high performance computing on graphics processing units: hgpu.org

Today graphics processing units (GPUs) are not only able to generate graphical imaging but also able to expose its multicore architecture to increase computationally heavy general purpose algorithms that can be adapted to the multicore architecture of the GPU. The study conducted in this thesis explores the efficiency of using the general purpose graphics processing unit (GPGPU) application programmer interface (API) Direct Compute for implementing the support vector machine (SVM) data mining algorithm. More specifically how does Direct Compute compare, efficiency wise, to two competing GPGPU APIs; CUDA and OpenCL. To answer the research question an artifact has been implemented that is able to run the same support vector machine algorithm, in both single and double floating-point precision, using Direct Compute, CUDA and OpenCL as well as solely using the CPU. Data has been collected using the artifact and includes; the time for training and testing, correctness of the solution when compared to the CPU implementation, in terms of number of support vectors and accuracy/enrichment factor, as well as the efficiency increase of the GPGPU implementations when compared to the CPU implementation. The results of the study show that the Direct Compute implementation has an efficiency advantage over both the OpenCL and CUDA implementations, in both single and double precision. The correctness of the solutions produced by the Direct Compute implementation are not significantly deviating from the CPU implementations solutions, in either single or double precision. It is concluded that Direct Compute is a beneficial API for implementing support vector machines, with efficiency and correctness surpassing both CUDA and OpenCL.