AMD: 32nm yields ramping worse than expected, Llano delayed, BD didn't sample in H1

[richierich]

He can say that until he KNOWS it will not ship for revenue in some form until after Dec 31, 2011. In fact, "on track" is even more flexible than that, really.

"Llano production shipments are still expected to occur in the first half of next year. In the second quarter this year we also taped out the first 32 nm product based on our new high performance Bulldozer CPU core. We plan to begin sampling our Bulldozer based server and desktop processors in the second half of this year and remain on track for 2011 launches. These new processors will deliver significant performance improvements to the AMD platform."

[richierich]

http://hothardware.com/News/AMD-Post...creased-Sales/

"Sunnyvale demo'd the first Llano parts at Computex this year, but is keeping mostly quiet about the APU's launch date. Our sources have quietly told us that Llano is ramping at least as well as Shanghai did. That's very good news for a company still struggling to prove it can execute its roadmap solidly over the long term and it implies Llano will arrive on time within its target thermal and performance envelopes."

[terrace215]

http://hothardware.com/News/AMD-Post...creased-Sales/

"Sunnyvale demo'd the first Llano parts at Computex this year, but is keeping mostly quiet about the APU's launch date. Our sources have quietly told us that Llano is ramping at least as well as Shanghai did. That's very good news for a company still struggling to prove it can execute its roadmap solidly over the long term and it implies Llano will arrive on time within its target thermal and performance envelopes."

Looks like their sources were F.O.S., doesn't it? It's like the author didn't listen to the CC.

Of course, AMD has now (re)defined "on time" for Llano to mean merely production shipments initiated on or before the end of June, 2011.

You caught that, right? Before it was "appearing in consumer products in H1" and now it is "STILL planned for *production shipments* in H1". A very liberal usage of the word "still".

And to think some here were saying, "hey Llano was demoed behind closed doors at Computex... it could start production shipments in Q3, and launch in Q4!!!!!!!1"

[terrace215]

tape out is a significant milestone in development. after that it's fairly easy to know how things will go. it's sort of like the last piece of the puzzle. the only thing left is for glofo to ramp 32nm.

OMG, you don't actually believe this, do you?

You think from tape-out, it's smooth, predictable sailing to a production MPU? The last piece of the puzzle? The engineering teams would die laughing.

It's not.

[Chumbucket843]

OMG, you don't actually believe this, do you?

You think from tape-out, it's smooth, predictable sailing to a production MPU?

It's not.

perhaps you can explain what could go wrong. im sure you know.

NRE costs for bulldozer are virtually done, maybe a metal spin iff necessary. it's only a manufacturing problem now.

btw dont take what i say out of context. a tape out doesnt guarantee anything. if AMD says they are on track after a tape out then that does mean something.

[Solus Corvus]

From tape out it sounds like Q2 2011 for BD is a more realistic estimate.

[god_43]

how is this not a perfect example of flame baiting?

because its terrance....lol too bad xtreme sys does not have a block feature like digg does?

[Xoulz]

OMG, you don't actually believe this, do you?

You think from tape-out, it's smooth, predictable sailing to a production MPU? The last piece of the puzzle? The engineering teams would die laughing.

It's not.

OMG... do you need a hug..?

[Serra]

I'm not sure what Diggs block feature is like, but you can ignore users. I think I'm on a few lists myself, despite my overwhelming charm


Anyway, I've gouged about half the posts in this thread out. I'm leaving the thread itself as I think a commentary on new AMD parts discussed in an official AMD communique from the top is valid... even if the discussion gets a bit rabid... but let's try to keep it in order, eh?

[terrace215]

From tape out it sounds like Q2 2011 for BD is a more realistic estimate.

Llano taped out in Q4 2009, and is due to ship in H1 2011.

BD taped out in Q2 2010, on the same process, but a more complex design, larger die, more difficult to yield, and requires more validation (at least for the server parts).

BD isn't going to ship in Q2 2011.

[savantu]

the thing is he does know. tape out is a significant milestone in development. after that it's fairly easy to know how things will go. it's sort of like the last piece of the puzzle. the only thing left is for glofo to ramp 32nm.

No its not. Validation is the most unpredictable part. Horrible bugs can ruin your day and they might show up at any time during the validation process.
People tend to forget this is AMD's entirely new core since like 10 years. Can you imagine all the caveats in the multithreaded front end ?
I think the TLB fiasco with K10 is still fresh in the mind of any AMD engineer ( apparently forgotten here ). You might need to respin some layers or worse, do a significant redesign. Add 6-9 months for every respin and $30-40m in costs.
It would be interesting to find out at what stepping is BD sampling to customers.

[superrugal]

At last, I haven't heard any new official launch date, all are speculations yet...

[Solus Corvus]

Llano taped out in Q4 2009, and is due to ship in H1 2011.

BD taped out in Q2 2010, on the same process, but a more complex design, larger die, more difficult to yield, and requires more validation (at least for the server parts).

BD isn't going to ship in Q2 2011.

I'm sure that AMD has separate engineering teams working on these obviously very different chips. Do you honestly believe that you can make predictions about one based on the other?

[savantu]

I'm sure that AMD has separate engineering teams working on these obviously very different chips. Do you honestly believe that you can make predictions about one based on the other?

