Dynamic Fused Multiply-Accumulate Posit Unit with Variable Exponent Size for Low-Precision DSP Applications
Nuno Neves, Pedro Tomás, Nuno Roma · 2020
The current processing power and energy efficiency demands have been pushing the research on low-precision floating-point formats alternative to the IEEE-754 standard. One such alternative is the Posit format, which has demonstrated to provide improved accuracy and larger dynamic ranges with lower precision arithmetic (when compared to the IEEE-754 standard). Such advantages are achieved by adopting a dynamic encoding of the value's scale factor, which combines a variable-sized regime field with a parameterized exponent field. Current hardware implementations are based on fixed exponent sizes. However, this limits the range of values that can be represented by the format for a given precision. Alternatively, this paper proposes a new posit Dynamic Fused Multiply-Accumulate (DFMA) unit with support for variable exponent sizes, capable of encoding an extended representation range, from values with high decimal precisions to very large integer numbers, within the same hardware. When compared to other posit implementations, the proposed DFMA unit shows that the overheads imposed by the improved representation range are marginal, making it suitable to a wide range of application domains.