Low-Power Scientific Computing

Introduction: Scientists and mathematicians are increasingly realizing the computational benefits of using modern, multi-core architectures. In response to this, manufacturers of traditional desktop graphics-processing units (GPUs) have evolved their architectures to create desktop and server GPGPUs (General Purpose Graphics Processing Units). These GPGPUs are quickly becoming the platform of choice for many high-performance, highly parallel applications. GPGPUs are also commodity hardware products commonly available in many desktop and laptop computers, making them rather inexpensive. The tools to program them are easily available as well; Nvidia’s Compute Unified Device Architecture (CUDA) package, for example, provides a small set of extensions to the C programming language, allowing for straightforward implementation of parallel algorithms on GPGPUs. Individual cores in Intel’s up-and-coming Larrabee processor implement the ubiquitous x86 ISA, allowing users to use a host of already-existing development tools to port their applications to it. Server products like the Nvidia Tesla S1070 with even more compute power are also available. Several applications, from a wide variety of domains, including medical imaging, electronic design automation, physics simulations, and stock pricing models, observe remarkable speed-ups on GPUs – at times, over 300X. Based on these dramatic performance increases, GPGPUs seem like an ideal computing substrate for high-performance, scientific computing. However, there are two major problems with GPGPUs – power consumption and an unbalanced ratio of compute ability to memory bandwidth.