Interconnect pushed China super to #1
SAN JOSE, Calif. -- A homegrown, ultrafast interconnect chip set was part of the secret sauce behind the Tianhe-1a, the first China-built system to be named the world's fastest supercomputer. The 160 Gbit/second Galaxy interconnect links thousands of the latest Intel Westmere and Nvidia Fermi processors in the system that will be used for a broad range of scientific research.
The Tainhe-1a in Tianjin was measured at 2.5 petaflops using the Linpack benchmark. The previous top supercomputer was the U.S.-based Jaguar built by Cray using six-core AMD processors to deliver 1.7 petaflops.
The accomplishment is a major milestone for China's advancing electronics industry. The Galaxy interconnect is twice as fast as the Infiniband QDR interconnects used on many of today's fastest supercomputers based on chips from Mellanox Technologies.
U.S. researchers said the news underscores the need for greater federal spending in the U.S. on research in high-performance systems. "It's a sign they are making a long term investment," said Jack Dongarra, University of Tennessee researcher who helps compile the twice annual Top 500 list of supercomputers.
"The next stage would be replacing the U.S. processors—and they are executing on chips replacing the Intel and Nvidia parts," said Dongarra, referring to China's Godson processors. "Then the machine becomes more interesting and of greater concern" to U.S. researchers, he said.
A China-built supercomputer in Shenzhen hit number two on the Top 500 supercomputer list published in June using Intel and Nvidia processors. Researchers on that project said their next-generation system will use China's Godson chips.
The U.S. has "not made adequate investment across the board on the key components of the supercomputer ecosystem," he said.
Inside the Tianhe-1a
The Tianhe-1a is believed to be fie first supercomputer to use China's Galaxy interconnect chip set. Dongarra said he expect the chip set could appear in other China supers in the future.
Apart from its speed, the interconnect appears to be similar to others in use. It is based on a standard fat tree architecture, Dongarra said.
Indeed the Tianhe-1a is itself not usual among supercomputers. It uses an increasingly popular hybrid approach of pairing standard microprocessors with graphics co-processors.
The system replaces an earlier supercomputer in Tianjin built from Intel processors and AMD graphics chips that was ranked #7 on the June Top 500 list. Graphics chips are credited with providing heavy duty processing for relatively little power.
As is typical of new supercomputers, the Tianhe-1a uses the latest processors to gain a speed edge. They include a 2.93 GHz version of the six-core Intel Westmere processor and the Nvidia Fermi M2050.
The Tianhe-1a is built up from 7,168 nodes. Each node contains two Westmere processors, one Fermi graphics card and 32 Gbytes memory.
The system came online in October and will be used for weather and climate modeling, materials research, biomedical research and oil exploration. Dongarra got an up close look at the system while in Tianjin last week for a joint conference of U.S. and China supercomputer researchers sponsored by the U.S. National Science Foundation and the China Academy of Science.
The meeting "was intended to start a dialog between [the U.S. and China supercomputer] communities, and hopefully we can continue the discussion," Dongarra said.
The Tianhe-1a is likely to retain its #1 status when the new Top 500 list comes out in early November. However, the list changes quickly.
For example, Cray and IBM are at work separately on next-generation systems expected to deliver tens of petaflops under a multi-year program sponsored by the U.S. Defense Advanced Research Projects Agency. The goals of that program include designing more balanced systems that are easier to program.
Dongarra was quick to point out the limitations of measuring supercomputer performance in petaflops using the Linpack benchmark.
"Linpack should not be used as the only way to measure the characteristics of a computer, it's just one computation," Dongarra said. "You need to understand other aspects to get a feel for how any application runs on a supercomputer, so it's just one pole in the ground," he said.
"I didn't get offered nor did I have time to look [at other aspects of the Tianhe-1a], but it looked like a reasonably balanced computer for scientific computing," he added.