The International Supercomputing Conference is underway in Leipzig, Germany, and Intel has used the occasion to reveal some new tidbits about its next-gen Xeon Phi processor. Dubbed Knights Landing, this massively parallel chip will pack more than 60 x86 cores. It's designed to plug into standard Xeon sockets, and the CPU isn't the only silicon on the package. Knights Landing also includes 16GB of on-package memory and a new Omni Scale Fabric that allows the processor to act as an "independent compute building block." Intel hasn't divulged too many details about the nature of the fabric, but the interconnect is based on IP acquired from Cray and QLogic, plus some of the firm's own special sauce.
There's also a rockin' promo video. Well, rockin' for Intel, anyway.
The Omni Scale Fabric will see action in more than just the next Xeon Phi. Intel plans a "full product line" that will include "adapters, edge switches, director switch systems, and open-source fabric management and software tools." The software will be compatible with Intel's current True Scale Fabric. Like the Xeon Phi processor, the Omni Scale Fabric controller will also be available separately on a PCIe add-in card.
As far as I can tell, the on-package fabric controller is the only truly new Xeon Phi detail revealed at the supercomputer show. The other elements of the processor are no less impressive, though. The compute cores are based on a version of the Silvermont architecture tweaked for high-performance computing. Micron collaborated with Intel on the integrated "MCDRAM" memory, which promises five times the bandwidth of DDR4 at one third the power. In all, Intel expects Knights Landing to boast three teraflops of double-precision horsepower and three times the single-threaded performance of the current-gen Xeon Phi.
The first Xeon Phi processors based on Knights Landing are scheduled to appear in HPC servers in the second half of 2015.