This year’s supercomputing conference, which is better-known as SC13 and takes in place in Denver this week, began with the announcement of the latest Top500 list of the world’s fastest computers. Not much has changed at the top since the list, which is updated twice a year, was last compiled.
China still holds the top with Tianhe-2 at the National Supercomputer Center in Guangzho, which has sustained performance of 33.86 petaflops (quadrillions of calculations per second). In fact, the top 5 systems remained the same and there was only one new addition to the top 10, Piz Daint, a Cray XC30 system at the Swiss National Supercomputing Centre in Lugano. Piz Daint also happens to be the most efficient among the top 10.
IBM and HP supply the bulk of these systems, followed by Cray. As in the PC world, Intel dominates high-performance computing with more than 82 percent of the systems on the list using Intel Xeon server processors. The rest use AMD Opteron (9 percent), IBM Power (8 percent) or SPARC.
The biggest change in recent years has been the adoption of accelerators to boost performance on certain high-throughput, parallel workloads while keeping the power manageable. Of the top 500 systems, 53 now use accelerators—mostly Nvidia's Tesla, though a couple have AMD FirePro and Intel’s Xeon Phi has been coming on strong since it was first announced at the International Supercomputing Conference in mid-2012.
Earlier today Nvidia announced the Tesla K40, its fastest board to date for general-purpose GPU computing. In comparison to the K20X (which is based on the same Kepler chip), the K40 has more stream processors running at slightly faster speeds and twice as much memory (12GB) pushing performance to 4.3 teraflops single-precision arithmetic and 1.4 teraflops on double-precision math. Nvidia also announced a deal with IBM to combine Tesla GPUs with Power-based systems used for data analytics. I would not be surprised to hear announcements from AMD or Intel this week as well.
But the rapid rise of accelerators may slow down thanks to some changes in the works. For two decades, the Top500 group has used a single test, the Linpack benchmark which measures the performance solving a dense system of linear equations, to rank the world’s fastest computers. Over time this led organizations to develop systems--most recently ones using accelerators such as GPUs—designed to post the highest possible scores on Linpack, even though this did not reflect the real-world performance on many high-performance applications.
In June, Jack Dongarra, a computer science professor at the University of Tennessee who developed Linpack and compiles the Top500 list, announced a new benchmark called the High Performance Conjugate Gradient (HPCG), designed to better reflect the way many supercomputing applications fetch data and work on it. It will take some time to phase-in--and HPCG won’t replace Linpack--but it should provide a better picture of the real-world performance of the world’s top computers.