Zarr Extensions

Appearance

bfloat16 data type ​

Defines the bfloat16 floating-point data type (https://en.wikipedia.org/wiki/Bfloat16_floating-point_format).

A bfloat16 number is a IEEE 754 binary32 floating-point number truncated at 16-bits.

  • 1 sign bit
  • 8 exponent bits, with bias of 127
  • 7 mantissa bits
  • IEEE 754-compliant, with NaN and +/-inf.
  • Subnormal numbers when biased exponent is 0.

Data type name ​

The data type is specified as "bfloat16".

Configuration ​

None.

Fill value representation ​

The fill value is specified in the same way as the core IEEE 754 floating point numbers: https://zarr-specs.readthedocs.io/en/latest/v3/core/v3.0.html#fill-value

The constant "NaN" corresponds to a representation of "0x7fc0".

Codec compatibility ​

bytes ​

Encoded as a 2-byte little-endian or big-endian value 0bSEEEEEEEEMMMMMMM.

See also ​

A Python implementation is available at https://pypi.org/project/ml-dtypes/

Current maintainers ​

  • Jeremy Maitin-Shepard (@jbms), Google