Bulldozer Opterons (Interlagos) as a server offering

[loophole]

I just read Anandtech's Opteron 6276 follow-on article:
http://www.anandtech.com/show/5279/t...-a-closer-look
and I figured I'd post it here in the AMD section because it had some good analysis on different workloads running on Interlagos and how well they scaled, how they were affected by disabling CMT (only one integer core per module enabled) and comparing it to SMT (see page 6). They also profile various server workloads and look at the resultant IPC which I haven't seen very often in a review.

I thought this snippet (from page 8) was interesting (they focus on few cores and threads because it seems the 1st gen Bulldozer are hampered by lock contention issues more than the Westmere Xeons):

First, we focus on the results with few cores and threads. Two Bulldozer Modules are capable of slightly outperforming four cores of the Opteron Magny-Cours. The ideas behind Bulldozer are sound: two modules are smaller (157 mm²) and more power efficient than four K10 cores (231 mm²). At the same time they perform equal to the Xeon X5650—which is clocked higher—with the same amount of threads. At eight threads this is still the case, and the gap between the newer and older Opteron widens in favor of the former.

It is perhaps the first positive thing I've read about AMD's module-based design (Clustered Multi Threading). (and it came from Anandtech which is often regarded as being pro-Intel...)

So the above article is the follow-on. If you're after the first look article, it's here:
http://www.anandtech.com/show/5058/a...nterlagos-6200

[bobbth]

Originally Posted by loophole

I just read Anandtech's Opteron 6276 follow-on article:
http://www.anandtech.com/show/5279/t...-a-closer-look
and I figured I'd post it here in the AMD section because it had some good analysis on different workloads running on Interlagos and how well they scaled, how they were affected by disabling CMT (only one integer core per module enabled) and comparing it to SMT (see page 6). They also profile various server workloads and look at the resultant IPC which I haven't seen very often in a review.

I thought this snippet (from page 8) was interesting (they focus on few cores and threads because it seems the 1st gen Bulldozer are hampered by lock contention issues more than the Westmere Xeons):


It is perhaps the first positive thing I've read about AMD's module-based design (Clustered Multi Threading). (and it came from Anandtech which is often regarded as being pro-Intel...)

So the above article is the follow-on. If you're after the first look article, it's here:
http://www.anandtech.com/show/5058/a...nterlagos-6200

The bulldozer architecture was claimed to be a server oriented architecture, its little surprise to see it doing well, it was partially aimed at clawing back some server market for AMD (they had as little as a 7% market share at one stage)

[gregpolk]

Has it done so?

[balckjok3r]

It does well at everything exept games....

[loophole]

Originally Posted by gregpolk

Has it done so?

It's hard to say, as It takes a long time to win OEMs over (it took them years to win 1st tier server vendors over with the original K8 Opteron too, and rather than just being competitive like Interlagos is, it was an absolute winner back then) so we'll probably see the results in several quarters time. And of course it'll depend on how much the server makers believe in AMD's Opteron roadmap - whether they think they can indeed deliver what they say they can.

The fact that Interlagos drops into the existing socket G34 infrastructure means that most server makers that came out with Magny Cours offerings will probably also do Interlagos systems. And Piledriver Opterons being a drop-in replacement will also help - less to qualify for the OEMs to make "new" products.