An application-oriented analysis of power/precision trade-off in fixed and floating-point arithmetic units for VLSI processors

Floating and fixed-point formats represent the most used representation in VLSI DSP systems. Fixed-point formats are often preferred for lower power and faster operation, when the algorithm permits the appropriate scaling factors. Floating-point formats require additional hardware and power, but guarantee better precision performance. This work compares the energy dissipated by the key arithmetic units in DSP systems, addition and multiplication, in both floatingand fixed-point scenarios, also considering reduced precision and power-aware architecture solutions. The evaluation has been performed referring to STMicroelectronics 180nm,130nm and 90nm CMOS processes. The results show that specialized floating-point formats can be the best energy-efficient solution in low power DSP.