Research Menu

Skip Search Box

The Next Wave | Vol. 20 | No. 1 | 2013

The Oak Ridge Leadership Computing Facility is home to Titan, the world's most powerful supercomputer for open science (as of November 2012) with a theoretical peak performance exceeding 20 petaflops. That kind of computational capability—almost unimaginable—is on par with each of the world's 7 billion people being able to carry out 3 million calculations per second. (Image courtesy of Oak Ridge National Laboratory.)

Supercomputers are extremely powerful, fast computers that are used for large-scale scientific calculations such as those found in quantum physics, weather forecasting, climate research, gas exploration, molecular modeling, and physical simulations (e.g., aircrafts and nuclear weapons). Twice a year since June of 1993, TOP500 has published a list that ranks the 500 most powerful commercially available supercomputers with help from high-performance computer experts, computational scientists, manufacturers, and the Internet community. TOP500 measures a supercomputer's performance based on its ability to solve a dense system of linear equations using floating-point arithmetic (i.e., the LINPACK benchmark).

In June of 1993, CM-5 was the number one supercomputer on the first ever TOP500 list. CM-5 performed 59.7 gigaflops on the LINPACK benchmark—that's approximately 59,700 billion floating-point operations (i.e., calculations) per second. Fast forward 20 years to November of 2012, and the number one supercomputer on the TOP500 list—Titan—performs 17,590 trillion calculations per second (i.e., 17.59 petaflops). Over the past 20 years, supercomputers have increased in performance at a rate of about 879,497 billion calculations per second per year. The systems listed in table 1 have occupied the number one position in the TOP500 list over that time.


Name Country Site Manufacturer Date in No. 1 Position
CM-5 US Los Alamos National Laboratory Thinking Machines Corporation 06/1993
Numerical Wind Tunnel Japan National Aerospace Laboratory of Japan Fujitsu 11/1993, 11/1994, 06/1995, 11/1995
Intel XP/S 140 Paragon US Sandia National Laboratory Intel 06/1994
Hitachi SR2201 Japan University of Tokyo Hitachi 06/1996
CP-PACS Japan Center for Computational Science, University of Tsukuba Hitachi 11/1996
ASCI Red US Sandia National Laboratory Intel 06/1997, 11/1997, 06/1998, 11/1998, 06/1999, 11/1999, 06/2000
ASCI White US Lawrence Livermore National Laboratory IBM 11/2000, 06/2001, 11/2001
The Earth Simulator Japan Earth Simulator Center NEC 06/2002, 11/2002, 06/2003, 11/2003, 06/2004
BlueGene/L US Lawrence Livermore National Laboratory IBM 11/2004, 06/2005, 11/2005, 06/2006,11/2006, 06/2007
Roadrunner US Los Alamos National Laboratory IBM 06/2008, 11/2008, 06/2009
Jaguar US Oak Ridge National Laboratory Cray, Inc. 11/2009, 06/2010
Tianhe-1A China National Supercomputing Center in Tianjin National University of Defense Technology 11/2010
K Computer Japan RIKEN Advanced Institute for Computational Science Fujitsu 06/2011, 11/2011
Sequoia US Lawrence Livermore National Laboratory IBM 06/2012
Titan US Oak Ridge National Laboratory Cray, Inc. 11/2012

FIGURE 1. Over the past 20 years, supercomputer performance has increased from gigaflops in 1993 to teraflops in 1997 and then to petaflops in 2008. TOP500 projects that by 2018, the highest performing supercomputer will reach about 1 exaflops (i.e., quintillions of floating-point operations per second; not shown in figure).

The • red data points show the sum LINPACK performance of all 500 supercomputers on the TOP500 list, the • purple data points show the LINPACK performance of the top supercomputer (i.e., number one on the list) and the • blue data points show the LINPACK performance of the bottom supercomputer (i.e., number 500 on the list).

FIGURE 2. Since 2004, cluster computing has been the dominant computing architecture of supercomputers in the TOP500 list.

The number of processors and their configuration determine how a computer reads and carries out program instructions. A single processor allows a computer to carry out one instruction at a time. A multiprocessor allows a computer to carry out two or more instructions simultaneously.

In symmetric multiprocessing (SMP), two or more identical processors are connected to a single shared main memory and controlled by a single operating system. In this kind of architecture, a computer system can execute multiple instructions simultaneously while drawing upon shared resources, which is useful for processes such as online transactions. In massively parallel processing (MPP), two or more identical processors are each connected to a separate memory and are each controlled by a separate but identical operating system. An interconnect arrangement of data paths allows messages to be sent between processors. In an MPP architecture, the workload is essentially distributed across separate computers that communicate with one another so, for example, a number of databases can be searched in parallel.

Single instruction, multiple data (SIMD) processing is a form of parallel processing that lets one microinstruction operate at the same time on multiple data items, which is useful for processes involving multimedia applications. Cluster computing is another form of parallel processing that uses multiple separate computers, each having an SMP architecture, to form what appears to users as a single system. A cluster computing architecture is useful in handling traffic on high-traffic websites. Constellation computing is a cluster of symmetric multiprocessors.

FIGURE 3. Since 2004, Intel Corporation has been the dominant manufacturer of chips in the supercomputers that made it into the TOP500 list.

FIGURE 4. Supercomputers on the TOP500 list are used primarily in industry, research, and academia. Over 50% of them go to industry.

*All figures are from TOP500 and have been modified for online viewing.

View PDF version of this article (300 KB)


Date Posted: May 17, 2013 | Last Modified: May 17, 2013 | Last Reviewed: May 17, 2013