Bulldoze that Sandy Bridge

[RaV666]

You have to excuse me and some other people JF.
AMD had a history for some time of using the same dies and architecture for desktop and server parts.
Started with MP, then first hammers.
I had athlon 64 san diego, and later opteron venus, and they were the same.Later x2 4800+ and opteron 165.And they were the same.
From what i read on many sites and forums amd used same masks for all these products.
It would be pretty good ending of this discussion if you would tell us at which point amd stopped using same masks for desktop and server products.
It would be awesome too if you could share some info on the exact differences between them (hardware wise, not feature wise).Maybe a post on your blog ?

And if this means that AMD is going with different designs/masks with the zambezi/valencia cores, thats very interesting too.Nobody touched on that before.
I bet dresden boy could use such info too.

[BeepBeep2]

You have to excuse me and some other people JF.
AMD had a history for some time of using the same dies and architecture for desktop and server parts.
Started with MP, then first hammers.
I had athlon 64 san diego, and later opteron venus, and they were the same.Later x2 4800+ and opteron 165.And they were the same.
From what i read on many sites and forums amd used same masks for all these products.
It would be pretty good ending of this discussion if you would tell us at which point amd stopped using same masks for desktop and server products.
It would be awesome too if you could share some info on the exact differences between them (hardware wise, not feature wise).Maybe a post on your blog ?

And if this means that AMD is going with different designs/masks with the zambezi/valencia cores, thats very interesting too.Nobody touched on that before.
I bet dresden boy could use such info too.

2009 and 45nm, when AMD stopped using the same sockets and CPU's for both server and desktop. There was a lot of internal confusion from 2003 and 2008 regarding server team vs desktop team IIRC and nobody really knew what was going on.

Recently, you will notice that except for the few Suzuka based chips, no Opteron chips are an AM2+/AM3 package.

[scorpidragon]

Well, going by some rough calculations BD would need:
982 pins for Dual channel (AM3+; not 942 pins) and 1207 pins for quad channel.

If 1207 pin desktop socket was made available. I would upgrade four of my computers to socket 1207, but if AM3+ came out with only 942 pins, then I would not upgrade my computers to AM3+.

[zir_blazer]

You have to excuse me and some other people JF.
AMD had a history for some time of using the same dies and architecture for desktop and server parts.
Started with MP, then first hammers.
I had athlon 64 san diego, and later opteron venus, and they were the same.Later x2 4800+ and opteron 165.And they were the same.
From what i read on many sites and forums amd used same masks for all these products.
It would be pretty good ending of this discussion if you would tell us at which point amd stopped using same masks for desktop and server products.
It would be awesome too if you could share some info on the exact differences between them (hardware wise, not feature wise).Maybe a post on your blog ?

And if this means that AMD is going with different designs/masks with the zambezi/valencia cores, thats very interesting too.Nobody touched on that before.
I bet dresden boy could use such info too.

Socket 939 Opterons were exactly the very same Processors that a comparable Athlon 64/64 X2, but with another brand name and higher quality bin. These weren't just similar, they were identical in package and everything, so they simply aren't good enough for a decent comparision. The best examples of a Core that had a very broad usage among many different platforms is the early Rev. SH-CG Core, also know as Clawhammer (Socket 754, 939) or Sledgehammer (Socket 940), as it was used in everything from Athlons 64, Athlons 64 FX to Opterons 8xx and Mobile Athlons 64. It was a one-size-fits-all, assuming that the Revision identifies an individual die piece design.
On the K8 Rev. E generation things started to get interesing, talking about variety. You had a Rev. SH-E1 Core (Single Core, 1 MB Cache L2) that was used in early Socket 940 Opterons, that later was replaced by the Rev. SH-E4 with same specifications, that had both Socket 940 versions in Single Core Opterons 2xx and 8xx while in Socket 939 it was used for Opterons 1xx and some Athlons 64. Also, you had an Rev. SH-E5 Core, that was used in some Mobile Athlons 64. I suppose that SH-E4 and SH-E5 coexisted as they feed two different lines, while SH-E1 was simply superceded by SH-E4. You also had two Single Core 512 KB Cache L2 dies in Socket 754/939, the Venice Rev. DH-E3 and Rev. DH-E6 ones. However, the main issues are that you need to get too technical to have at least a minor idea of what differences there are bewthem these Revision of a Core with the same base feature set (Erratas, manufacturing process, Voltage/Frequency/Power Consumption curves, etc).

