AMD cuts to the core with 'Bulldozer' Opterons

[JF-AMD]

JF-AMD i am a bit confused as to when documentation on sockets are sent to mobo manufactures, i know someone in a reputed mobo maker and he told me that they have yet to receive any documentation for the G34 socket but have all-ready received some info on LGA 1567.

Is it because AMD will have limited partners making G34 socket based mobo's or is it that the documentation is not released?

Not sure what they could be talking about. We go to Taiwan quarterly and have a local office there that works with all of the ODMs and board manufacturers.

I displayed G34 boards last september, Supermicro was showing them off at supercomputing 09 (along with other vendors). Taiwan has had G34 details for close to 24 months by my estimation. I can't say the names of the partners that I personally met with, but I have been there talking about G34 since early 2008.

The folks we talk to are all the usual suspects. All of our OEM partners have been testing G34 for some time now.

[JF-AMD]

Not to nit-pick, b/c I agree with what you are saying, but the "who" that GF could sign would be ARM customers. ARM chips are like grains of sand compared to desktop CPUs. They are everywhere, in everything.

Sign a big enough ARM design (say Qualcomm and cell phone products), and they could potentially load up the fab with no need for AMD.


Just speaking hypothetically, don't flame me for it.

No, you're totally correct, ARM could be a big customer for GF. One of the challenges in having someone like ARM is that on a "per wafer" basis, the revenue *could* be lower. But I don't know any GF details, I am not that close to manufactring operations to know.

[Shadov]

No, you're totally correct, ARM could be a big customer for GF. One of the challenges in having someone like ARM is that on a "per wafer" basis, the revenue *could* be lower. But I don't know any GF details, I am not that close to manufactring operations to know.

Thanks for all the answers JF, great to have you here!

[Opteron146]

No, you're totally correct, ARM could be a big customer for GF. One of the challenges in having someone like ARM is that on a "per wafer" basis, the revenue *could* be lower. But I don't know any GF details, I am not that close to manufactring operations to know.

You indirectly admitted with that statement that there are risks in the Fab-Light strategy.

As the analyst wrote on eetimes that I quoted in my earlier post, not everybody could be a FAB's No.1 customer.

It is not too far fetched to assume a future situation, where GF has to choose to invest in either a SHP (super high performance) SOI process or a HP bulk process first.

If they choose the HP process and AMD is still using SOI, then AMDs process gap to intel will increase even more ...

I hope that this will never be the case, because GF is now starting to be a big FAB company with many employees; thus they should have sufficient personnel to work in parallel on both approaches ... but who knows.

I hope AMD will sell lots of CPUs to remain GF's No.1 customer ;-)

cheers

O.

[LesGrossman]

It is not too far fetched to assume a future situation, where GF has to choose to invest in either a SHP (super high performance) SOI process or a HP bulk process first.

If they choose the HP process and AMD is still using SOI, then AMDs process gap to intel will increase even more ...

And GF would make less money because of it.

I hope AMD will sell lots of CPUs to remain GF's No.1 customer ;-)

GF hopes that too. That's why they will always try to make AMD as competitive as possible, with the bonus of being able to use whatever process they invested for AMD, with any other client.

[JF-AMD]

You indirectly admitted with that statement that there are risks in the Fab-Light strategy.

Every strategy has risks, there is no risk-free strategy. Building a fab is probably more of a risk than using a foundry.

[nn_step]

If I go from an Opteron 2435 D0 to an Opteron 2435 D1, my partners have to do a complete re-qual on the processor, that is how they work.

There are notifiable and non-notifiable changes. Doing a rework with HKMG would be a notifiable change to partners and that would force them into a requal. Requals take several months (complete reression testing on all platforms).

There are minor changes that we do that are non-notifiable (typically only touching one layer), but an all-layer change would force more work for partners. Most would not want to take the part based on the amount of work vs. the short life cycle. If I give them Magny Cours in Q1, then HKMG in say Q3 or Q4, with Bulldozer coming in 2011, they would either wait for HKMG (a bad option for us) or they would skip HKMG.

It's an economic decision, not a technical decision. You may see the change as low risk, but OEMs will not. There is too much money on the line.

thank you for clearing that up for me.

[w0mbat]

JF: There are already (exists) first samples for server Bulldozer? Thx.

yes, there are bd samples up & running.

