June 07, 2013, 1:58 PM — Intel is planning a two-day server-oriented event in San Francisco next month, which would likely be the next phase of the Haswell generation rollout. Traditionally the company always rolls out the new desktop/mobile line first and the server product always comes shortly after that.
The server market is an entirely different animal for Intel and the pressures there are very different and complex. The challenges of the desktop market were rather clearly defined: PCs were being heavily pressured by the tablet market and a poorly-received new OS from Microsoft.
The server side is a little different. First, Intel is in a safer position. AMD, which once held 20% of the server market, has shriveled to around 5% of the server market. The RISC/Unix market is in a freefall. If people are buying servers, it's x86 servers, and with the cloud buildout that Microsoft, Google, Amazon and other vendors are engaged in, they are buying servers in units of 10,000.
But there are still challengers. First, you've got Nvidia, a company that has laid waste to more competitors than Intel. Looking to expand beyond the desktop PC market, the company has set its sights on high performance and heavy compute loads, arguing that its GPUs, essentially giant math coprocessors, are better suited for many mathematical calculations like visualization.
The Top 500 supercomputer list is due this month and you can bet there will be more GPU-powered systems on that list than there were the last time. Nvidia has been on that upward slope for a while now.
On the low end, Intel is facing power challenges from ARM and its own Atom line. Several vendors, starting with SeaMicro, have argued that there are a lot of server tasks that don't require a big, powerful CPU like the Xeon. A file and print server or spitting out HTML pages can be well-served by a low-power ARM or Atom processor using a fraction of the wattage.
So there's no smooth sailing for Intel in servers, either. It has to compete with itself, among others. It has a considerable advantage over the ARM processor in that it's got a 64-bit offering, something ARM won't have for at least another year. Intel launched the 64-bit Centerton family in December of last year, with features like Intel VT virtualization technology and support for ECC memory, something ARM and other Atoms don't have.
I'm curious to see how the new Xeons turn out, because the new desktop generation of CPUs aren't much faster than the old Ivy Bridge. Tests from Tom's Guide showed that as a CPU, Haswell was only slightly faster than Ivy Bridge. Most of the advances were in graphics performance.
In the end, Haswell may prove a competitive advantage over AMD's Opteron and ARM because of its GPU technology after all. According to Cisco, Internet video/TV/VoD and other video traffic will account for 80% to 90% of Internet traffic by 2017. Besides the fact that's a downright scary stat, something has to process all that video, and Haswell is far better equipped than any prior generation of Xeons to handle it.
Intel has already released a few products for this market. The E3-1275 v3 processor provides the capabilities needed to deliver an improved video content with the data center network. Intel's upcoming E-3 1230L v3 is expected to have a significant performance per watt improvement.
Then there's the Xeon Phi, the final product from the Larrabee boondoggle. Larrabee was a mess but at least with Xeon Phi, Intel got a payoff. The first system deployed with Xeon Phi is a very impressive beast, the Texas Advanced Computing Center (TACC) supercomputer Stampede, a 10 petaFLOP supercomputer. Suddenly, Intel had an answer to Nvidia and its Tesla GPUs for HPC.
It will be interesting to see how things turn out, because Intel isn't going to be able to sell Haswell Xeons when they are only 5-10% faster than Ivy Bridge, like the desktops are.