A 30% manufacturing process shrink does by no means shrink the die by 30% in each direction, ESPECIALLY processors. Logic devices do not shrink very well. Cache on CPUs or memory chips do shrink much better than logic devices. In our last manufacturing process change in DRAM, we had a larger manufacturing shrink than 90nm->65nm but still did not reduce our chip size by 50% like you mentioned...let alone a CPU..it has no chance.

Most likely AMDs Fab36 is well on its way to producing more than 30% of AMDs chips.

The term 'Yield' can be used different ways. Most likely the way a manufacture expresses yield is not the way a2jfreak did by determining speed. A 6000+ CPU could very well come from the same wafer as 4600+ as an example. Manufactures do not manufacture them to be a certain speeds from the start, they are binned out to different speeds well afterwards at a testing facility, after they all yielded. Yield = a functional die. What that die sells for does not determine 'yield'.

Of course Intel has gone to 300mm before any other CPU manufacture, they have Billions of cash laying around to buy all new equipment. Going from 90nm to 65 is totally different from going from 200mm to 300mm. A new wafer size requires a COMPLETE new Fab! A new process change is mainly limited by the Photolithography equipment, the most expensive equipment in the fab in the range of $20-30Million/each.

Much of your 20% lower cost is negated by the overjudgement of the die size reduction. And die size does not change over generations to begin with. Manfuactures use new technology and then add more cache or more memory so the size doesnt shrink anyway. Remember how much extra cashe Intel CPUs have compared to AMD CPUs and that most of a core is just that, cache. (Heh I just read the article linked...sounds like a broken record.)

a2jfreak

mirage wrote:

You made my day thanks, I am still laughing ....

Glad you find it funny.

Still, you make statements that are not backed up by any facts that I've seen. When asked to clarify you blow it off as if I don't understand what you meant by packaging or that I don't know what SOI means. I don't need to use Google. You should link your arguments to documentation that supports your claims or else not be surprised when people ask you to provide such support.

Sledgehammer70

Agreed... Most of your claims are not backed by anything. I would provide my source of information on my above post but I am under an NDA... thanks to my pals at Nvidia

mirage

mmonnin wrote:

A 30% manufacturing process shrink does by no means shrink the die by 30% in each direction, ESPECIALLY processors. Logic devices do not shrink very well. Cache on CPUs or memory chips do shrink much better than logic devices. In our last manufacturing process change in DRAM, we had a larger manufacturing shrink than 90nm->65nm but still did not reduce our chip size by 50% like you mentioned...let alone a CPU..it has no chance.

Much of your 20% lower cost is negated by the overjudgement of the die size reduction. And die size does not change over generations to begin with. Manfuactures use new technology and then add more cache or more memory so the size doesnt shrink anyway.

Just to be exact, the difference between 90nm and 130nm is 41.6%:

Athlon 64 3500+ (90nm)
Freq/Cache: 2.2GHz / 512KB-L2
Process Technology: 90nm Silicon on Insulator (SOI)
Approximate Die Size: 84mm squared
Nominal Voltage: 1.40 V
Max Thermal Design Power: 67 W
Max Icc (processor current): 45.8 A

Athlon 64 3500+ (130nm)
Freq/Cache: 2.2GHz / 512KB-L2
Process Technology: 130nm Silicon on Insulator (SOI)
Approximate Die Size: 144mm squared
Nominal Voltage: 1.50 V
Max Thermal Design Power: 89 W
Max Icc (processor current): 57.4 A

I found the info here. And it is pretty close to 50%. You are wrong.

Sledgehammer70

mirage wrote:

Just to be exact, the difference between 90nm and 130nm is 41.6%:

I found the info here. And it is pretty close to 50%.

8.6% is by no means close.. that is a huge number in regards to die shrinking...

Its almost like saying my Intel Core 2 Duo is 9% faster... yeah well my AMD Athlon 64 is pretty close to it...

mirage

Sledgehammer70 wrote:

8.6% is by no means close.. that is a huge number in regards to die shrinking...

Its almost like saying my Intel Core 2 Duo is 9% faster... yeah well my AMD Athlon 64 is pretty close to it...

Sledge, this is about die size after 70% shrink from 130nm to 90nm.

100*(144-84)/144=41.6%

mmonnin

Hehe funny. First you mention the die shrink from 90 to 65nm or 30% like you said. Then you multiple the reduction, 0.7, by itself as if the die were square....dont think so. Now you use a 130nm to 90nm example.

Using your numbers...70% of the size(NOT a 70% shrink).....thats 90/130 = 67%...close enough to 70%.

A new revision, a new arrangement of structures....more effcient. Its more than a die shrink so its not even comparable. We also did a new revision of our DDR2 part type and gained a hefty gain in die/wafer.

Qeldroma

mmonnin wrote:

Yield = a functional die. What that die sells for does not determine 'yield'.

That is what I thought too. And since the transistor count on the Penryns (45nm) is supposed to be 410 million semi-conductors for the duals and 820 for the quads. Do the math with these:

Penryn dual size- 107mm^2

Conroe Dual size- 144mm^2 291 M Transistors

Barcelona Quad size=283mm^2, 463M Transistors

Since Penryn quad is basically takes a page from Kentsfield in design but does it monolithically, the die size should be a modest 220mm^2

Still, no matter how you slice these guys- bottom line for me is going to be max performance/price. I'll reserve my judgement as to the pass/fail of Barcelona until then. From the article:

Sources tell us that the chip is cranking through software at an unreal clip in the labs, beating out AMD’s oft-cited 40 per cent performance gain figure.

and don't forget

AMD expects its ATI buy to pay serious data center dividends when the company shifts to 45nm chips in 2008. The chipmaker is looking to combine x86 processors with graphics engines to make products capable of cranking through scientific workloads with unprecedented performance. Such products will be complemented by third-party accelerators that plug right into Opteron sockets - an attack [EDIT note- I think she meant "area"] where AMD still enjoys a healthy lead over Intel.

I think Intel may find itself catching up to AMD technically again soon.

mmonnin

Qeldroma wrote:

I think Intel may find itself catching up to AMD technically again soon.

In a different arena I think. Like AMD was the first to go 64-bit, and the first to put the memory controler on die, AMD will be the first to integrate the GPU and CPU on a package, this being with Fusion. With a chipset, or even a 3rd party chipset to cover other I/O it could make for some cheaper server machines. Since AMD has build a server chip and is going down through the markets, the next step is into HPC with Fusion.

Intel has the technology to do the same as the do manufacture CPUs, integrated video (largest manufacture in the world), and chipsets.

RWB

This is what I think: :banghead: