Arithmetic performane of floating point formarts available in VLSI

This paper reports on the results obtained by simulating 1024, 256, and 128 point complex-to-complex Fast Fourier Transforms with four different floating point formats: the IEEE P754 Rev.10 standard, the MIL-STD-1750A standard, the TRW 22-bit DSP format, and the Hitachi 16/20-bit format. Since many existing systems use a sixteen-bit format, performance of a sixteen bit format (with block floating point where appropriate) was used as a baseline for comparison. The results were evaluated for bias and variance (in the form of signal-to-noise ratio) for all functions. The performance of FIR and IIR filters was analyzed for the same numerical representations.