AMD cuts to the core with 'Bulldozer' Opterons

[Particle]

#1 is already known. AMD (in a blog, if I recall) has stated that the interconnect in their MCM product uses an even faster version of the normal HT link used in MP systems. Yes, Istanbul has a snoop filter. What isn't known is how it will apply. Given the extra apparent bandwidth, MC may conceivably not use any filter between the two modules. It would be completely unnecessary for two CPUs in one package just as it is for current 2P systems. (CPU0 goes, "Hey, I don't have that memory address. Since there are only two of us, it must be at CPU1.") I would imagine how that works inside the package will be dictated by how they've rigged up the newly-combined (quad-channel DDR3) memory controller. If each chip in the MCM is set up as one NUMA node, I suspect we'll see the potential for the snoop filter to exist between both dies on the same package when there are 2+ CPUs installed. It makes the most sense though, I think, that they are treated as one. Otherwise it isn't so much "quad-channel" as they've been marketing it as it would be "dual, dual-channel". There's a real difference there.

[Lightman]

#1 is already known. AMD (in a blog, if I recall) has stated that the interconnect in their MCM product uses an even faster version of the normal HT link used in MP systems. Yes, Istanbul has a snoop filter. What isn't known is how it will apply. Given the extra apparent bandwidth, MC may conceivably not use any filter between the two modules. It would be completely unnecessary for two CPUs in one package just as it is for current 2P systems. (CPU0 goes, "Hey, I don't have that memory address. Since there are only two of us, it must be at CPU1.") I would imagine how that works inside the package will be dictated by how they've rigged up the newly-combined (quad-channel DDR3) memory controller. If each chip in the MCM is set up as one NUMA node, I suspect we'll see the potential for the snoop filter to exist between both dies on the same package when there are 2+ CPUs installed. It makes the most sense though, I think, that they are treated as one. Otherwise it isn't so much "quad-channel" as they've been marketing it as it would be "dual, dual-channel". There's a real difference there.

And that's the question I would like answer to

AMD might have something more than Istanbul like Snoop filter.
They also might decide to upgrade HT links on MCM to full 16-bit, double pump freq. compared to external links, etc... There are a lot of things which can be done if they give enough benefits in targeted usage patterns.

[Nedjo]

[Xoulz]

Is it me, or is that drawing confusing? Or, incomplete.

[doompc]

Xoulz, each Sao Paulo core inside the Magny Cours have 4 HT links, 3 of them are Cache Coherent. (the 4th one is to connect the cpu to the chipset)
Both dies are connected by one and a half HT link (x16+x8) and there is another one and a half HT link for each die to connect to one of the dies on the other Magny Cours (in a 2P system).

[Dresdenboy]

According to a presentation, the internal links work at 6.4 GT/s. Their way of connection looks like this:

[ajaidev]

According to a presentation, the internal links work at 6.4 GT/s. Their way of connection looks like this:

[ig]http://www.planet3dnow.de/vbulletin/attachment.php?attachmentid=17117&stc=1&d=12592393 85[/img]

Thanks for that, it does seem that Lisbon "This is MCM Lisbon rite" is quite different as compared to Istanbul...



Also since Thuban is suppose to be based on Lisbon this is quite exciting, where is the L3 by the way on the slide?

[Oliverda]

Thanks for that, it does seem that Lisbon "This is MCM Lisbon rite" is quite different as compared to Istanbul...



Also since Thuban is suppose to be based on Lisbon this is quite exciting, where is the L3 by the way on the slide?

I can't see any major difference.

[informal]

I can't see any major difference.

Me neither.

[ajaidev]

I can't see any major difference.

Me neither.

Not major difference but things do seem to have move around, i even think that this maybe small than 346mm2 "Istanbul die". Time will tell if that is true or not tough.

[demonkevy666]

thought this thread about bulldozer not something we've already seen >_> ?
I don't get this bulldozer anymore.

[blindbox]

I can't see any major difference.

That's an Istanbul die, check the image information. No difference between two istanbuls, yeah?

I don't get what ajaidev is trying to say though.

[Helmore]

thought this thread about bulldozer not something we've already seen >_> ?
I don't get this bulldozer anymore.

That's better than the other way around. I mean, you wouldn't want a bulldozer to get you...

[JF-AMD]

Xoulz, each Sao Paulo core inside the Magny Cours have 4 HT links, 3 of them are Cache Coherent. (the 4th one is to connect the cpu to the chipset)
Both dies are connected by one and a half HT link (x16+x8) and there is another one and a half HT link for each die to connect to one of the dies on the other Magny Cours (in a 2P system).

The die is not a Sao Paulo.

There are 4 coherent links per processor, this allows you to create a fully connected 4P with single hop latencies by utilizing a split link.

