At the recent Intel Developer Forum show in China, Intel showed off more information on Knight's Landing, the third generation of its Xeon Phi coprocessor that it launched in 2010 as a competitor to Nvidia's GPUs in high-performance computing.
The information comes from a session guide (PDF) hosted by WCCF Tech and was meant as a supplement for attendees of its recent IDF China show.
The next generation Xeon Phi will come with 72 Silvermont cores connected with a high speed mesh and 36MB of shared L2 cache. This is a pretty big change because previous generations of Xeon Phi used 36 or 50 Pentium 1 cores. Yes, Pentium 1, a design from 1995. Silvermont is a much more advanced architecture used in the Atom processor. Silvermont will support four threads per core and have dual 512-bit vector units per core.
The coprocessor will come with up to 16GB of high-bandwidth memory on-board plus will support six channels of DDR4 2400 memory and can access up to 384GB of system memory total. With new on-package memory jointly developed with Micron, Intel claims the on-board memory will have five times the bandwidth of DDR4 and five times the power efficiency. Knight's Landing will also support a whopping 36 PCIe Gen 3 lanes.
Intel claims this will give up to three times single thread performance over the previous generation, called Knight's Corner, while at the same time be three times more power efficient as the previous generation. All told, it predicts up to teraFLOPS of performance per card. Older cards averaged 1 teraFLOP.
The talk actually spent little time discussing the hardware and was more focused on parallelization and optimization of code for multithreaded and parallel processing. The discussion was focused around using Intel compilers and optimized libraries along with removing data transfer bottlenecks, which Knight's Landing will certainly do with all its high-speed interconnects.
The Xeon Phi cards have become increasingly important in supercomputing, with many of the top supercomputers in the world using it, including Tianhe-2, the fastest supercomputer in the world. Tianhe-2 is based in China, and it won't be getting any more upgrades now that the government has nixed Xeon sales to China.