At 31 petaflops of sustained LINPACK capacity, the new Chinese Tianhe-2 supercomputer will be the fastest supercomputer in the world when this month’s Top 500 list comes out, as we reported previously in HPCwire. While the system uses American-made chips, the Chinese supercomputer’s interconnects and software will hail from a homegrown IT industry that has made remarkable progress.
The new 16,000-node TH-2 supercomputer will be a monster of an HPC system. A total of 32,000 Ivy Bridge sockets and 48,000 Xeon Phi boards will give it a total of 3.1 million cores. With 88 GB of memory per node, that gives the full 125-rack system a total of 1.404 petabytes of memory, to go along with 12.4 petabytes of shared storage. It will consume 24 MW of electricity at full capacity, and use a liquid cooling system.
TH-2 is being built by the Chinese company Inspur in tandem with the National University of Defense Technology. Inspur has manufactured some of the front-end chips, while NUDT has worked on interconnect design. The TH-2’s interconnect, called TH Express-2, is an “optoelectronics hybrid transport technology” that utilizes a fat tree topology, and is capable of 2.56 Tbps per device, according to Dr. Jack Dongarra, who is a member of the Oak Ridge National Lab and a founder of the Top500, and who met with Chinese researchers last month and subsequently published details on TH-2.
The new supercomputer’s software will also be made by a Chinese supplier, according to Peter Beckman, a computer scientist at the U.S. Department of Energy’s Argonne National Laboratory and head of the Department of Energy’s exascale initiative.
“They designed the interconnect from scratch, and they designed a software stack,” Beckman says in an interview with Computerworld. “They are taking their own approach on how to do parallelism. The two items that make the supercomputer super, the software and the interconnect, they are growing at home.”
Beckman says it’s a point of national pride for the Chinese to build their own IT industry that is capable of delivering all the components of an HPC system–from CPUs all the way to interconnects and memory. “They will really work, in my opinion, to make a top machine that will be all [homegrown] tech from top to bottom, the software, the interconnect and the CPU,” he says.
As the HPC community progresses toward exascale systems, the Chinese are marking to their own drum. “Inside the messaging layers, there were pieces that they were inventing, that they were doing over–doing a different way,” Beckman says in the Computerworld piece. “My impression is that their intent is obviously to collaborate and work with the community, but they really want to grow many of the components in-house.”