[hyc]

Yeah, I know that you're a server guy.

According to the "tradition" (since K8) AMD always uses its server dies on (high end) desktop.

Italy = Toledo
Santa Rosa = Windsor
Barcelona = Agena
Shanghai = Deneb

The difference is the package and the desktop versions don't support the registered DIMMs.

The memory controller is identical, therefore both versions support both registered and unbuffered DIMMs. (I've run both registered ECC and unbuffered ECC DIMMs with a Venice core Athlon64, X2 3000+, and Opteron 185. No problems either way.) It's a matter of motherboard support, really.

[JF-AMD]

yes, there are bd samples up & running.

Where are you getting this data?

[Particle]

Last I heard BD was supposed to start sampling this quarter or the next, but I haven't seen any more specific info.

[Oliverda]

The memory controller is identical, therefore both versions support both registered and unbuffered DIMMs. (I've run both registered ECC and unbuffered ECC DIMMs with a Venice core Athlon64, X2 3000+, and Opteron 185. No problems either way.) It's a matter of motherboard support, really.

You're right. It's only a BIOS dependent feature.

[JumpingJack]

I don't think AMD will be paying the equivilant of the $4-5 Billion on wafers, that it costs to build and equip a fab and develop a new process.

http://www.semiaccurate.com/forums/s...t=1226&page=27

Foundries, though, pay these costs and, as business, they also want to be profitable. Therefore, the fab build (Capital expenditures) are paid, depreciated, and recouped by billing those costs into the costs of the wafers they sell to their customers.

The GloFo maneuver only works if the resulting foundry can keep utilization at high as possible spread over many customers (the more the better). If AMD were GloFo's only customer (this is hypothetical, because this is certainly not the case) and GloFo was on a mandate for profit, ultimately costs to AMD would be higher not lower due to the spin off. This is why many analysts were skeptical of the spin off approach, if GloFo did not retain any new customers, ultimately AMD would be paying that 4-5 billion for a fab and development costs as it would have been built into the price GloFo would charge them per wafer.

The concept is that the capital expenditure costs necessary to run a fab is spread over several customers, such that the total wafer costs are lower per customer than if they tried going it alone. This requires as many customers as you can secure and selling all your capacity to run at high efficiency.

[god_43]

this website has a lot of information, they also have some nice slides....pdf's only seem to work if you access them from the non translated page?


http://translate.google.ca/translate...%26tbs%3Dqdr:d

[Mats]

The pics work in the translation if you click on them, and not the PDF link in the text.

[w0mbat]

Where are you getting this data?

incredible india

hope u cant check my ip

[Oliverda]

incredible india

hope u cant check my ip

[FlanK3r]

wombat: whau! and :-)

[LightSpeed]

incredible india

hope u cant check my ip

LOL

[JF-AMD]

incredible india

hope u cant check my ip

Don't need to check your IP, but if I were you I would check your sources.

[Shadov]

Don't need to check your IP, but if I were you I would check your sources.

JF may I ask you what are your biggest hopes & goals when it comes to the new CPU architecture?

[wuttz]

JF may I ask you what are your biggest hopes & goals when it comes to the new CPU architecture?

prolly that he can sell his wares without his competitor coercing his customers to limit volume/platforms to a certain number.

[JF-AMD]

JF may I ask you what are your biggest hopes & goals when it comes to the new CPU architecture?

Scalability
Power efficiency
Value

That is what I care about. Not really concerned about how it handles cinebench or povray or the other "benchmarks".

When I talk to customers they just want stuff that works. And they want value. And they want things to be easy. It's actually a pretty simple world out there from that aspect.

My favorite example is the fact that I changed from tivo to windows media center. WMC is 10X more powerful, does such a better job of handling movies, music and other things. Much easier to use netflix.

But tivo was rock solid.

I just like things that work.

[Chumbucket843]

would modularity fit under scalability? i.e. changes in cache size, adding an APU, or other uarch tweaks.

i totally agree with you opinion on scalability. an architecture is useless with out it. if designs are not reused or recycled in some way then you will never make a profit.

prolly that he can sell his wares without his competitor coercing his customers to limit volume/platforms to a certain number.

before posting, ask yourself if your post contributes to the thread.

[JF-AMD]

Yes, modularity is important, that was the basis of the design.