The supercomputer, deployed at ORNL facilities in Tennessee, is capable of processing 20,000 trillion calculations per second. The supercomputer is more than 10 times faster than its predecessor called Jaguar, which was deployed in 2009 and was considered the world's fastest supercomputer in June 2010 until it was dethroned a few months later by a Chinese supercomputer called Tianhe-1A at the National Supercomputer Center in Tianjin.
Titan's deployment comes just a few weeks ahead of the release of the Top500, which lists the 500 fastest supercomputers in the world. The fastest supercomputer in the Top500 list released in June this year was Sequoia, an IBM BlueGene/Q system deployed by the U.S. DOE at the Lawrence Livermore National Laboratory in Livermore, California.
"American competitiveness is very important from a global security and national security perspective," said Jeffrey Nichols, associate laboratory director for the computing and computational sciences directorate at ORNL, in an interview. "It's absolutely important that we are competitive in this high-tech field so the science solutions we are solving are competitive and put us on the leading edge of where we need to be in solving these problems."
Countries such as Japan and China are swiftly scaling their computing capabilities with supercomputers in the top five, Nichols said. But the U.S. makes more effective use of available computing power to solve some of the top science problems in the country, Nichols said.
"If you look at Oak Ridge and what we bring to the table is that we have application developers that can use these machines at scale. China cannot. They have an economic development model that says we'll put this hardware on the floor and people can come in and pay for using the machine in order to do their scientific research," Nichols said.
ORNL welcomes proposals from scientists and 40 projects are selected every year to use computing facilities in the lab. The proposals are selected on merit by scientific experts, but also by ensuring the applications are scalable so resources are not wasted. Scientists using Titan don't have to pay for usage.
More computing power boosts knowledge discovery, and helps in more realistic simulation and experiments, Nichols said, adding that Titan will help the U.S. in research areas like biosciences, climate, energy and space.
For example, ORNL conducts research on neutron sciences, which includes the building of combustion engines that deliver lower emissions and higher efficiency. Those experiments are related to the country's economy, environment and national security, and faster supercomputers like Titan are advancing research quicker in that area, Nichols said.
Titan is a Cray XK7 supercomputer, which pairs 18,688 Advanced Micro Devices 16-core Opteron 6274 CPUs with 18,688 Nvidia Tesla K20 GPUs (graphics processing units). Graphics processors provide faster execution of some scientific and math applications, while CPUs are better for serial processing. By harnessing the joint computing power of CPUs and GPUs, supercomputers are able to provide results in the most power-efficient way, said Steve Scott, chief technology officer of the Tesla product line at Nvidia.
Titan is built into 200 server cabinets, which is the same size as Jaguar. ORNL upgraded by moving to 16-core CPUs and the latest graphics processors, which are faster and more power efficient. Titan has 700TB of memory.
Titan consumes about 9 megawatts of power and the energy costs for running the supercomputer could add up to $10 million a year, Nichols said. The DOE is willing to absorb the cost because it knows that a sophisticated research program is needed, Nichols said.
The next milestone for supercomputers is to reach exaflop performance, which is about 1,000 petaflops, by 2018. ORNL upgrades the supercomputer at a three to four year clip, and Nichols expects a major upgrade to Titan in 2016. He also hopes an exaflop system will be in place at ORNL by 2020, though he added nothing was certain.
"We have to think about making the case for the 2020 exascale machine or 2016 machine. We have to start talking to vendors about those machines," Nichols said.