But it's not the cpus, it is the process. Llano is being delayed because the 32nm process isn't ready. A natural logical inference would be to assume that BD is also impacted in some way by this. Technically, they only committed to 2011 launch. Because of 32nm issues, it could very well shift from June 2011 to October 2011. For the masses, no delay occurred, it is still 2011.
But for the company internal road maps it is a slip which impact revenue and the competitive position.

[Solus Corvus]

Right, the only thing they have in common is the process and apparently that will be ready 1H.

[informal]

perhaps you can explain what could go wrong. im sure you know.

NRE costs for bulldozer are virtually done, maybe a metal spin iff necessary. it's only a manufacturing problem now.

btw dont take what i say out of context. a tape out doesnt guarantee anything. if AMD says they are on track after a tape out then that does mean something.

I think you are wasting your time.He now believes it takes 1.5-2 years from tape out to bring out a design like this to market
He also seems to forget how modular this thing is.They need a single module part to work like it should. Desktop is 4 of those on a die,server is a MCM just like MC.If underlying design has no showstopper bugs,Q2 2011 is a valid target for desktop and Q3 for servers.

[FlanK3r]

some guys like flame AMD vs Intel think...:-( And again the same people...

[Oliverda]

an useless thread... again...

[FlanK3r]

maybe ban for flame guy Terrace? We have on one czech forum cards. First u can get Yellow card, than second admonition and Red card. With red card u cant writing to forum think one month (than u can normal user and should be carefuly, because again u can get Yellow card etc etc). And the biggest sh*ts, spammers etc get black card and blocked IP :-)

[savantu]

I think you are wasting your time.He now believes it takes 1.5-2 years from tape out to bring out a design like this to market

?? 12 months is normal. A new uarch with a troubled process can easily get you to 15 months. It's doing so well that altough it taped out some time ago it will be sampled to customer sometine this year. Everyone is free to drawn their own conclusions.

He also seems to forget how modular this thing is.They need a single module part to work like it should. Desktop is 4 of those on a die,server is a MCM just like MC.

They should hire you to show them how the work is done. Probably the uncore part doesn't matter, "it's enough for a single module to work like it should". Care to tell me where K10 floped ? Was it the core or the uncore ?

If underlying design has no showstopper bugs,Q2 2011 is a valid target for desktop and Q3 for servers.

That puts it on top of Llano. Would be quite a feat to get it so soon out.

[pokipoki]

If it's true, it can't be helped. I am still looking forward to BD though.

Does anyone know the cache sizes for BD product line-up? I hope it's more than 6MB for L3, 8-12MB would be nice. Wouldn't mind paying a bit more for them.

[Zed_X]

yes its true, 32 mn process is problematic for AMD now, every products are delayed for now, we will see next month maybe will be situation different. But Bulldozer was expected in Q2/2011 and still is there for launching.

PS. Look, how long Intel prepare his 32nm, first ES samples was out for ONE YEAR before Retail product on shelves.

[pokipoki]

Some questions about Zambezi, 2 models 4-core & 8-cores, will they appear on different timelines? Will they use different sockets and cache sizes ie 8-core launching later on a different socket/mb? The current socket is quite old, it doesn't make sense to stick to it, maybe just for BDv1, but later BD & beyond should migrate to a new socket. Server counterparts are already using G34 sockets.

[Hans de Vries]

As a side note, Intel's adding 32nm capacity to ramp Sandy Bridge faster and "huge customer interest" makes perfect sense now

Q4, 2005: Intel introduces affordable 65nm quad core processors.
Q4, 2007: Intel introduces affordable 45nm quad core processors.
Q4, 2009: oops...
Q1, 2011: Intel introduces affordable 32nm quad core processors.

Good to hear all this "much better than expected" "faster ramp", "flawless
execution" talk.....

Regards, Hans

[informal]

?? 12 months is normal. A new uarch with a troubled process can easily get you to 15 months. It's doing so well that altough it taped out some time ago it will be sampled to customer sometine this year. Everyone is free to drawn their own conclusions.



They should hire you to show them how the work is done. Probably the uncore part doesn't matter, "it's enough for a single module to work like it should". Care to tell me where K10 floped ? Was it the core or the uncore ?



That puts it on top of Llano. Would be quite a feat to get it so soon out.

Barcelona erratum was L3 related,but now AMD has plenty of experience with the new shared cache concept and one would expect them to learn from their mistakes.I again state that if the module part has no showstopper bugs then this thing could easily launch in Q2 next year,on desktop. Server stuff one quarter later.
And the process node troubles can be dealt in time,it's just a learning curve.Dirk also said that in just the last couple of weeks they made great progress with the yields on the 32nm node,so I'd expect by the end of the year they will hit close or relatively close to the mark they would be comfortable with when it comes to yields.

Some questions about Zambezi, 2 models 4-core & 8-cores, will they appear on different timelines? Will they use different sockets and cache sizes ie 8-core launching later on a different socket/mb? The current socket is quite old, it doesn't make sense to stick to it, maybe just for BDv1, but later BD & beyond should migrate to a new socket. Server counterparts are already using G34 sockets.

Nobody knows ATM will the 2 variants launch at the same time,my guess is yes.Both will work in the AM3 socket(present and future AM3r2 boards),provided BIOS support is there.