Last time that I checked, AMD seems to be producing three different Hexacore dies, the Rev. HY-D0 for Socket F Istanbuls, the Rev. HY-D1 for Lisbons and Magny Cours, and the Rev. PH-E0 for Thubans. Without very specific data is pretty much impossible to know puntual differences, though it would be interesing just for the record.

[scorpidragon]

Last time that I checked, AMD seems to be producing three different Hexacore dies, the Rev. HY-D0 for Socket F Istanbuls, the Rev. HY-D1 for Lisbons and Magny Cours, and the Rev. PH-E0 for Thubans. Without very specific data is pretty much impossible to know puntual differences, though it would be interesing just for the record.

AMD needs to simplify there product; like the way they used too as it would cut their production costs.
Like; dump AM3/AM3+/C32 and move all to a 1207 pin compatible socket for desktop and budget servers.
Then target G34/6* at the server market, rather than using C32 and G34 to target that market.
AMD needs to save money from production as your current setup waists money!

[zalbard]

AMD needs to simplify there product; like the way they used too as it would cut their production costs.
Like; dump AM3/AM3+/C32 and move all to a 1207 pin compatible socket for desktop and budget servers.
Then target G34/6* at the server market, rather than using C32 and G34 to target that market.
AMD needs to save money from production as your current setup waists money!

I agree. Using too many sockets is inconvenient and quite expensive. Plus us, consumers, would have more upgrade paths using the same boards. Unfortunately, both AMD and Intel do this, probably to make more money on motherboard upgrades...

[RaV666]

2009 and 45nm, when AMD stopped using the same sockets and CPU's for both server and desktop. There was a lot of internal confusion from 2003 and 2008 regarding server team vs desktop team IIRC and nobody really knew what was going on.
Recently, you will notice that except for the few Suzuka based chips, no Opteron chips are an AM2+/AM3 package.

Shanghai and barcelona seems to share same core too, RB-C2.So the more prominent differentation is on the hexacores like blazer said.
Packaging does not guarantee a different die, as zir blazer explained.

...

Last time that I checked, AMD seems to be producing three different Hexacore dies, the Rev. HY-D0 for Socket F Istanbuls, the Rev. HY-D1 for Lisbons and Magny Cours, and the Rev. PH-E0 for Thubans. Without very specific data is pretty much impossible to know puntual differences, though it would be interesing just for the record.

I agree with you blazer, AMD shares most of the cores until hex core`s came along.
But it still doesnt mean valencia and zambezi wont be same dies again.

[JF-AMD]

From what i read on many sites and forums amd used same masks for all these products.
It would be pretty good ending of this discussion if you would tell us at which point amd stopped using same masks for desktop and server products.
It would be awesome too if you could share some info on the exact differences between them (hardware wise, not feature wise).Maybe a post on your blog ?

And if this means that AMD is going with different designs/masks with the zambezi/valencia cores, thats very interesting too.Nobody touched on that before.
I bet dresden boy could use such info too.

Remember that for a processor each layer has a mask. So Lisbon and Thuban *could* share several layer masks, they don't share all layer masks. Zambezi and valencia will essentially be the same die, but a.) I cannot verify that and b.) that doesn't mean that they won't vary in the future.

AMD needs to simplify there product; like the way they used too as it would cut their production costs.
Like; dump AM3/AM3+/C32 and move all to a 1207 pin compatible socket for desktop and budget servers.
Then target G34/6* at the server market, rather than using C32 and G34 to target that market.
AMD needs to save money from production as your current setup waists money!

