Supercomputing, There's an App For That 66
aarondubrow writes "Researchers at MIT have created an experimental system for smart phones that allows engineers to leverage the power of supercomputers for instant computation and analysis. The team performed a series of expensive high-fidelity simulations on the Ranger supercomputer to generate a small "reduced model" which was transferred to a Google Android smart phone. They were then able to solve engineering and fluid flow problems on the phone and visualize the results interactively. The project proved the potential for reduced order methods to perform real-time and reliable simulations for complicated problems on handheld devices."
Re:Not even... (Score:3, Interesting)
What the phone is doing is "reduced order modeling", which (if the article is using the term accurately) finding a simple set of equations whose solution provably approximates a far more complex system of equations. It's not just interpolation (lookup tables, or machine learning from a training ensemble). Reduced order models actually have dynamics in them.
The potential innovation here, as I see it, is that it takes some supercomputing effort to build the reduced order model for a specific problem. So you can input a description of your problem into the phone, communicate it to the supercomputer and let it analyze the problem for a couple hours, and it spits out a custom approximation to your problem that you can run in the field on your phone.
Now, what I'm not sure that they've developed is some convenient way to specify the problem using the phone's interface. Maybe the problems they're working with are too complex to be easily specified with a few keystrokes. If all they're doing is loading a precomputed reduced model onto the phone, and there's no interaction with the supercomputer to let it handcraft solutions to problems in semi-realtime, then it's not nearly as interesting.
Re:Put CUDA on a phone, then we can talk (Score:3, Interesting)
The lattice QCD people, at least, are porting their code to CUDA just as fast as they can. The bottleneck right now is that there's no good way to get multiple GPU's to communicate (quickly). So, for the largest problems (simulating a 64x64x64x192 lattice), you still need a conventional supercomputer (like Ranger, the one in the article here), because it's just too huge to put on a single GPU and multi-GPU doesn't scale well.
But for smaller problems (like a 24x24x24x64 lattice), GPU's will be great, and people are developing this capability as fast as they can.