AMD's Six-Core Istanbul

[Idontcare]

Johan's AT Review of Istanbul

The Tech Report's review of Istanbul

At 346mm², I'm kinda surprised they didn't up the L3$ cache to keep the cache/core the same as Shanghai and then add some more to compensate for the cache loss incurred with HT Assist.

I'm also surprised they didn't go for an 8-core version to harvest down to the lesser SKU's. 346mm² is still pretty small in the scheme of things when comparing to Dunnington/Beckton/Tukwilla/GT200/etc.

I've got a request in to the gents at The Tech Report to see if I can get the actual data values for the Stars benches so I can do a breakdown of the scaling results as done in the past.

I got to admit, 75W ACP for 6 cores at 2.6GHz just flat impresses me for AMD considering they have no HKMG at 45nm. Unless they are binning silicon and tossing all the higher wattage stuff into the reclaim bin. I really thought 6 cores at 2.6GHz would need a 95W ACP rating. They did good considering their boundary conditions and budets :thumbsup:

 

[ShawnD1]

Originally posted by: Idontcare
I got to admit, 75W ACP for 6 cores at 2.6GHz just flat impresses me for AMD considering they have no HKMG at 45nm. Unless they are binning silicon and tossing all the higher wattage stuff into the reclaim bin. I really thought 6 cores at 2.6GHz would need a 95W ACP rating.

I think they already do that with other opterons. Shanghai chips are all 75W. Deneb are as low as 95W and as high as 125W.

Intel does the same thing. The 2.93ghz Xeon is 95W, i7 is 130W.

 

[Gikaseixas]

Anandtech should include X5550 numbers as this is the direct competitor to the 2435

 

[Idontcare]

Originally posted by: Gikaseixas
Anandtech should include X5550 numbers as this is the direct competitor to the 2435

I'm not so concerned about direct performance comparisons being made, I'm sure they did their best with the time they had and the hardware they had at their disposal.

The spirit of the message though is they should crunch some of the results as price/performance comparisons. After all that is exactly what end-users of the enterprise gear are going to do.

Also we can't really just compare absolute prices of just the CPU's as the price for price/performance comparisons. It's really the system price that a customer is looking at.

How does a system populated with X5550's compare to a system populated with 2435's when the system prices are used to normalize the price/performance.

Even if you want to just look at upgrading existing systems (which X5550 is not going to be for its an entire new platform) with the Istanbul you still can't use (should not just use) the 1ku prices as that fails to comprehend the actual installation costs (people/time/etc) and upgrade efforts that are going to go into the cost of upgrading the systems so as to deliver the performance sought after.

So should Johan have included X5550 in his review? Sure if he had one it would have been nice to have more data splattered across the graphs and the tables...but would it have changed the Istanbul performance results at all (answer: no) or would it have materially impacted the conclusions of the review (answer: not hardly)?

The story with Istanbul is very similar to that of PhenomII...if you already own an AMD system then it makes perfect price/performance reasoning to upgrade to the latest AMD CPU...but if you are in the market for an entirely new server/rig, be it Intel based or AMD based, then you need to factor in the total cost of the system for your price/performance and the new Nehalem-EP's (and i7's) make a compelling decision point.

 

[Gikaseixas]

Total cost is affected also by the cpu price but anyway, i just wanted to say that they should include the X5550 in this review as it is the more direct competitor to the 2435. I mean if you're on the market for a specific item it is very natural to look at other options, meaning other stuff around the same price range.

It was a constructive critique, i loved the review as usual, it is my fav site for that reason:beer:

 

[magreen]

Hey idc, what do you think will be the standard core count of a standard server cpu before 4P platforms start disappearing?

What I mean is, right now the standard is 4 cores/chip, with some 6 core chips available, meaning a 4p server would have 16 or 24 physical cores, plus SMT.

I know you've said Amdahl's Law limits how useful many-core systems will be when they get into the tens of cores per system. Whatever value of maximum useful cores is, let's call it x, so once the standard chip has x/2 cores, there's no purpose to a 4 socket system.

Or does virtualization tech make the many-core systems useful still, and negate the effects of Amdahl's Law?

(Hope I made sense, typing on a blackberry.)