AMD posts controversial virtualization value math

Posted

Opinion – AMD has its hands full to counter Intel’s Xeon 5500 processor, which is considered the highest performing server processor these days. Intel’s chip has some obvious weaknesses which AMD could go after, but we haven’t seen much of that yet. Instead, there is a new post discussing the dollar value of virtualization performance delivered, in a way that is easy for Intel to attack and for server customers to scratch their head over.

The core idea of AMD’s post “Simply Spectacular Virtualization” is to answer an interesting question. How much dollar value do you get out of a server in a virtualized environment? How much does a certain kind of virtualized performance cost? And: What is cheaper? An Intel Nehalem-based server or AMD’s Shanghai?

You already know the answer to this question since it was AMD’s post. And reading such a post on AMD’s site you can expect that the evaluation environment is skewed towards AMD’s advantages. For example, AMD admits that Nehalem offers more performance, but claims that few server buyers opt for the highest performing chip and that real world environments do not run the 100 VMs per server the highest performing servers are capable of today. Instead, most customers, the company claims, are running just 15 to 20 VMs - which means that lower performing (and cheaper) processors are just fine for today’s IT environments.

In this example, AMD compares two Dell (AMD/Intel) and two HP (AMD/Intel) servers in ways they can be configured directly from the firm’s websites. AMD said that it configured the systems as closely as possible to the systems used in public VMWare benchmark results. AMD released a presentation with detailed information of its comparison, which concluded that the Intel HP servers are up to 103% more expensive than the AMD system in terms of cost for virtualized performance, while the Dell systems are 36% more expensive.

Makes sense, right? Well, sort of. The problem is that the initial post compares systems that are that were tuned for maximum performance, not price. As a result, the Intel systems used 96 GB of memory and 12 expensive 8 GB memory modules costing close to $10,000 in upgrades over the standard 4 GB memory in those servers. The AMD systems, however, only have 64 GB memory and use 4 GB memory modules. So there is an obvious inflation of price that puts AMD at an unfair advantage because of the claim that performance is not the primary concern, while the VMWare benchmark was focused on performance.

When we talked to AMD, the company said that it would update its post in the coming days, since Dell and HP are now offering 8 GB memory modules for AMD’s servers as well and that should level out the playing field. Well, almost. Intel’s servers are still equipped with 96 GB and AMD’s servers with 64 GB, in an environment where a much cheaper Intel system would reach that 15-20 VMs performance easily as well. And yes, Dell, for example, allows customers to configure their Nehalem server with less memory, for example with 72 GB of memory using 18 4 GB memory modules costing $2880 in upgrades instead of the $9880 upgrade for 12 8 GB devices. There is even a $6480 option for eight 8 GB modules. Memory is the most expensive factor in AMD’s cost-focused comparison, which makes this whole blog post look, well, strange.
 
Even if AMD said it will update its post, it will not include an apples-to-apples comparison with the same amount of memory, simply because AMD claims there are no official VM benchmarks available for Intel 64 GB Nehalem systems. However, we talked to industry sources who claim they do exist in a non-public manner – and which show that Nehalem servers may actually offer an advantage over Shanghai in a cost per virtualization performance discipline. AMD agreed that a 64 GB vs. 64 GB comparison makes sense, but it would take someone to actually run such a public benchmark.

I see AMD’s point to look at available systems, actual required real world performance and what the cost of that performance would be. But it is easy for Intel to attack that equation, especially in the way it was published. And yes, I would also say that the comparison isn’t fair. The implications or information included in this post is, for actual customers, virtually irrelevant. Large customers test such servers on their applications anyway before they buy the hardware and may come up with entirely different results based on their performance and power requirements. But AMD asked an interesting question about the actual cost of performance. So, are there any takers who would compare a 64 GB Nehalem system against a 64 GB Shanghai system in the VMWare benchmark? Chime in to let us know what you think!
 

Wolfgang Gruener is the founder of TG Daily. The opinions expressed in this commentary are solely those of the author.