Our current strategy optimizes the money that YOU pay for platforms.

Moving client products to server sockets would significantly increase the cost of client motherboards and not add any additional benefit to customers.

C32 (1207 pins) does not support quad channel memory, only G34 does.

Putting a client processor in a C32 socket would push up the cost of the platform without increasing any features.

Shanghai and barcelona seems to share same core too, RB-C2.So the more prominent differentation is on the hexacores like blazer said.

Barcelona has a 2MB cache and Shanghai has a 6MB cache. Just looking at the dies you can see that they are radically different.

[zir_blazer]

Barcelona has a 2MB cache and Shanghai has a 6MB cache. Just looking at the dies you can see that they are radically different.

If he is talking about Rev. RB-C2, I suppose that he intended to mean Shanghai and Deneb. Barcelona based Opterons were Rev. DR-BA, DR-B2 and DR-B3, while Agena used the last two Revisions for Desktop Phenoms.

[jeremyshaw]

If he is talking about Rev. RB-C2, I suppose that he intended to mean Shanghai and Deneb. Barcelona based Opterons were Rev. DR-BA, DR-B2 and DR-B3, while Agena used the last two Revisions for Desktop Phenoms.

Another way of seeing it is barcelona is 65nm K10. shanghai is 45nm K10.5.

[RaV666]

Sorry JF, i of course meant shanghai/deneb.
I learn something every day here .
So one more question, when similar (almost identical) chips share some of the layers,but not all.They get different steppings ?
Which would mean, that IF cpus have the same stepping/revision/node/family thats when theyre identical for sure ?
That seems to be the case.Which means many (not all) opterons share the same dies with many (not all) desktop cpus.
And that would mean that its not that desktop cpus (with same die) dont have the extra HT links, but that theyre inactive and same goes for any other functionality differences.
Of course athlons and other segment dedicated cpus dont have to follow that rule.

[scorpidragon]

The way AMD is currently going, is that they will never make the money they need! Just look back at the past and learn from your mistakes!

[Mats]

A lot of people here think they know better than AMD, that's very interesting. Not.

I think it's possible that we'll see this kind of dual socket platform in the future, but probably not before PCIe is integrated.

[JF-AMD]

Sorry JF, i of course meant shanghai/deneb.
I learn something every day here .
So one more question, when similar (almost identical) chips share some of the layers,but not all.They get different steppings ?
Which would mean, that IF cpus have the same stepping/revision/node/family thats when theyre identical for sure ?
That seems to be the case.Which means many (not all) opterons share the same dies with many (not all) desktop cpus.
And that would mean that its not that desktop cpus (with same die) dont have the extra HT links, but that theyre inactive and same goes for any other functionality differences.
Of course athlons and other segment dedicated cpus dont have to follow that rule.

Now it gets into the area that I prefer not to discuss because the conversation could hit the edges of areas I am not allowed to talk about.

[scorpidragon]

Our current strategy optimizes the money that YOU pay for platforms.

Moving client products to server sockets would significantly increase the cost of client motherboards and not add any additional benefit to customers.

C32 (1207 pins) does not support quad channel memory, only G34 does.

Putting a client processor in a C32 socket would push up the cost of the platform without increasing any features.

Well, since G34 is not available in New Zealand, and since your company has banned overclocking on G34. It means it's gonna be socket 2011 for my server and maybe my main computer too.
Two of my other machines, since DDR4 is coming they will stay as is (AM2+), and they will not be upgraded until DDR4 hits the market.
All I want is memory bandwidth.

[RaV666]

Now it gets into the area that I prefer not to discuss because the conversation could hit the edges of areas I am not allowed to talk about.

Thanks for your sincerity, i appreciate that.

[scorpidragon]

What let AMD down is that useless NUMA memory controller of theirs and if BD has that in it, then I will not be buying AM3+ as it is nothing but a stop gap - until FM (whatever) socket arrives.

What I need is memory bandwidth and a multi-threaded CPU.

