The Larrabee project was something of a mess for Intel, but it's finally paying off for the firm. A Xeon Phi-powered supercomputer sits atop the Top 500 list of supercomputers in the world, and Intel plans to bring out new chips that will serve not only as accelerators to the Xeon but as replacements, too.
Larrabee was an attempt at a GPU that Intel reinvented as a high-speed, many core co-processor similar to how Nvidia is using its GPUs in supercomputing. The Larrabee GPU project became the Xeon Phi, a co-processor for servers that comes on PCI Express cards. The cores are x86-based but highly simplified from the standard x86 chip and connected with a very high-speed interconnection.
The all-Intel Tianhe 2 cluster at the Chinese National University of Defense Technology uses a mixture of Xeon processors and Xeon Phi accelerators, which come on large PCI Express cards that resemble a GPU card. All told, Tianhe-2 has a Linpack score of 33.86 petaFLOPS.
It's not alone. This is the first list issued since Xeon Phi hit the market, and already there are 11 systems on the Top 500 list use the it. Nvidia GPUs power 39 machines, a number that continues to grow with each list as the GPU maker aggressively pursues that market, while AMD's Radeon is in three machines.
The Xeon Phi line will continue its momentum with the new Xeon Phi 7100, the highest performing card of the lineup, according to Rajeeb Hazra, vice president if the Datacenter & Connected Systems Group and general manger of the Technical Computing Group.
The 7100 is for maximum performance, the Xeon Phi 5100, currently on the market, is a high density form factor and the Xeon Phi 3100, also on the market now, will be for price/performance-sensitive customers.
The 7100 and 3100 are PCI Express cards but the 5100 lets OEMs add their own innovation. "So the card is an enabling form factor along with OEMs already building this in," said Hazra. "Where are we headed? The rate of innovation has to continue. We believe the fundamental element that addresses many of the challenges of that innovation continuing around power efficiency lie rooted in the ability to integrate key features into silicon."
Intel has already done that with the new generation of desktop processors, Haswell. Those chips have features, like voltage regulators, on the CPU instead of on the motherboard.
Hazra said Intel wants to bring supercomputing to the large scale data-centric world, from health care to understanding social behavior of communities to other functions that require supercomputing or significant levels of computing on data. Intel is taking part in a European project to simulate the human brain, a 1 billion Euros project to do sequencing, imaging and microscopy to identify mathematical relationships that enable the brain to acquire and processes information.
Clearly Intel has Big Data on its mind. It introduced an optimized version of Hadoop a few months back that utilized Xeon instructions, and last week announced the Enterprise Edition for Lustre software that it says will simplify the implementation and management of the performance-based storage applications, including Hadoop.
When combined with the Intel's Distribution for Apache Hadoop software, it allows Hadoop users to access Lustre data files directly, reducing time and resources. Intel Enterprise Edition for Lustre is expected to ship in the third quarter of this year.
In addition to the Top 500 boasting, Hazra announced Intel's next generation of Xeon Phi, codenamed Knight's Landing. It will be a 14nm either stand-alone CPU or PCI Express co-processor. With the new chip, Intel wants to change how it's programmed.
"Offloading acceleration is an unnatural way for programmers to think of their code. What we will be doing in the next generation is making it operate like the CPU. It will come with integrated on-package memory that can also function as a stand-alone, bootable CPU," he said. A release date for Knight's Ferry was not given, but Intel is expected to move to 14nm some time next year. It's already begun the process of preparing the factories to build 14nm parts.