Non-linear Approximation Intellectual Property Targeting Edge Inferencing in Field Programmable Gate Array

Non-linear function is the core computation for every neural network's activation function. Many researchers attempted different methods to implement non-linear functions onto hardware to achieve better system performance. Despite its popularity, non-linear functions' implementation methods available today are still costly for edge inferencing. This paper proposes an approximation Intellectual Property (IP) to implement non-linear functions in FPGA with high accuracy, fast computation, controllable error rate, adjustable granularity, and yet light in resource utilization. Test result shows that the proposed approximation IP has significant reduction in computation time (improvement rate up to 71%), DSP utilization (improvement rate up to 97%) as well as number of logic utilization (improvement rate up to 74%) and total register count (improvement rate up to 69%).