Innovative arithmetics for efficient DNN computing: HW and SW solutions and their integration in RISC-V platforms

This thesis provides a comprehensive analysis of the various features of real number representations, ranging from the IEEE floating point standard to fresh and innovative alternatives. We focus on the Posit\texttrademark format, in particular, commenting on its primary pros and limitations as well as its essential qualities. We demonstrate novel advances in posit forms using optimised non-linear function implementations. We describe our implementation of a posit library, complete with a high-level application programming interface and interaction with popular machine learning frameworks like Tensorflow. On this point, we also presented our findings on posit accuracy when used in various deep neural network tasks. On the hardware side, we show our RISC-V core-integrated lightweight posit processing unit that enables data compression between posits and IEEE floats. Finally, we show a pipelined full posit processing unit that allows algebraic operations between posits.