high performance computing on graphics processing units: hgpu.org

In the world of artificial intelligence (AI) at the edge, we need to focus primarily on the energy efficiency with which we approach deep neural network (DNN) applications. In many applications, the speed of obtaining an inference can be critical; but many applications easily meet their time requirements, and the energy needed to calculate the huge numbers of multiplication and addition operations of DNNs becomes the essential element. We have provided systolic architectural solutions written in C++ and OpenCL that are highly flexible and easily tunable to take full advantage of the resources of programmable devices and achieve superior energy efficiencies. We focused on low-cost solutions with soft macro microprocessors (Nios2) and hard macro microprocessors (ARM cortex A9).