Bulldozer for Servers: Testing AMD's "Interlagos" Opteron 6200 Series

[exar333]

It did better than I expected in those benchmarks, but the real SB Xeons (not the rebranded desktop stuff that you can buy now) are out Q1 of 2012.

I'm not sure where this pricing is coming from, and I can't divulge the actual pricing I see on systems due to contract reasons, but the price for the X5675 in the review is substantially higher than what I pay for them.

Real world price delta on the test systems with ONLY proc, chassis, memory, psu is around 8.5%. Add in storage, HBAs, etc, and your price delta is going to drop to 4.5-5% (or less) in favor of the AMD.

I don't really see a compelling reason to buy these at the moment, especially with the impending release of the SB Xeons. I'm a little disappointed in the way price is presented in this review, as when we look at purchasing servers, we don't look at the cost of individual components, but instead the entire platform cost. The price disparity when you're actually purchasing systems is very, very low (as I gave above)

How exactly are the desktop parts 're-branded' when the Xeon versions dont even exist?

Back on topic - This definitely means no price reductions for Intel Xeons. Expect the SB-Xeons to replace the existing products at the same price point with a small bump in clock-speed.

 

[beginner99]

Real world price delta on the test systems with ONLY proc, chassis, memory, psu is around 8.5%. Add in storage, HBAs, etc, and your price delta is going to drop to 4.5-5% (or less) in favor of the AMD.

Exactly. Wanted to point out the same thing, that total server cost isn't that much different. And if you then consider that BD needs a lot more power and hence cooling must be better /consumes more power (=more expensive) the price advantage basically doesn't exist anymore.

 

[Ferzerp]

How exactly are the desktop parts 're-branded' when the Xeon versions dont even exist?

Back on topic - This definitely means no price reductions for Intel Xeons. Expect the SB-Xeons to replace the existing products at the same price point with a small bump in clock-speed.

I'm talking about the Xeon 1xxx series that came out in the first half 2011. 1 socket, desktop proc called Xeon.

 

[Marcelo Viana]

I´m just wondering why they did´nt benchmark a 4 way interlagos yet! it´s not about a processor but platform. like: what is better, a system with intel, or a system with amd? they insist in things like core per core, and i ask: what it really means?
please don´t try to explain to me what is core vs. core benchmark, i work with it.
what i mean is put what amd have best against what intel have best, booth on servers, and try to find the better way to go, not simply compare parts, or you never go anywhere.
sorry for my bad english.

 

[Ferzerp]

I´m just wondering why they did´nt benchmark a 4 way interlagos yet! it´s not about a processor but platform. like: what is better, a system with intel, or a system with amd? they insist in things like core per core, and i ask: what it really means?
please don´t try to explain to me what is core vs. core benchmark, i work with it.
what i mean is put what amd have best against what intel have best, booth on servers, and try to find the better way to go, not simply compare parts, or you never go anywhere.
sorry for my bad english.

If you compare best Intel to best AMD, you're going to have to use an 8 socket E7-8870 on the intel side. That would be 80core/160thread vs. 32core/64thread on the AMD side. I don't think that's a very useful comparison

 

[Marcelo Viana]

it is, but you have to put costs on the counter in this way you fast find the market ninche of each one. Ferzerp i really believe that you can´t fearlly compare this two processor family, one say amd have 16 cores other 8 modules, intel have 6 cores, no 12(multitread), you see? they are diferent, but a system over linux windows, whatever, can be used by a company.
for example, a company need to build a new renderfarm, and the benchmarked system(not only cpu) intel(with 80 cores) can be 40%faster(just saying, i don´t know) that a amd system(32 cores), but maybe intel cost is about us$20,000,00(i know that is more expensive, just saying) and amd system is us$10,000.00. you can think with 2 amd system at the same price i can get 20% more performance
resume:
amd 10,000.00(60%performance) + 10,000.00(60%performance) = 20,000.00(120%performance)
intel 20,000.00(100%performance)
Ok now, the group of render softwares that is charge by system, not cpu or cores is example: mentalray 1,500.00, vueXstream 1,000.00 and others totaling 3,000.00.

Now the costs for a system is:
intel 20,000.00 + 3,000.00 = 23,000.00
amd (10,000.00+3,000.00)+(10,000.00+3,000.00) = 26,000.00
multiplying by a need number of systems, the amd numbers won´t be great, and i even talk about montly energy bill.
of corse, have situations where the amd will shine against Intel, but i can only talk about this things that the companys needs at a systems level, not processors level.
i hope you undestand my point.

 

[SolMiester]

BS got canned in the virtualisation benches and AFAIK, that what most servers are purchased for nowadays...

 

[brybir]

I agree. Llano indicates this, zambezi indicates this, interlagos indicates this.

But I don't think GlobalFoundries is really to blame. Before GF was spun-off, AMD execs made the decision (along with all the other members of the IBM fab eco-system) that 32nm was going to be gate-first.

This decision was made at the time knowing full-well that gate-first sacrificed parametric performance (clockspeeds and power efficiency) for lower production cost (fewer process steps in the fab, higher xtor density in the design).

