Fig. 4. Compression factors versus compression errors.
a–d, The maximum absolute and decimal errors for 24-bit linear and 16-bit logarithmic quantization (LinQuant24, LogQuant16) with 1D round + lossless and 3D Zfp compression are shown. Every marker represents, for one variable, the global maximum of the normalized absolute error (a,c) and decimal error (b,d) for 100% preserved information (a,b) and 99% preserved information (c,d). The geometric means of compression factors over all 68 variables are provided as horizontal lines. The medians of the errors across all 68 variables are given as vertical lines.