Intel recently released a new generation of the enterprise CPU. These new CPUs, version 4 of the venerable E5 and E7 product lines, have some pretty impressive specs: more cache, more cores, faster RAM and more RAM, too. The interesting part about this release, though, is that nobody seemed to notice.
Maybe it's all this cloud stuff. Maybe we've moved past caring about hardware. The higher you get in an IT command structure, the more you hear that hardware doesn't matter, that applications and the cloud are the future. I don't disagree -- applications are the tip of the spear in IT, because applications are what get things done. Problem is, the underlying hardware has everything to do with both the performance and the economics of those applications -- even in the cloud.
Nothing punctuates the idea that hardware matters more than looking at a cloud provider's offerings. If cloud hardware doesn't matter, why does Amazon have instance types, all with drastically different specifications? No matter how much a CIO doesn't want it to be true, someone on their staff must be adept enough to mate an application with an instance type. A wrong matchup between apps and instance types means poor performance, higher costs and wasted staff time -- not unlike the risks associated with matching apps to on-premises equipment.
If we look closely at the descriptions of cloud instances we also see the generations of hardware underneath the instances, too. In Amazon's case, we see cloud hardware as old as the first version of the E5 CPUs. Given the processor types they list, we know that its version 1 servers have eight cores per socket, or 16 per server. These new CPUs that Intel just released have, at the maximum, 22 cores per socket, or 44 per server. With the additional cache and manufacturing improvements, we'll call it a 5x improvement in performance, an estimate backed by CPU throughput benchmark listings on SPEC.org.
Five times the performance is nothing to sneer at, and I'd guess there will be much less than a 5x price difference between the two instance types. Maybe your organization is smart enough to choose instance types wisely.
Five times the performance is also nothing to sneer at in your own data center. The original E5 CPUs were released in the first quarter of 2012, so that fivefold increase in performance comes over the span of only four years. Along with improved performance, the newer enterprise CPU offers better manageability and lower power consumption. As a result, I've started to think that perhaps the problem with on-premises equipment is that we don't replace it often enough. Public cloud providers don't run VMware software, they run open source hypervisors, largely because they want to reduce license costs. Public cloud providers also license other software, like operating systems, at a per-VM level, whereas we typically assign socket-based licenses to hosts for our on-premises equipment.
We tend to hang on to our servers, thinking that if we can run them for five, six or seven years they become very cost effective. It's really the opposite, though. After four years, I could have five times the computing power in the same footprint as my old servers, at the same price point. What does that do to my license costs? It should halve them, at least. Then what does that do to my overall cost of ownership? And that is what I'm afraid of, because it means in our rush to distance ourselves from our hardware -- mistakenly thinking that the cloud means no hardware -- we've removed ourselves from important cost savings discussions. Intel's enterprise CPU is the latest example of this.
Maybe the private cloud and public cloud comparison is more apples to apples than we thought.
Bob Plankers is a virtualization and cloud expert at a major Midwestern University.
Determine the right cloud hardware for your IT applications
Discover how to choose the right CPU features for virtualization