Nvidia today unveiled a system for high-performance computing that uses four graphics processors to provide 1 teraflop of computing power, and multiple units can be easily combined to form a GPU-based computing cluster. The system competes with CPU-based clusters that employ Intel or AMD chips, but offers faster performance on some tasks while using less energy, Nvidia says. The system uses about 800 watts, meaning a 1-petaflop computer composed of GPUs would consume about half the power of Jaguar, the No. 2 fastest supercomputer on the planet.
Nvidia has signed deals with several HPC system manufactures, such as Appro, to get its Tesla S1070 into the hands of researchers and scientists. As Nvidia has focused on scientific computing, it has made GPU-based cards to be used in HPC systems, for desktop supercomputers and even for some corporate settings. This is its first GPU-system that comes pre-assembled, so companies or researchers don’t have to build the GPU clusters themselves. Its graphics processors, which excel at parallel processing for CAD design, creating software algorithms or Monte Carlo simulations, are used in several industries. Today’s announcement also underscores an emerging trend of using a broader range of processors in HPC and even data centers as the tradeoff between performance and energy becomes a bigger consideration.