SVM Modeling. Related work has shown that SVM is a problem suitable for hardware. In particular, in [48] Xxxxxxx, Xxxxxxx, et al. achieve 20x speedup with the use of a Virtex 5 FPGA, compared to a 2.2 GHz CPU processor. Furthermore, Papadonikolakis and Bouganis in [49] utilize an Altera Stratix III FPGA to reach 7x speedup compared to other hardware-based implementations. Additionally, in [49] Xxxx-Xxxxxxx et al. use a Virtex II to implement the SVM method, but do not achieve speed up compared to a 550 MHz CPU processor. The aforementioned results indicate that SVM can be accelerated with the use of hardware for the purposes of the QualiMaster project. In order to perform SVM Modeling we used Version 3.20 of the C/C++ open source LIBSVM project. Modeling a system entails a functional (commonly referred to as behavioral) prototype with such considerations as data operation precision, functional units, sequence of operations, etc. In this context the model is a reference design done in software (typically with MATLAB or C/C++, but it can also be in Java or Python) which gives the designer a feeling for the cost vs. performance tradeoffs. To illustrate, Papadonikolakis and Bouganis‟ work on SVM [49] uses fixed point precision rather than floating point precision because fixed point takes fewer resources vs. floating point. The evaluation, however, of the quality of the results was performed prior to the development of an architecture and a detailed hardware design – if the quality was poor there would be no need to proceed with the time-consuming design. Similarly, if one needs a hardware system with the same accuracy as the software (and we assume that this is a realistic scenario) the system operation has to be modeled from the beginning, in the case of our example (and the QualiMaster work on the same algorithm) with floating point operations. Based on the bibliography, the majority of works on SVM compare their performance to the results of LIBSVM, both in terms of accuracy and speed([42], [43], [44], [45]). During the modeling process first we analyzed the data inputs and outputs of this implementation. Then we used specific data sets to perform profiling of the software code, and finally we identified important data structures and operations. A very useful additional result of modeling is not only that it allows for comparisons against pure software implementations of an algorithm, but it also provides detailed datasets and expected results for the actual hardware de...