Comparing GF's 32nm to Intel's 32nm, the delta between the two is "baked in" and doesn't really give the current fab engineers all that much room to finesse and optimize the stuff they traded-away 4yrs ago when their R&D engineers at Fishkill decided to go with gate-first HKMG integration.

GloFo's 22nm will be their first internally managed node, no influence from the AMD execs before the spin-off. But 22nm will still be hamstrung by the conflicting priorities of the IBM fab ecosystem. It will be gate-last but still problematic from a yield perspective.

IMO 14nm will likely be internalized and developed solely by GF in their Malta fab if they can get the staff they so desperately need. The jobs aren't getting filled though, headhunters keep calling. Its difficult to rebuild a top-notch R&D development team and the infrastructure needed once you've dismantled them. But GlobalFoundries is never going to best TSMC so long as GF's process roadmap is tied to IBM's decision maker IMHO.

One of the things that is interesting about the GloFo and IBM relationship is that IBM has the resources, if it wanted, to develop and "have" one of the best foundries in the world. I can only presume that since they do not, the business strategy that they are going forward with in PowerPC and custom IC designs does not require it to be "bleeding edge" in the way that AMD needs them to be. Is that your take on it?

Also, I do wonder if AMD knew a few years ago when planning to spin off the foundry that they would essentially giving up the high end consumer and server markets with that decision? They had to know that they could not compete with Intel as a stand along company and were probably hoping to leverage other companies to allow GloFo to become a world class foundry. But they had to know it could be 5-10 years for GloFo to become independent and truly competitive with Intel or TSMC and the others? So, why would they go forward with designs like BD that require impressive fab performance when they probably should have known that it would not be there?

 

[Idontcare]

One of the things that is interesting about the GloFo and IBM relationship is that IBM has the resources, if it wanted, to develop and "have" one of the best foundries in the world. I can only presume that since they do not, the business strategy that they are going forward with in PowerPC and custom IC designs does not require it to be "bleeding edge" in the way that AMD needs them to be. Is that your take on it?

Yep, its all about the confidence management has in the ability to sustain gross margins as time marches on, and the tradeoff in prioritizing the strategies that seem to have the best risk vs. reward in terms of delivering the projected ROI.

(that may sound like I am channeling the pointy-haired-boss in Dilbert, but I assure you it is not)

You can't pursue everything all at once, and often times it is easy to have more candidate projects that all claim to be able to turn a profit if pursued but you can't fund them all at the same time. So you rank-sort them.

Obviously the foundry effort at IBM didn't pan out in terms of its ranking once the projections were tallied.

To state it differently, IBM found more lucrative things to do with its money

Also, I do wonder if AMD knew a few years ago when planning to spin off the foundry that they would essentially giving up the high end consumer and server markets with that decision? They had to know that they could not compete with Intel as a stand along company and were probably hoping to leverage other companies to allow GloFo to become a world class foundry. But they had to know it could be 5-10 years for GloFo to become independent and truly competitive with Intel or TSMC and the others? So, why would they go forward with designs like BD that require impressive fab performance when they probably should have known that it would not be there?

AMD may have gotten exactly as they intended from GloFo, but they may have underestimated what the competive landscape was going to look like come the summer of 2011.

As such they have been in scramble mode trying to find another 10-20% performance and it is this scramble that makes it look like they didn't know what they were doing 4 yrs ago when they made the decisions they made.

Consider that when Zambezi was planned, Thuban was not intended for the desktop markets. Comparing Zambezi to Thuban was not planned when Zambezi was formulated. Once Istanbul was released to the server markets then it was repurposed for competion with Nehalem in the desktop markets.

Fast forward to 2011 and you've got the unintended consequence of having Thuban eating Zambezi's lunch on launch.

 

[hooflung]

BS got canned in the virtualisation benches and AFAIK, that what most servers are purchased for nowadays...

Their virtualization benchmark is pathetic. I'd call it a one trick pony benchmark. Old kernels. Old guests. No .NET apps to take care of testing enterprise use of AVX and nothing more than an Oracle Database test to test threading for java?

Did we just start drinking coolaid all of a sudden and not use our objectivity to actually read what the benchmarks do say vs what we want them to say (that amd is a failure lololol trollolololol it seems).

Even the reviewer said their is more testing to be done here...

 

[JohanAnandtech]

Their virtualization benchmark is pathetic. I'd call it a one trick pony benchmark. Old kernels. Old guests. No .NET apps to take care of testing enterprise use of AVX and nothing more than an Oracle Database test to test threading for java?
.

What do you think is running inside those VMs in enterprises? The latest and greatest?

And BTW, power management and CPU scheduling was handled by ESXi 5.0... which is the latest software. You do understand that the Hypervisor is in charge, not the guest OS?

And what is "enterprise use of AVX" ???

 

[CHADBOGA]

I don't know why Anandtech or any review site bothers with Benchmarks any more, they have been irrelevant ever since Conroe came out.

 

[zlejedi]

I don't know why Anandtech or any review site bothers with Benchmarks any more, they have been irrelevant ever since Conroe came out.

Ehh come on look at this forum. Even in place like this there's so much idiotism and misinformation coming from lack of knowledge