Is x86 still relevant in the datacenter space?

IN the last 20 years we have seen the data-center landscape transform massively. From a bunch of racks dotted across the globe with per application hardware and single core CPUs - to vast sprawling data-centers, consuming as much power as a small country, located in obscure locations around most countries in the world. From bare metal, through virtualisation (many servers/applications running on a single piece of hardware) and then into cloud (individual services on demand, running on whatever infrastructure they are needed on at the time).

Do the traditional server giants like Intel with their somewhat confusing Xeon product line up still remain relevant in these new, automated datacenters with their "as a service" based cloud infrastructures?

A change is happening recently, driven by increased knowledge around silicon design and programming - the trend very much seems to be "Accelerate all the things, and do it in hardware". Something that Intel may well be positioned to capitalise on, but an area at the moment we are seeing relatively new players driving some real innovation in.

For the uninitiated, the process is quite simple whilst actually hugely complicated at the same time. Great! Stay with me - I will try to keep this as simple as possible!

For years and years we (and I say we, in the loosest sense - much smarter people than me!) have made programs and ran them on general purpose CPU's like those which Intel make. Way back since we had things like DOS and Windows 3.1, if you are old enough to remember.

These Intel, and compatible CPUs (now really only AMD but there used to be people like Cyrix and others I forget) have an underlying codebase, or function set called x86. All instructions must eventually end up running in this x86 "language" if you like. Whether programming in Python, HTML or C , it all ends up in this x86 format in the end before the CPU can do anything with it. The thing is that as we all know, any translation is not hugely efficient and over time new (and old) ways of doing things have matured. So x86 is no longer the "be all and end all" of the compute industry.

These days many competitors to x86 exist - without even mentioning remnants of the old competing "Power" architecture. One of those competitors generated a completely new market for low power devices - ARM, who are now the incumbent in the mobile phone space, and have shown that x86 is not the only way. Indeed, it is not just in the mobile space. Nvidia (and AMD with their Radeon line) have become a huge player in the semiconductor market over the last 20 years as well - driving significant revenue and market share with another non-x86 product - the GPU.

Part of it is about design, and the origins of the designs engineers find themselves with today. The GPU is almost inherently parallel by design, featuring thousands of small cores and having more and more each generation. Whereas the CPU has always basically morphed from its single core origins into a multi-core version of itself to allow it to do more things at once. Think of it a bit like throwing another man at a job every time you want it to go faster. Eventually the room is full and nobody can do anything very quickly because they are all either waiting for the bathroom or queuing for food. The analagy suits itself well to memory bandwidth and IO starved cores in my mind.

This parallelism found in GPU architectures has lent itself well to modern programming models and demanded a large share of wallet in the data-center. Parallelism becomes ever more important in driving performance increases and new discoveries - particularly in the fields of high end research, AI and Machine Learning.

However there is an old dog lurking around, with some fancy new tricks.

The FPGA has been around for a long time but recently has become far more accessible to the datacenter. With recent developments in FPGA manufacturing, increased speed, lower cost, less real-estate, and increased possibilities for re-programming 'on-the-fly'" rather than fixed with a set of functions like an x86 CPU. Whilst this may be difficult to get your head around as a traditional x86 programmer, FPGA solutions can provide flexible acceleration for certain workloads, and operate at higher throughput's than competing general purpose solutions like CPUs and GPUs. FPGA's are also becoming far easier to program, with companies like Xilinx driving leaps ahead in the programming tools and APIs necessary to work with these devices. Another great advantage is that once you have an FPGA with the set of features you require, it is easier to turn it into an ASIC (Application Specific Integrated Circuit), and potentially enable even higher performance. Once a programmer, now a chip designer - but that is a whole other story.

Recently we see huge developments in this space particularly relevant in large datacenter and the cloud. As FPGA solutions find their way into the Smart NIC world. We start to see reduced latency for fixed functions such as trading become commonplace. However it wont stop there. NIC based acceleration of TCP has been around for a while. "Kernel Bypass" has also become a standard used to avoid the OS altogether when doing complex functions involving both the CPU and NIC. Now we have things like "Cloud Onload", from Solarflare, who were recently bought by Xilinx, an FPGA company. Or Exablaze, who's NIC products blur the lines between FPGA and NIC altogether. Interested in NVMeOF - probably coming to an FPGA near you soon!

Oh, and don't count Intel out yet, especially with their recent acquisition of one of the largest FPGA companies in the world, Altera, an acquisition that has been long in the works. Did I mention Intel have x86 CPUs with FPGAs inside? Oh and there is that Intel Xe GPU in the works I keep hearing about...........

Just when you thought it was all starting to make sense.........