Intel bolsters massive parallel processor
Intel announced at the International Supercomputer Conference in Germany the development on its Xeon Phi. In addition, the company unveiled a new interconnection fabric based on the company's silicon photonics advances and an educational programme designed to allow new programmers to learn how to code for parallel processors.
"In just 16 years, we've seen the fastest supercomputer in the world at 3TFLOPS migrate down to a single socket," Charles Wuischpard, VP and general manager of Intel Workstations and High Performance Computing Data Center Group told EE Times in a conference call.
Intel has been talking about a new version of its massively parallel Xeon Phi processor¡ªcurrently with 60 cores per chip¡ªbut at ISC-14 unveiled many more, but not all, of the details about the new chips, which will be available in the second half of 2015. Its current-generation Xeon Phi is a 1TFLOP chip cast in 22nm CMOS and sold on a PCIe board in several versions. The Green500 list pronounced it the most power-efficient parallel processor in the world. The Top500 supercomputer list just announced the Xeon Phi powered Tianhe-2 (Milky Way 2) supercomputer at the National Supercomputing Center in Guangzhou, China, as the fastest in the world for the third time running.
"We have a new Xeon Phi processor coming in the future called Knights Landing," said Wuischpard. "The first thing to note is that it will be 3TFLOPS in a single package, which will be available in the second half of 2015, with at least as many processors, but based on the Silvermount architecture and connected by a low-latency mesh."
Figure 1: Intel predicts that high-performance computers will grow at a rate of 20 per cent per year as prices drop, inducing more segments to purchase them. (Source: Intel)
The previous MIC processor¡ªcalled Knights Corner¡ªwas based on a special microarchitecture created just for it, but the new-generation Knights Landing Xeon Phi will be based on Silvermont modified for Intel's 14nm process. The Silvermont architecture has also been heavily modified to add key features, such as an AVX512 vector processor and four threads per core, versus the current Silvermont, which has no AVX512 support and just one thread per core.
|Related Articles||Editor's Choice|