Bio-mimetic classification on modern parallel hardware: Realizations in NVidia CUDA and OpenMP
Thomas Nowotny, Mehmet K. Muezzinoglu, Ramón Huerta · Figshare · 2011
Both the brain and modern digital architectures rely on massive parallelismfor efficient solutions to demanding computational tasks, such as pattern recognition. Inthis paper, we implement a parallel classi cation scheme inspired by the insect brain intwo popular parallel computing frameworks, namely as an NVidiarCUDATMimplemen-tation on a TeslaTMdevice and a brute force OpenMPTMparallel implementation on aquad-core CPU. When evaluating the systems on the MNIST data-set of handwrittendigits, we can report that, compared with a standard serial implementation on a singleCPU core, CUDATMimplementations of the bio-inspired classi cation provide a 7-to-11fold speed-up, whereas the OpenMPTMimplementation is 2-to-4 times faster. Our re-sults are a proof of concept that suggests that modern parallel computing architectures andbio-mimetic algorithms are compatible and that the CUDATMsolution on an NVidiarTeslaTMC870 device at the time of writing has a small edge over an OpenMP solutionon a recent quad core processor (3 GHz AMDrPhenomTMII X4 940)