IT & Network Infrastructure : Eight Ways Nvidia, AMD and Other GPUs Boost HPC
Eight Ways Nvidia, AMD and Other GPUs Boost HPC
Coprocessing is Here
New operating systems such as Apple's "Snow Leopard" Mac OS X with OpenCL and Microsoft's Windows 7 with DirectCompute interfaces are being designed to utilize GPUs for tasks that go beyond graphics. Coprocessing leverages the sequential architecture of CPUs and the massively parallel architecture of GPUs to enable splitting of an application's workload between the two.
A GPU in a laptop can do face recognition or video processing 10 times faster than the CPU in the same laptop. Such capabilities are important as consumers rely more on media than text to tell their stories. A GPU's parallel architecture is well-suited for many of those tasks, such as video and photo editing. Apple will begin using more of AMD's ATI Radeon HG 4870 video cards in its Mac Pro systems. The iMac systems also will offer AMD's HD 4850 cards.
GPUs can hold many more processing cores than can CPUs. For example, the latest Nvidia Tesla GPUs contain 240 cores and are poised to double every 12 to 18 months, far outpacing CPUs and Moore's Law.
Low Power Consumption
The modern GPU offers more performance per watt of any processing architecture, which means it can be deployed within a much smaller power footprint than rival architectures. The Dell Precision T7500 workstation comes with Nvidia Tesla GPUs, creating what officials from both vendors call a "personal supercomputer."
Write Once, Run Everywhere
The same GPU computing program can run from netbooks to laptops to supercomputing clusters. Cell phones are next. Appro's HyperPower HPC compute cluster combines Intel's "Nehalem EP" processors with Telsa GPUs from Nvidia.
GPUs are good at accelerating mainstream applications, including video encoding, fast media conversion, video editing, DVD upscaling, real-time video enhancement and photo touch-ups, and soon will include database acceleration and business intelligence. HP's Z800 workstation holds up to two Nvidia Tesla GPUs.
From Games to Science
Many math-intensive applications can leverage the GPU to dramatically accelerate calculations: scientific research, seismic exploration, financial trading, virus modeling, etc. The Tsubame supercomputer at the Tokyo Institute of Technology, the 41st fastest supercomputer in the world, according to Top500.org, uses GT200 GPUs from Nvidia.
AMD in 2008 introduced the ATI Radeon HD 4890 graphics card, which includes the RV790 chip. The chip, shown here on a wafer, is the size of a dime and has broken the 1 teraflop (1 trillion floating point operations per second) barrier.