The Xeon Phi chip, code-named Knights Landing, offers an array of new technologies that collectively deliver performance breakthroughs. The chip is also a springboard for new memory, I/O and storage technologies destined to reach desktops and laptops in the coming years.
Intel didn't provide details on the first supercomputers with Knights Landing. The U.S. Department of Energy, however, said that the chip will be used in Cori, a 9,300-core supercomputer that will be deployed in the latter half of 2016 at the National Energy Research Scientific Computing Center in Berkeley, California.
More than 50 manufacturers will ship systems using Knights Landing, said Charlie Wuischpard, vice president of the Data Center Group and general manager of the HPC group at Intel.
Xeon Phi chips are used in the world's fastest supercomputer, code-named Tianhe-2, which is installed in China. Knights Landing will succeed the current version of Xeon Phi, code-named Knights Corner, which has 61 cores.
Like its predecessors, Knights Landing will be a coprocessor alongside CPUs and assist in complex calculations. However, it is different in design from graphics processors like Nvidia's Tesla, which are used for the same purposes.
The purposes for which supercomputers are used has changed over the years, and machine learning is one of many new ways to get the best out of high-performance machines. Intel is providing the software, networking and other tools to help in such workloads, said Wuischpard.
Knights Landing will deliver double-precision peak performance of more than 3 teraflops and single-precision performance of more than 8 teraflops. That's in the same range of some other graphics chips used in the world's fastest supercomputers. Double-precision is important for supercomputing due to the higher accuracy rates in floating-point calculations than single-precision calculations.
Knights Landing has 16GB of on-package MCDRAM memory, in which modules are stacked. It offers five time more bandwidth than emerging DDR4 memory technology. Intel claims the memory is five times more power efficient and three times denser than GDDR5, which is used on graphics cards.
Intel is also adding a new interconnect called OmniPath, a fabric that the chip maker says can boost overall supercomputing performance. The technology helps move data faster between storage, memory, processors and other components. It will first appear in supercomputers, though Intel is making the technology available to server makers. OmniPath is a proprietary interconnect, and Intel has not detailed the technology.
The goal of new technologies like Xeon Phi is to squeeze more performance out of supercomputers without drawing more power. The software element is also important in achieving that goal, said Dean McCarron, principal analyst at Mercury Research.
Intel and Nvidia provide programming tools for developers to distribute processing over the hundreds of thousands of cores in the supercomputers.