Comparison of OpenCL based design for a medical device on heterogeneous architectures with CPU, GPU and FPGA

Digital design is moving in the high performance era with fast steps. The ability to manufacture rich components, with large numbers of resources at relative low cost, has made design cost a considerable figure in the lifetime of modern circuits. Design cost can be reduced by reducing design time, which can be achieved with appropriate methodologies, improving designer productivity like high-level synthesis. After many years of research, such toolflows have lately reached mainstream industry, offering quality-of-results especially in the FPGA domain. While most of the times C/C++ is used as input language, OpenCL, a C/C++ extension introduced for application development in heterogeneous multi-core architectures, utilizing different CPUs, DSPs and GPUs, is gaining acceptance by major FPGA vendors. In this paper an OpenCL based environment is used for the design of the Back Projection algorithm, found in medical devices. It is shown that the fine granularity of parallel computational blocks, found in large amounts in modern FPGA devices, can be captured and exploited in a natural way with OpenCL's constructs and offer implementation improvements (5X-10X average speedup) over more complicated coding techniques. Compared to other implementation platforms, FPGAs offer performance, scalability and a low power alternative, at reduced cost.