It's Istanbul (not Constantinople)...

[fellix]

AMD puts on the ritz with six-core Opteron demo

4S box on the bench, 24 cores at work!
"HT Assist" dedicates part of the L3 cache for snoop filtering -- aids SMP scaling.

 

[3dilettante]

The improved scaling is a good thing.
Istanbul at least pushes AMD's 4-socket products out of the embarassing situation of being outclassed by dual-socket Nehalem.

Perhaps with 6 cores, the chip will also return to the square shape of Barcelona, as Shanghai is pretty rectangular.
The chip is going to be bigger, and we'll have to see how AMD manages with another big chip, and how well it can clock with two extra full cores.

What does this mean for Shanghai?
That chip is significantly less compelling with Istanbul for a lot of workloads (particularly with the refusal to add the snoop filter to Shanghai), so this will segment AMD's product lineup with negative consequences for the smaller but higher volume chip.

There's always Phenom to take up the slack.

 

[fellix]

Perhaps with 6 cores, the chip will also return to the square shape of Barcelona, as Shanghai is pretty rectangular.

Not a chance, mate:

 

[3dilettante]

I'm taking note of the "not a real die" disclaimer in the image.
AMD could still use that configuration, so it might be right anyway.

 

[fellix]

Looks like AMD is already on the "stretched" path, as we see it here.

 

[3dilettante]

Which would happen if packaging two dies together, which is what the G34 socket is for.

 

[Ailuros]

A completely useless note on the thread title (and entirely OT), but Constantinopel (standing for Constantine's city) was often referred to as the "Polis" ("the City"). Istanbul means nothing else than "in the City" in greek or to be more precise InTheCity.

 

[fellix]

It is/was known also as Tsarigrad or The King's City -- a popular Slavonic nickname of the city, since the days of the Eastern Roman Empire.

 

[Tchock]

I have a feeling that Istanbul will have to come to desktops, and kick Deneb off its pedestal.


X6 would nudge K10.5 closer towards Nehalem in MT applications (depending on task characteristics it might be an overshoot or just catch-up), while harvesting for X4/X3, perhaps even an X5, would help AMD out. All for another 50mm^2 or so on diespace, yay or nay?

 

[fellix]

Which would happen if packaging two dies together, which is what the G34 socket is for.

So, G34 could in fact be a 2P system on a single package -- two 6-core dies each one wired for a total of four DDR3 channels.

 

[Tchock]

So, G34 could in fact be a 2P system on a single package -- two 6-core dies each one wired for a total of four DDR3 channels.

Could? It IS on the roadmap!

(Ohwait AMD revises roadmap faster than you can say wha-)


I wonder if they had any improvments on Sao Paolo which is after Istanbul (it's also a hexacore, as that makes sense WRT K10's core size) in terms of branching and prefetch. It's been long enough since Barcelona that they should have done something else besides just the probe filter. For starters, an implementation of Intel's SSE4/4.1 would be more than welcome for the DP market.

Interesting sidenote: The future G34 platform can use both unbuffered and registered DDR3. Looks like some aiming for the workstation platform too. Hmm.

 

[fellix]

Could? It IS on the roadmap!

(Ohwait AMD revises roadmap faster than you can say wha-)

I generally don't bother much with road-maps -- it's obvious, that a 1944 pin-out package isn't just for more robust grounding.