You can scratch the Xeon Phi off your shopping list. And if you deployed it, don’t plan on upgrades. That’s because Intel has quietly killed off its high-performance computing co-processor because forthcoming Xeon chips have all the features of the Phi, no separate chip or add-in card needed.
Intel quietly ended the life of the Xeon Phi on July 23 with a “Product Change Notification” that contained Product Discontinuance/End of Life information for the entire Knight’s Landing line of Xeon Phis.
The writing was on the wall when Avinash Sodani, chief architect of the Knights Landing chip, took a job at Cavium in September 2016. But really, he likely saw what was coming: that the Xeon Phi would be obsoleted by forthcoming Xeons.
Intel has a new line of Xeons coming, called the Xeon Scalable Platform. These are based on the newer Skylake architecture, which represented a pretty big leap over the prior generation of Haswell and Broadwell.
Xeon Phi innovations rolled into Xeon Scalable Platform
Many of the innovations going into the Phi were effectively rolled into the Xeon Scalable Platform, such as AVX512 extensions and other acceleration capabilities. So, the Phi was effectively redundant. This is history repeating itself going on 30 years. I assume many of my readers are old enough to remember the 80387 math co-processor. I was doing computer sales at the time, sold a bunch with PCs to offices doing Lotus 1-2-3 work. Then came the 80486 with a built-in math co-processor. The same process has repeated itself over and over in the last three decades.
Intel has also decided to mess with its branding. Instead of the Xeon E3/E5/E7 for low/mid/high end, it now has four “metal” brands: Bronze, Silver, Gold, and Platinum. Gold and Platinum are for HPC customers who would otherwise have used a Xeon Phi.
Intel is still offloading work to other processors. It has the Stratix line of FPGAs, plus AI-oriented chips from Nervana and Movidius. One has to wonder how long before they get rolled into the Xeon chip, as well.
The Xeon Phi wasn’t exactly a huge hit for Intel. Only seven systems in the June Top500 list of supercomputers had Xeon Phi co-processors compared to 80 running Nvidia GPUs, and that was down from 10 on the November 2017 list.