Humankind is a curious creature. Collectively, we seek answers to questions to questions thought unknowable in another age. Advances are being made on many scientific fronts, especially through the contributions of everyday people through distributed computing projects like Folding@Home, World Community Grid, SETI@Home and many others.
Anyone who has ever attempted to set up a home distributed computing farm or designed a state-of-the-art supercomputer knows, the two main limiting factors are cost of the machines and power they consume. GPU-based computing appears to be coming into its own.
According to an nVidia press release, National Taiwan University [NTU] is carrying out this work on the first GPU-based supercomputer in Taiwan, the 128-GPU cluster at CQSE, which uses 16 nVidia Tesla S1070 1U systems [64 Tesla C1060 processors]. The system plays a key role in large-scale computations for quantum physics, ranging from the strong interaction at the subatomic scale to the strongly correlated electrons in condensed matter physics, and to the cosmology at the astronomical scale.
A research team at National Taiwan University, led by Ting-Wai Chiu, Professor of Physics and Associate Director of the Center for Quantum Science and Engineering [CQSE], is "achieving breakthrough results in learning about the early evolution of the universe by harnessing nVidia Tesla parallel processors – which provide the computational horsepower of an IBM BlueGene/L supercomputer" [Not exactly, Blue Gene/L, the first Blue Gene series was twice as fast in 2001 posting 36.01 TFLOPS and that was before the supercomputer?s cabinets were doubled to 32 and speed increased to 135.5. Eventually, Blue Gene/L would top out at over 200 TFLOPS compared to NVIDIA?s current 15 TFLOPS], "at just 1% the cost and 10% the power consumption." How does that deal sound?
NTU’s system plays a key role in large-scale computations for quantum physics, ranging from the strong interaction at the subatomic scale to the strongly correlated electrons in condensed matter physics, and to the cosmology at the astronomical scale. In addition, the lattice QCD group [TWQCD] based at National Taiwan University is now the first group in the world to use a GPU cluster to perform large-scale simulations of lattice QCD with exact chiral symmetry.
"We are excited to see our GPU-based cluster outperform many conventional supercomputers in both cost and energy use," said Chiu. "With our GPU-enabled supercomputer, we are delivering 15 teraflops at a price of US $200,000, 1% the cost of a conventional supercomputer like IBM BlueGene/L."
The link between enthusiast community and university is now confirmed, as none of this would probably happen if there wasn’t for a small research team at a Belgian university. Dan Vivoli, Senior Marketing Vice President at NVIDIA, acknowledges, "If it wasn?t for FASTRA, NVIDIA would not have developed the TESLA supercomputer." I seriously want one of these machines, my F@H ranking would be enormous. [Chris, University of Antwerp launched FASTRA II, now based on GTX295 boards. Ed.]
The cost benefit analysis must have been very easy for NTU when deciding on which HPC solution [High Performance Computing] to choose. Perhaps the budget presentation went something like this: "We can purchase a more powerful IBM BlueGene/L or comparable solution for millions or we can procure a more feasible solution that suits our needs for $200,000." The trend of massive increases computational power in smaller form factors continues. And it looks like nVidia’s investment in the enterprise sector may be pay off handsomely from adoption of its own CUDA [http://en.wikipedia.org/wiki/CUDA] based development platform.
GPU Computing is looking like it will provide a lower price of entry into to the HPC across the board and hopefully, this will translate into faster resolution of real world problems – such as cancer, AIDs, global warming as well as physical questions about the universe.