20 Petaflops Titan Supercomputer Summit
In August, the Oak Ridge Leadership Computing Facility hosted a summit to discuss progress towards 20 quadrillion floating point operations per second. Currently the top mainframe in the world runs at around 8 petaflops, so this newer one would be over twice as quick as that record holder. The machine is expected to be inaugurated in the year 2012. This ultra-fast number cruncher appears to be a sure bet, although IBM had to recently cancel plans for constructing a different supercomputer. Titan will use a Gemini XE interconnect, a Nvidia Tesla X2090 processor running at 665 gigaflops, an AMD Opteron 6200 Interlagos 16 core CPU and a Tesla X090 memory with 6 gigabytes GDDR5 capacity. Recent developments may enable Titan to become the fastest device in the world. That’s assuming China or Japan don’t have something new up their sleeves to challenge it. See the file “introduction and timeline“ (PDF) for an overview on the project.
Titan could be used to better comprehend the universe. A lot of data is coming in from mapping a number of galaxies. Large models will be needed to make sense of all the information that is being collected. In 1969 an early simulation of this type was run on a mainframe that could perform at about megaflop/s and had 32 kilobytes of random access memory. Since then, there have been many orders of magnitude improvement. Large survey volumes of the sky may need 100 terabytes to a petabyte of memory (see computing the dark universe). The enormous parallelism could also be utilized to create supernova explosions in silico.
Better weather modeling is also being explored as a method to exploit the tremendous calculating power. The goal of researchers is to create a 2 kilometer resolution global forecast model in 2013. This requires a high degree of computational efficiency. In forecast mode it needs to be able to run it at fifty times real time. This is done so they can predict things in advance with enough accuracy. Climate simulations only have to be a thousand times real time. They would be used to discover longer trends such as global warming (see a unified model for forecasting climate and weather).
Titan can also be employed to reveal biochemical processes. It is a difficult task to model large molecules over a long period. Even emulating a fast conformational change of a single protein for a micro to millisecond takes many days. So they need to overcome a billion fold impasse in order to better understand certain biological processes. Specialized software and the 20 petaflops or higher can help a great deal in this area (see statistically motivated linear scaling for protein dynamics). A lot of the biotechnology revolution has ended in failure. The US has poured close to a trillion dollars in funding since 2000, but new pharmaceutical approvals have hardly budged. A lot of novel therapies have never come to fruition. The complexities of biological functions mean that simple cures for diseases will not be found. Machine intelligence and speedy computers may be the only hope of getting approval for real breakthroughs.
The rest of the presentations cover a variety of software and hardware technical challenges that are needed to take advantage of these future resources. The titan summit website can be found here.
