Bulldoze that Sandy Bridge

[cesariuth]

8 server cores in less die space than current lisbon.

But, the 8 cores server are the same 8 cores of bulldozer???

[Apokalipse]

Server: Interlagos (2 x 8 core MCM) and Valencia (8 core) support quad channel DDR3
Desktop: Zambezi supports dual channel DDR3

[savantu]

Server: Interlagos (2 x 8 core MCM) and Valencia (8 core) support quad channel DDR3

?
How ? Since the dies are the same, you get Quad-Channel in Interlagos by using 2 dual channel dies. Valencia is dual channel.

Desktop: Zambezi supports dual channel DDR3

Which btw is the same die as Valencia.

[FlanK3r]

no, is to same, if people say this "Istanbul die is same as Thuban die". And of course, its not right!

[jeremyshaw]

no, is to same, if people say this "Istanbul die is same as Thuban die". And of course, its not right!

Obviously not . Thuban has Turbo Core and DDR3. Maybe Lisbon is closer, though it still lacks Turbo Core.

[FlanK3r]

its a bit diferent die.

[informal]

?
How ? Since the dies are the same, you get Quad-Channel in Interlagos by using 2 dual channel dies. Valencia is dual channel.


Which btw is the same die as Valencia.

There will be 8 core model on G34 with 4 channel support. It will be a "special" part(MCM),with higher clock.It's not Valencia,but still it's effectively 8 core and on G34.

[JF-AMD]

But, the 8 cores server are the same 8 cores of bulldozer???

Bulldozer is the name of the core. There is an 8-core client zambezi and and 8-core server valencia. They are generally the same, but I cannot commit that the die sizes are exactly the same, so I only make a server comparison of valencia relative to lisbon. They may be the same, but I just don;t know their die size.

Server: Interlagos (2 x 8 core MCM) and Valencia (8 core) support quad channel DDR3
Desktop: Zambezi supports dual channel DDR3

Valencia supports dual DDR-3

[savantu]

no, is to same, if people say this "Istanbul die is same as Thuban die". And of course, its not right!

What's the difference then ? Size is the same, transistor count is the same, maybe microcode is a bit different, package is different,testing might be more thourough, other than that they are virtually identical.

[JF-AMD]

Lisbon features HT Assist and has 3 HT links for connecting to other processors and I/O devices.

Thuban has Turbo CORE and some other features that Lisbon does not have.

The point that we have been making is that Valencia is smaller than Lisbon. We have not made any statements regarding thuban. The statement MAY be the same for Thuban, but nobody has officially made that statement.

[BeepBeep2]

its a bit diferent die.

Thuban/Lisbon


Istanbul

[savantu]

Lisbon features HT Assist and has 3 HT links for connecting to other processors and I/O devices.

Thuban has Turbo CORE and some other features that Lisbon does not have.

These are features which are ondie, but are not active, as a differentiators, just as Intel enables or disables HT,VT,etc. Such features are controlled by fuses in the die.

[JF-AMD]

Not always, they are both different steppings. You do not understand what you are talking about.

As for the photos BeepBeep 2, keep in mind that you are only seeing the top layer and you are only seeing to a certain level of granularity. Lookingthe same is not the same as being the same. The agrument thrown out here is that they are the same die, and they are not. They may be very similar, but they are not the same.

[savantu]

Not always, they are both different steppings. You do not understand what you are talking about.

...

Steppings aside, what is the major physical difference which would make those 2 dies different ?
Feel free to point at least one.

[BeepBeep2]

Steppings aside, what is the major physical difference which would make those 2 dies different ?
Feel free to point at least one.

Please remember that dies are not just made of one layer. There are several layers in these pics that have either been purposely eroded or are under the visible one...

Also, if you want a hint, look at the die outline. They aren't the same.

[savantu]

Please remember that dies are not just made of one layer. There are several layers in these pics that have either been purposely eroded or are under the visible one...

There are 11 layers. I am fully aware of that.
I'm also aware of the fact the AMD uses the same design for different markets. They might be tuned and have different features active/disabled, but it is the same design, same layout,etc.
It is simply a matter of resources and payback. Having different designs for different markets isn't always worthwhile.

Intel does the same with desktops and 2S Xeons. There is no difference between a I7 980 and a Xeon 5680.
But Xeon MP has its own design with a different die. That's why it took 1 year longer to get it to market. You need to wait for the uarch to be done for the generic parts, than you take it in implement it for the specific part. Not only that, but you validate much more thoroughly and that takes extra time again.

Also, if you want a hint, look at the die outline. They aren't the same.

I have quite good eyesight and I say you look again. They are identical, bar for the the differences in colours and blurring.

[JF-AMD]

Steppings aside, what is the major physical difference which would make those 2 dies different ?
Feel free to point at least one.

I pointed out several above and you even quoted that response. Please go back and read it.

[BeepBeep2]

I have quite good eyesight and I say you look again. They are identical, bar for the the differences in colours and blurring.

I'll tell you for the third time here to look again.