[jeremyshaw]

What let AMD down is that useless NUMA memory controller of theirs and if BD has that in it, then I will not be buying AM3+ as it is nothing but a stop gap - until FM (whatever) socket arrives.

What I need is memory bandwidth and a multi-threaded CPU.

You can disable NUMA (ccNUMA)... jsut go into the BIOS and tell it not to initiate the ccNUMA related tables, and NUMA will be disabled

[zir_blazer]

Now it gets into the area that I prefer not to discuss because the conversation could hit the edges of areas I am not allowed to talk about.

This explain why I sometimes don't get concreted answers from you about these topics

[Kobaltrock]

I don't understand AMD's logic at all, why not release AM3+ boards now, this will generate revenue for both AMD + mobo partners, and this will allow all the fence sitters waiting for BD to purchase a AM3+ board (and maybe even a CPU if they have a socket 7 / 939 rig) to tie them over until BD arrives...

If you go by what the local wholesale distributor says, they don't expect shipments until May. That is a ton of potential sales AMD could be generating, but instead they are bulldozing those fence sitters to intel's playground.

[haylui]

I don't understand AMD's logic at all, why not release AM3+ boards now, this will generate revenue for both AMD + mobo partners, and this will allow all the fence sitters waiting for BD to purchase a AM3+ board (and maybe even a CPU if they have a socket 7 / 939 rig) to tie them over until BD arrives...

If you go by what the local wholesale distributor says, they don't expect shipments until May. That is a ton of potential sales AMD could be generating, but instead they are bulldozing those fence sitters to intel's playground.

traditionally, AMD is not a weak chip maker but really not good in marketing and sales I believe.

[zalbard]

I don't understand AMD's logic at all, why not release AM3+ boards now, this will generate revenue for both AMD + mobo partners, and this will allow all the fence sitters waiting for BD to purchase a AM3+ board (and maybe even a CPU if they have a socket 7 / 939 rig) to tie them over until BD arrives...

If you go by what the local wholesale distributor says, they don't expect shipments until May. That is a ton of potential sales AMD could be generating, but instead they are bulldozing those fence sitters to intel's playground.

Maybe they are trying to get rid of old stock of boards.

[zoomee]

Maybe they are trying to get rid of old stock of boards.

Not what customers want to hear when they've just dropped £150+ on a Crosshair IV!!!! grrrrrrr

Out of principle I will be refusing to buy AM3+ no matter how good bulldozer is - waiting for AM4 or intel equivelant (if bulldozer doesnt live up to the hype).

[scorpidragon]

Not what customers want to hear when they've just dropped £150+ on a Crosshair IV!!!! grrrrrrr

Out of principle I will be refusing to buy AM3+ no matter how good bulldozer is - waiting for AM4 or intel equivelant (if bulldozer doesnt live up to the hype).

The socket will most likely be called FM4 as the AM name is being dumped by AMD. That socket better be bigger than 942 pins or it's not sale!
They (AMD) better make sure that BD has better memory bandwidth than the Phenom CPU, otherwise people will end up moving to Intel. So AMD get your act together or prepare to lose sales.
The memory bandwidth sucks as it is.

[Kobaltrock]

Maybe they are trying to get rid of old stock of boards.

I somehow doubt this, they would have stopped producing boards way before '11 swung around.

It could be a chipset issue of some kind though. End of May is a very long way away...

The socket will most likely be called FM4 as the AM name is being dumped by AMD. That socket better be bigger than 942 pins or it's not sale!
They (AMD) better make sure that BD has better memory bandwidth than the Phenom CPU, otherwise people will end up moving to Intel. So AMD get your act together or prepare to lose sales.
The memory bandwidth sucks as it is.

I am also wondering if AMD will have IOMMU on all chipsets, instead of only the fx series (990fx?). That, with increased memory bandwidth would be a killer.

Intel screwed up by not allowing VT-d on their 'k' series of CPUs, so I hope AMD enables this for all their line.