[JF-AMD]

Lisbon will be very similar to Istanbul, but it is a new stepping with some new features not found in Istanbul. Thuban is a desktop die that has features you won't find in server.

Snoop filter is enabled in all server chips (assuming 2P or greater combinations.)

[JF-AMD]

Apart from obvious ones?

- will Bulldozer incorporate improved interconnects for MCM packaging.
- what will be maximum memory speed supported by IMC.
- how many DIMMs IMC will be able to drive (single, not MCM)
- which extensions will be supported (SSE(S3,4,4.1,4.2,AVX,3DNow,etc) and to what degree

Just a few

1. All communications between the die happen at package speed.
2. That is not being disclosed yet.
3. For server we are supporting 12 DIMMs on MC and 6 DIMMs on LS (platforms for DB are the same, but could be greater)
4. Not sure on that, would need to look it up.

[JF-AMD]

1. Istanbul has snoop-filter, hope Bulldozer has this
2. Rumoured to be DDR3-1866
3. Don't know (K10: 2 for desktop, 3-4 for server)
4. SSE1-SSE4.2, AVX, AES-NI, FMA3, FMA4, etc.

1. Yes
2. Not commented publicly

[w0mbat]

Oouu, JF is here so i cant talk freely ;-)

[blindbox]

Oouu, JF is here so i cant talk freely ;-)

This made me lol

[Oliverda]

Lisbon will be very similar to Istanbul, but it is a new stepping with some new features not found in Istanbul. Thuban is a desktop die that has features you won't find in server.

I think that the two dies will match but the server features will be disabled in Thuban and the desktop features will be disabled in Lisbon.

[Opteron146]

Lisbon will be very similar to Istanbul, but it is a new stepping with some new features not found in Istanbul. Thuban is a desktop die that has features you won't find in server.

How accurate is that information ?
Last time you commented on desktop products was something like "there wont be any desktop 6 core CPU"

I appreciate all your server information, but it seems to me that your desktop knowledge is limited. Has anything happened that changed this since the last 6core comment ?

Thanks


Edit:

I think that the two dies will match but the server features will be disabled in Thuban and the desktop features will be disabled in Lisbon.

I hope he is referring to more than unbuffered & registered memory

regards

Opteron146

[JF-AMD]

I think that the two dies will match but the server features will be disabled in Thuban and the desktop features will be disabled in Lisbon.

They may be different designs, not necessarily a pop/de-pop situation. I am less familiar with their plans and that is a general indicator that they are different designs. Or that I am just out of the loop

[JF-AMD]

How accurate is that information ?
Last time you commented on desktop products was something like "there wont be any desktop 6 core CPU"

I appreciate all your server information, but it seems to me that your desktop knowledge is limited. Has anything happened that changed this since the last 6core comment ?

Thanks


Edit:


I hope he is referring to more than unbuffered & registered memory

regards

Opteron146

Well, since I am in the server world, I don't comment on desktop plans. I have personally said in the past that I didn't see a need for a 6-core desktop processor. There were requests to take the Istanbul die and turn it into a Phenom and I said this would not happen because Istanbul is a server-only die. That continues to be true.

As a server guy I know little about the client world.

And bulldozer will support both unbuffered and registered memory. Those DIMM counts are for registered memory.

[Opteron146]

Well, since I am in the server world, I don't comment on desktop plans.

You just did

There were requests to take the Istanbul die and turn it into a Phenom and I said this would not happen because Istanbul is a server-only die. That continues to be true.

Ahh .. that is too much bean counter / marketing slang. Yes - Istanbul is a server die, because you marketing guys named it like that in the server segment / Socket-F package. Let's use the more practical engineering term "Hydra". A Hydra would fit nicely into an AM3 package, AMD's designs are very flexible, you should know that better than me

If it would be a great success in the desktop segment without a Turbo mode is another question.

Anyways, I guess your desktop colleagues would came up with a nice "desktop die" code name very soon - and of course that desktop die would be desktop only and be never be usable in servers.

It would have been the same as with the Greyhound die, which was used exclusively as Deneb desktop die. Once upon the time there were plans for special server dies named Shanghai and Suzuka. However it was not possible

And bulldozer will support both unbuffered and registered memory. Those DIMM counts are for registered memory.

Of course, because Bulldozer is not the exclusive server die code name. I told you that AMD designs are flexible, didnt I ?
However, we spoke about Lisbon and Thuban before. That are again the "special" desktop / server die names. I wonder if the difference is more than the memory support.

Thanks for the reply !

[Oliverda]

They may be different designs, not necessarily a pop/de-pop situation. I am less familiar with their plans and that is a general indicator that they are different designs. Or that I am just out of the loop

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.