float8_e5m2fnuz data type

Defines an 8-bit floating point representation with:

• 1 sign bit
• 5 exponent bits, with bias of 16
• 2 mantissa bits
• Extended range: no infinity, NaN represented by 0b1000'0000.
• Subnormal numbers when biased exponent is 0.

The suffix fnuz is consistent with LLVM/MLIR naming and is derived from the differences to IEEE floating point conventions. F is for "finite" (no infinities), N for with special NaN encoding, UZ for unsigned zero.

Data type name

The data type is specified as "float8_e5m2fnuz".

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 "0x80".
• The constants "Infinity" and "-Infinity" are not supported since there is no representation of infinity.

Codec compatibility

bytes

Encoded as a 1-byte value 0bSEEEEEMM. The "endian" parameter has no effect.

See also

• Bit layout described by https://arxiv.org/abs/2206.02915
• Python implementation available at https://pypi.org/project/ml-dtypes/
• Implemented in LLVM/MLIR as Float8E5M2FNUZ.

Current maintainers

• Jeremy Maitin-Shepard (@jbms), Google