Hardware support for variable precision floating point formats exploration

International audienceVariable Precision (VP) Floating Point (FP) is a solution to compensate accumulation and rounding errors during computing. Hardware approaches to VP FP are often preferred, since software solutions result in algorithmic performance degradation and computational instability. We propose three new VP formats and evaluate them using a RISC-V platform running on a FPGA. Preliminary results show that our VP formats are beneficial for certain applications with low-precision requirements