"Performance Evolution of Different SYCL Implementations on the Basis of PLSSVM"

We developed PLSSVM, a GPU accelerated Parallel Least Squares Support Vector Machine, able to classify dense data sets with hundreds of thousand data points and more than thousand features beating the state-of-the-art SMO implementations like LIBSVM. Additionally, we support a plethora of different hardware architectures like any CPU and GPUs from Intel, NVIDIA, and AMD using different backends written in OpenMP, CUDA, HIP, OpenCL, and SYCL. In this talk, we want to compare these backends on different architectures with respect to their implementation and performance characteristics.