Look at the edge of each picture. What is on the outer edge of each die is NOT the same, not at all.

[jeremyshaw]

There are 11 layers. I am fully aware of that.
I'm also aware of the fact the AMD uses the same design for different markets. They might be tuned and have different features active/disabled, but it is the same design, same layout,etc.
It is simply a matter of resources and payback. Having different designs for different markets isn't always worthwhile.

Intel does the same with desktops and 2S Xeons. There is no difference between a I7 980 and a Xeon 5680.
But Xeon MP has its own design with a different die. That's why it took 1 year longer to get it to market. You need to wait for the uarch to be done for the generic parts, than you take it in implement it for the specific part. Not only that, but you validate much more thoroughly and that takes extra time again.



I have quite good eyesight and I say you look again. They are identical, bar for the the differences in colours and blurring.

They have different pinouts, even. I heavily doubt 1207 & C32 are somehow, magically pin compatible with AM2+/AM3. Fact is, Istanbul made it to the market long before Thuban. So by your own logic, Thuban =/= Istanbul. So far, only the E0 stepping chips have turbocore (AMD has had per core clock modulation based on workload since Agena, mind you ), an "anti-Agena CnQ" setup, in a twisted sense. Yeah, try to argue they had the functionality on every AMD hexcore design. Fact is, they don't.


And *just* incase if you are ignorant to the fact (I'm waiting for a staunch denial...), JF-AMD is John Fruehe, AMD's Server (and a few other things) PM. And if you think Product Marketing is drawing advertisements.... then... "wow" is all I will care to say in response.

[RaV666]

AMD always made same dies for desktop/server cpus, and differentiated both by disabling/enabling parts of it,packaging and testing, but most importantly masks on which they were made were the same.
AMD did it because it was most sensible approach for them ,it saves a lot of money and engineering which they dont and didnt have.
It was the case with athlon MP`s, opterons based on athlon 64 and barcelona/shanghai dies, and as far as i know with phenom/thuban dies.
Thats why server parts never had bigger caches for example.Same production line,same mask,same die.Ive never seen a proof of it being different.
I know JF you work for the company, and for some (retarded) clients it may seem as a downside for the product.But we would have to see some proof to believe different.
Server parts dont HAVE more ht links, they dont have them disabled as desktop ones have ;-)
Its like intel is squeezing more money from people over HT, intel wants quite more money to NOT castrate 2600K over 2500K.This extortion type of business is pretty widespread these days.And AMD doesnt do it that much.

@Jeremy,
Do you believe TESLA cards from nvidia use different chip than desktop GFX cards ?
Think about it for a sec.

[savantu]

I pointed out several above and you even quoted that response. Please go back and read it.

Yes, I've read it first time. As I've said then, the features you mention are found in all dies, in some they are active in others they are not.
That does not make it different designs.

I'll tell you for the third time here to look again.

Look at the edge of each picture. What is on the outer edge of each die is NOT the same, not at all.

Maybe I can't see the picture because of the trees, but frankly, except the blurring and different color tones, to me they seem identical.
Feel free to point with an arrow where you see the differences please.

They have different pinouts, even.

What has pinout to do with the design ? Do you think a K8 on socket 939 is different from a K8 on socket 940 ?
Or Prescott on socket 478 was different from a Prescott on socket 775 ?

I heavily doubt 1207 & C32 are somehow, magically pin compatible with AM2+/AM3.

What I'm arguing has nothing to do with compatibility. I thought it was obvious.

Fact is, Istanbul made it to the market long before Thuban. So by your own logic, Thuban =/= Istanbul.

Which reinforces the point. If server made it out before desktops, it means it has no extra features which needed more time to be implemented and validated. It was just a matter of marketing; supplying the highest margin market segment until production ramped up and desktops could be supplied.

So far, only the E0 stepping chips have turbocore (AMD has had per core clock modulation based on workload since Agena, mind you ), an "anti-Agena CnQ" setup, in a twisted sense. Yeah, try to argue they had the functionality on every AMD hexcore design. Fact is, they don't.

You don't add turbocore with a stepping. Either that functionality is there from the uarch phase, when all the control logic, power delivery and measure points were thought out, or you're in for a full respin ( the designed is changes, layouts, number of transistors, etc ).
So the point is that all hexacore had turbocore in, but wasn't enabled in all products. Either because of time-to-market restrictions or complexity, it could be validation of the turbocore feature wasn't ready when the design was thought fully functional w/o turbocore. Thus they decided to ship it with the feature disabled and when the validation is done, they enable it.

And *just* incase if you are ignorant to the fact (I'm waiting for a staunch denial...), JF-AMD is John Fruehe, AMD's Server (and a few other things) PM. And if you think Product Marketing is drawing advertisements.... then... "wow" is all I will care to say in response.

I'm fully aware of who he is. That doesn't mean he knows everything about server cpus or be completely open about all issues. Do you suppose he knows the instruction latencies in current CPUs or other really technical stuff ? I bet he doesn't. And frankly, he doesn't need to. He's not selling circuit design techniques, but functionality. He's selling performance per watt or per $, not L1 nanosecond access time.

