The Cranky Admin
Shakeup in the CPU Market Coming This Year
Things are moving in a positive direction for admins.
2017 is going to see some disruption to the normally staid and boring CPU market. Intel's Xeon v5 chips come out for 2P and larger servers. AMD is promising Zen-based CPUs that can challenge Intel, and 64-bit ARM server CPUs are being pushed in combination with Linux and containerization to become serious contenders.
Of all the new silicon available to datacenters in 2017, Intel's offerings are probably the least interesting. The v5 Xeons look to be reasonable and capable incremental improvements over their predecessors, with no big surprises or new segments expected.
Intel continues to crank out the Xeon Phi processors aimed at challenging NVIDIA for machine learning and supercomputing tasks; however, these are not expected to have any meaningful impact on virtualization or other, more mainstream workloads in 2017.
On the other side of the spectrum, the Xeon-D chips are gaining a cult following, and we're about due for a proper refresh. These low power, single socket chips support 128GB RAM per node and make excellent SMB starter clusters; especially given that they tend to be found on boards designed for 5- or even 7-year life cycles.
AMD's Huge Hurdles
AMD has a lot to prove. After falling dramatically behind Intel on Instructions Per Clock (IPC), and flubbing the floating point capabilities of the previous generation, AMD functionally disappeared from the datacenter. It has been years since I've seen anyone discuss using AMD servers in the wild. Indeed, even with all the many and varied technological communities that I am part of, for several years when discussing new servers it was always taken for granted that Intel would be the CPU vendor.
This intellectual inertia, more than anything, is what AMD has to combat. Intel isn't simply the primary supplier. They aren't merely the default choice. Intel has become so absolutely dominant that choosing another supplier isn't merely radical, isn't just heresy: replacing Intel in the datacenter is so taboo that it is inconceivable.
To overcome this, AMD needs more than just good chips that perform well: AMD needs mindshare. They need to seed the market by capturing vendors' attention, the interest of "thought leaders," community leaders and internal champions within large organizations. Preferably large organizations that like to talk a lot about why they chose to buy what they bought.
AMD has a marketing and PR war to engage in that I am not convinced it is capable of waging. It has been traditionally stingy with community engagement, and not very inclined to seed product where it would do the most good. AMD's success in 2017 -- and the likelihood of making it into our datacenters -- will depend entirely on its ability to change how it approaches reputation management. That may be harder for it than redesigning its chips.
ARM Wrestling
ARM processors face a similar challenge. Unlike AMD, however, using ARM server CPUs isn't heresy. AMD has to fight the good fight against Intel, seeking a like-for-like replacement. ARM can get into datacenters through the side door: they're for Linux and containers, you see. That new stuff that requires you rewrite all your code and change your approaches to application management anyways. Why not run it on a different CPU architecture?
The ARM CPU vendors don't even have to make that big a splash. If they can get a handful of Fortune 500 companies to bite, they have enough to go to the public cloud vendors and say "if you want that business, you need to light up our chips on your clouds." To survive, AMD needs to sell a whole lot of chips; ARM vendors, on the other hand, can sell a fraction of AMD's requirements and call it a huge win.
For systems administrators, all of this means good times on the CPU side. Additional choice means the ability to hold vendors' feet to the fire. More importantly, it also means we can stop thinking of two socket servers as the only building block for our datacenters.
The Power of One
You'd be surprised what you can get done with single socket servers these days, especially if your thoughts on virtualization encompass not just hypervisors but containers as well. Servers are getting more and more capable, but our applications aren't necessarily getting more demanding. This may well be the year we rethink the minimum size of our compute infrastructure ... and hopefully pay less for each piece.
About the Author
Trevor Pott is a full-time nerd from Edmonton, Alberta, Canada. He splits his time between systems administration, technology writing, and consulting. As a consultant he helps Silicon Valley startups better understand systems administrators and how to sell to them.