News - AMD's unannounced dodeca-core Opteron processor goes up for sale

[badass]

No, the Nehalem EX products have not been announced yet, so there are no benchmarks there.

We'll launch before the end of the month. We'll be able to compare benchmarks at that point and I think you'll find that when you compare price, performance and power consumption (the 3 things that matter the most for customers), you'll see AMD clearly in the lead.

But I wouldn't put too much faith in VMmark benchmarks. Even VMware has said they are outdated. I have never met a customer who has 100 VMs configured with ~500-700MB of memory on a single platform.

Most 2P customers have 5-10 VM's with ~4-8GB of memory configured; 4P customers tend to have 10-20VMs.

What will make Magny Cours shine for virtualization is the large number of real cores. 10 VMs with 2 vCPUs per VM means that on a 6-core 2P you have 20 vCPUs fighting for 12 total cores, so you are emptying cache more often. For AMD, you'll have 20 vCPUs with 24 total cores, so cache data gets to stay in place an a lot of the latency in virtualization is reduced.

http://it.anandtech.com/IT/showdoc.aspx?i=3769

Looks like a 2 CPU, 12 Core Westmere server is about as fast as a 4 CPU, 24 Core Opteron Server. including in virtualisation.
The Dual CPU server is cheaper and uses less power. You can use all of the technical explanations in the world about how AMD's processors will be faster, but the benchmarks do not back that up.
AMD can only compete on price against intel now. At everything else, they are worse.

[JF-AMD]

Looks like a 2 CPU, 12 Core Westmere server is about as fast as a 4 CPU, 24 Core Opteron Server. including in virtualisation.
The Dual CPU server is cheaper and uses less power. You can use all of the technical explanations in the world about how AMD's processors will be faster, but the benchmarks do not back that up.
AMD can only compete on price against intel now. At everything else, they are worse.

Those benchmarks reflect 6-core processors in a 2006-era platform. Our new platform will launch before the end of the month and you should compare at that time.

Based on the Westmere benchmarks that have been released, you'll be real interested to see how our new platforms perform in terms of power efficiency, price, and, most importantly, performance.

[badass]

I have just read that the 12 core Opterons will have 16 or 18 MB cache.
This makes them much more interesting, particularily if that cache is used properly and just maybe they will be significantly faster per core per MHz than the old 4 core 6MB opterons.

[JF-AMD]

The processors have 12MB of L3 and 512MB of L2 per core, which is how you get to those numbers.

We launched officially about an hour ago.

Compared to a westmere (X5680) our N-1 (2.2GHZ, NOT our top bin) was faster than the X5680, drew less power and is ~$500 less expensive.

Looks like someone misjudged how good MC was going to be when they priced their products.

[Funkstar]

The processors have 12MB of L3 and 512MB of L2 per core, which is how you get to those numbers.

Really? impressive

[badass]

The processors have 12MB of L3 and 512MB of L2 per core, which is how you get to those numbers.

We launched officially about an hour ago.

Compared to a westmere (X5680) our N-1 (2.2GHZ, NOT our top bin) was faster than the X5680, drew less power and is ~$500 less expensive.

Looks like someone misjudged how good MC was going to be when they priced their products.

Now you have my interest. /me goes benchmark hunting.

[badass]

Now you have my interest. /me goes benchmark hunting.

Massive dissapointment
A Dual X5670 beats out Dual 6174's in almost everything, including virtualisation.
Looks like AMD is back to playing second fiddle in the server market still.

[pricey4907]

Competition is healthy for the heart, but Gulftown is here!!!

[JF-AMD]

Massive dissapointment
A Dual X5670 beats out Dual 6174's in almost everything, including virtualisation.
Looks like AMD is back to playing second fiddle in the server market still.

I have to disagree.

If you look at raw performance data, you see that our N-1 (6174) outperforms a dual x5680 in integer and outperforms in FP by ~20%. (both SPEC).

Most of the reviews are not using server benchmarks, they are using client benchmarks like cinebench.

One used the Oracle Calling Circle benchmark, but that one only scales to 16 threads, so it is a bad indicator of how a 24 thread application is going to run.

Tec Channel has a better review.

Also, don't forget that the 6174 is $1165 and the x5670 is $1386 with the x5680 being $1663.