• NetCDF-3 uses a canonical format approach in which data stored on disk is represented in a standard way (big-endian)
• Writing and reading requires conversion between standard on-disk representation and native (in memory) representation
• This big-endian bias favors platforms that are big-endian or bi-endian set to big-endian, such as ARM, PowerPC, DEC Alpha, SPARC V9, MIPS, PA-RISC, and IA64. It always requires byte-swapping conversions when writing or reading for little-endian processors such as Intel x86.
• NetCDF-4 uses "reader-makes-right" approach, which eliminates big-endian bias.
• With "reader-makes-right" approach
  • Writer always uses native representations, so no conversion is necessary on writing
  • Reader is responsible for detecting what representation is used and applying a conversion, if necessary, to reader's native representation
  • No conversion is necessary if reader and writer use same representation
• NetCDF-4 also lets writer control endianness explicitly, if necessary.

2007 Unidata NetCDF Workshop for Developers and Data Providers > Introduction to NetCDF-4