Intel’s manycore (MIC) chip, formerly known as Knights Corner, will begin shipping next quarter as Xeon Phi. Although Xeon Phi can serve as a standalone processor, Intel is positioning it as a coprocessor because single-threaded applications will perform better on a standard Xeon. Manufactured in 22nm technology, Xeon Phi offers double-precision performance of greater than one teraflops (one trillion floating-point operations per second).

Integrating more than 50 CPU cores, Xeon Phi is aimed at applications that require a high level of parallelism, such as big-data analytics and financial services. Early benchmark results place systems based on Xeon and Xeon Phi among the top 500 supercomputers in the world. Although Intel has withheld the operating frequency, we estimate a 50-core Xeon Phi would need to operate at 1.25GHz to reach one teraflops.

Instead of selling Xeon Phi chips, Intel plans to offer the coprocessor only on PCI Express cards. The card consists of a single Xeon Phi chip and multiple GDDR5 SGRAM channels. This card virtualizes TCP/IP over PCIe, enabling Xeon Phi to run tasks independently of the host Xeon processor. It can also serve with InfiniBand for supercomputing applications.

By maintaining x86 compatibility, Xeon Phi is able to use the same software tools that Intel has developed for Xeon processors. Customers can recompile existing parallel software for the Xeon Phi coprocessor for nominal gains. To take full advantage of Xeon Phi’s vector processing engines, however, customers need to hand-optimize their software. For its target supercomputing market, programmers expect to spend time vectorizing and optimizing their code. Outside supercomputing and scientific workloads, however, the adoption of Xeon Phi may be limited.