Slashdot Mirror
Simulating Supernovae with Graphics Cards
astroboy writes "As graphics cards get more powerful, Los Alamos and Utah scientists have developed a package, Scout, to use those usually-languishing FLOPs to do simulations, and to visualize of them on the on the run. As an example, they have released
movie of part of the evolution of a core-collapse supernovae"
I know this is slashdot, and I appreciate all the Beowulf cluster jokes, especially since they're actually appropriate here, but nobody is asking any meaningful questions. By my calculations, the noise-to-signal ratio is illegal div_by_zero.
Where can I get Scout? What is the license? What platforms are supported? I'm working on an open-source scientific computing package for doing quantum simulations, and I'd like to use Scout for visualization, but this article provides no information on where to get Scout or even if the licensing would allow me to use it.
It's also not clear exactly how you'd link Scout up with an existing app. Does Scout produce machine code that you stick into your app somehow? Are there C or C++ wrappers for using Scout?
IEEE 754 compliance makes fp operations slower, which is why hardware doesn't often support it (famous example Cray where SQRT(1-COS(X)) could return with an error root of a negative number).
Roundoff errors might not matter for graphics (who cares about being one pixel off?), but it is a huge problem for numerical computations.
Also, does GPU signal overflow/underflow/division by zero??
I did RFTA, and they don't actually run the simulation on the graphics card. On the contrary, they had to downsample the data from 320^3 to 256^3 just to fit it into the GPU's memory. All they did in the GPU was a bit of post-processing (and the rendering, which looks nice enough).
In a more general sense, I wouldn't "trust" the result of a hydro-only simulation of a SN explosion in any detail. Too much physics left out, and a lot of chaotic dynamics which are only barely resolved (or not at all). An experiment like this is indicative of the likely behavior of a general class of models, but it's not a prediction.