Parallel processing on a supercomputer is the dominant approach to data-intensive science, where large data pipelines are demanded by increasingly complex and multi-scale models and by the deluge of data coming from sophisticated instruments and nets of sensors.
New approaches to these challenges are emerging, however, and IIC collaborators have been at the forefront of experiments with lower-cost, efficient solutions to both processing and storage problems in science. In 2008, the SciGPU collaboration was launched with a $2 million grant from the National Science Foundation.
This project recognizes that most modern desktop computers contain a parallel processing device of almost unprecedented power, with hundreds of processors linked to high-bandwidth memory banks. However, unless used for playing games, the graphics processing unit (GPU) in a modern machine is little used.
General-purpose programming of GPUs began in 2006, when NVIDIA released their Compute Unified Device Architecture (CUDA). This provided a very small set of extensions to the C language, enabling scientists to harness graphics cards for their own general-purpose computations. A core technical activity of the IIC, the SciGPU project draws on challenges in neuroscience (Connectome), quantum chemistry and astronomy (MWA) and includes development of code libraries, experiments with cluster configurations and desktop supercomputers, student projects and community-building to generate and disseminate knowledge about this new paradigm.
Harvard’s leadership in this area and in GPGPU education have been recognized by NVIDIA, which designated the university as a CUDA Center of Excellence in early 2009.
