skip to: online tools | main navigation | content | footer

Computer Sciences R&D

Home » Research & Development » Computer Sciences » Visualization & Scientific Computing » Reconfigurable Computing Using FPGAs

Reconfigurable Computing Using FPGAs

To meet the scientific community’s constant demand for increased computing power, researchers at Sandia National Laboratories are actively investigating emerging technologies that can be leveraged to improve the capabilities of high-performance computing (HPC) systems. While the traditional approach to enhancing performance has been to scale the number of compute nodes in an HPC system, an alternative method is to add hardware accelerators to individual nodes in a system.

These accelerators can perform certain scientific computing operations much more efficiently than general-purpose CPU and therefore represent opportunites to boost node performance. Multiple commercial products can be used as hardware accelerators, including the Cell (IBM/Sony/Toshiba), XMONARCH (Raytheon), and CSX (ClearSpeed) architectures. However, these technologies are still emerging and lack widespread support. A more attractive technology for accelerator research that is available today is reconfigurable hardware.

Reconfigurable hardware devices such as field-programmable gate arrays (FPGAs) can be programmed to emulate custom digital hardware circuitry. Current generation FPGAs house up to 10 million logic gates in a single chip and can be reprogrammed with new circuitry in a few milliseconds. With this in mind, reconfigurable computing (RC) researchers are using FPGAs as an affordable way to implement an algorithm in hardware instead of software and have achieved significant application speedups. In support of this work, HPC vendors are beginning to include FPGA accelerators in their system architectures.

Sandia’s RC research effort has demonstrated that FPGAs can be used as hardware accelerators for scientific-computing applications. Early results have produced speedups that are 2 to 10 times faster than approaches that rely solely on software implementations of algorithms. We expect performance will continue to improve as the tools and hardware developed in this research effort are refined and as additional applications are adapted to hardware. This work, combined with research into other emerging accelerators, will help propel the scientific and engineering communities at Sandia National Laboratories as they continue to answer the challenging questions facing our nation.