I've met people on different forums which can eat JF for breakfast when it comes to the insides of AMD CPUs. They are SW designers, HW engineers, you name it. They know those quirks either because they need to in their daily activities or as a hobby.

Anyway, I prefer to leave the topic where it is and agree to disagree.

[JF-AMD]

AMD always made same dies for desktop/server cpus, and differentiated both by disabling/enabling parts of it,packaging and testing, but most importantly masks on which they were made were the same.

The word "always" is very dangerous word to use because it paints you into a corner.

I happen to know the differences between the parts. If I run out of Lisbon parts I can't put a Thuban die in a C32 package and sell it as a Lisbon. And if I run out of Istanbul I REALLY cannot put a Thuban die in a socket F1207 package.

Physically the dies may fit, but there are logic level changes between each of those dies. Each of the three is a different stepping, Istanbul is D0, Lisbon is D1 and Thuban is E0.

[jeremyshaw]

What has pinout to do with the design ? Do you think a K8 on socket 939 is different from a K8 on socket 940 ?
Or Prescott on socket 478 was different from a Prescott on socket 775 ?

Thought: how is 1207(F) different from.... 1207(C32)? Must be the same CPU/Socket? Interestingly enough, they aren't. For that matter, there are different revisions of 1207F... (or TYAN is really as bad as they seem) with and without split power planes.

Which reinforces the point. If server made it out before desktops, it means it has no extra features which needed more time to be implemented and validated. It was just a matter of marketing; supplying the highest margin market segment until production ramped up and desktops could be supplied.

So one can grab any two Thuban dies (hey, they have later stepping - *must* be later and more advanced with even MORE "unlocked" functionality!!) and slap them onto a MCM package?

You don't add turbocore with a stepping. Either that functionality is there from the uarch phase, when all the control logic, power delivery and measure points were thought out, or you're in for a full respin ( the designed is changes, layouts, number of transistors, etc ).
So the point is that all hexacore had turbocore in, but wasn't enabled in all products. Either because of time-to-market restrictions or complexity, it could be validation of the turbocore feature wasn't ready when the design was thought fully functional w/o turbocore. Thus they decided to ship it with the feature disabled and when the validation is done, they enable it.

So basically, what AMD is doing doesn't fit your predefined idea of a stepping. That's fine. AMD is run on your opinion, anyways. It's heavily unlikely AMD was already exploring the idea with Agena's CnQ, nor is it likely that stepping D1 is different from E0. It's just sheer impossible, according to you. Just because you think some nonexistant functionality is there, doesn't mean it's even remotely similar to another idea in the next stepping.

Anyway, I prefer to leave the topic where it is and agree to disagree.

Not surprising. People with nothing to offer, love pretending to have the high horse to ride upon (and twisting words, but who knew?).

EDIT: I don't like dragging other people into my own junkpile, but post above me also adds in a lot of value.

[RaV666]

The word "always" is very dangerous word to use because it paints you into a corner.

I happen to know the differences between the parts. If I run out of Lisbon parts I can't put a Thuban die in a C32 package and sell it as a Lisbon. And if I run out of Istanbul I REALLY cannot put a Thuban die in a socket F1207 package.

Physically the dies may fit, but there are logic level changes between each of those dies. Each of the three is a different stepping, Istanbul is D0, Lisbon is D1 and Thuban is E0.

In this instance i would say its pretty possible, they are different masks and istanbul never made it to desktop.Different revisions.
And just a question, could you exchange barcelona and shanghai "parts" ?
And if not, isnt it because after initial production, they got fused/programmed with corresponding features and got their different code names ?.Same revision.
My logic is, yea, they cant be changed one for another, but its because they already got fused differently.Which doesnt change the fact that theyre the same physically.

Its like, phenom 920/940.They were the same as 910, BUT, 920/940 didnt work with am3.910 did.They were same chip with same revision, but different fusing/microcode.

And yes, i shouldnt use the word "always" :-) .But "most of the time" ;-P

[JF-AMD]

In this instance i would say its pretty possible, they are different masks and istanbul never made it to desktop.Different revisions.
And just a question, could you exchange barcelona and shanghai "parts" ?
And if not, isnt it because after initial production, they got fused/programmed with corresponding features and got their different code names ?.Same revision.
My logic is, yea, they cant be changed one for another, but its because they already got fused differently.Which doesnt change the fact that theyre the same physically.

Its like, phenom 920/940.They were the same as 910, BUT, 920/940 didnt work with am3.910 did.They were same chip with same revision, but different fusing/microcode.

And yes, i shouldnt use the word "always" :-) .But "most of the time" ;-P

This started with the discussion that these were the SAME die, not interchangeable.

They are not the same die. I know this to be a fact.

I am not sure where all of these people get the ideas that every chip has the exact same features and it is just a fusing option. There are several things that are fuseable, but some of the things that I have mentioned are not fuseable. Many of the layers may be the same, but not all of the layers are the same.

Can we all please drop this discussion?