Performance Comparison of Dlfloat16 and Bfloat16 Binary Formats

The 16-bit floating-point format is a popular alternative to the widely used 32 -bit floating-point format (FP32). This study analyzes the performance of two commonly employed 16-bit versions, dlfloat16 and bfloat16, for data with Laplacian and Gaussian distribution. The applied performance metric is the signal-to-quantization noise ratio (SQNR), allowed by using the connection between the floating-point format and piecewise uniform quantization. It is shown over a wide range of input data variances that SQNR results for Laplacian and Gaussian data differ slightly, indicating that both observed formats are robust to changes in the data distribution. SQNR analysis also suggested that dlfloat16 is a better option when increased calculation precision is required, while bfloat16 is better from a dynamic range perspective.