universal_dtypes

Status — early development. The repository is scaffolded and the design is settled (see docs/design.md); the NumPy dtype implementation has not started. The installation and usage snippets below describe the intended API — universal_dtypes is not yet published to PyPI.

universal_dtypes is a stand-alone Python package that registers the Stillwater Universal number systems as first-class NumPy dtypes, so alternative real-number formats can be used directly in NumPy (and pandas) arrays. It is modeled on ml_dtypes, which does the same for machine-learning float formats used by JAX and TensorFlow.

The following dtypes are planned:

• posit8, posit16, posit32, posit64 — tapered-precision posits (Posit Standard, es = 2)
• fp8, fp16 — configurable low-precision floats (Universal cfloat)
• fixpnt8, fixpnt16 — fixed-point
• lns16, lns32 — logarithmic number system

Linear algebra over these types (dense/sparse solvers, and the quire-based exact accumulation policy) lives in the sister package mtl5-python, which consumes these dtypes. See Relationship to MTL5.

Installation

universal_dtypes will be published to PyPI as universal-dtypes (PEP 503 normalizes the underscore to a hyphen; pip install universal_dtypes also resolves):

pip install universal_dtypes

To run the tests:

pip install universal_dtypes[dev]
pytest

To build from source (the build fetches the Universal C++ headers via CMake FetchContent; a C++20 compiler is required):

git clone https://github.com/stillwater-sc/universal_dtypes
cd universal_dtypes
pip install -e ".[dev]"

Example usage

import numpy as np
import universal_dtypes

a = np.array([1.0, 2.0, 3.0], dtype=universal_dtypes.posit16)
b = a * 2  # element-wise arithmetic, in posit16
c = np.sum(a)  # NumPy reductions (naive posit accumulation)
n = np.sqrt(a)  # NumPy ufuncs

The dtypes also register under their string names, so NumPy's usual APIs accept them:

np.dtype("posit16")
np.zeros(4, dtype="posit16")
np.arange(8, dtype="fp8")

Specifications of implemented number formats

Posits — posit8, posit16, posit32, posit64

A posit posit<nbits, es> encodes a real number with a sign bit, a run-length regime field, up to es exponent bits, and the remaining bits as fraction. Precision is tapered: it is highest for magnitudes near ±1 and decreases toward 0 and ±∞. Posits have no subnormals and no separate infinities or NaNs — a single exceptional value NaR ("Not a Real") represents overflow, division by zero, and other undefined results. The standard sizes use es = 2 (posit8 = posit<8,2>, posit16 = posit<16,2>, and so on).

Configurable floats — fp8, fp16

Universal's cfloat is a parameterized IEEE-754-style float (sign / exponent / fraction). fp8 and fp16 are low-precision configurations for mixed-precision work. Unlike posits, they follow the familiar exponent/mantissa layout and can carry subnormals, infinities, and NaN depending on the configuration.

Fixed-point — fixpnt8, fixpnt16

Fixed-point fixpnt<nbits, rbits> is a two's-complement integer interpreted as a value scaled by 2^-rbits — a fixed absolute resolution across its whole range (in contrast to floating formats), with configurable rounding and saturation/wrap-around on overflow.

Logarithmic number system — lns16, lns32

The logarithmic number system stores a sign and the base-2 logarithm of the magnitude. Multiplication and division become addition and subtraction of the stored logarithms (cheap and exact in the exponent), at the cost of more expensive addition and subtraction.

Encodings and exceptional values

Each type is a true NumPy dtype with a fixed item size — posit8/fixpnt8/fp8 occupy one byte, posit16/fixpnt16/fp16/lns16 two bytes, posit32/lns32 four, and posit64 eight. Exceptional values differ by family and are not interchangeable with IEEE semantics:

• posits have a single NaR and no ±inf/NaN; overflow saturates toward the largest-magnitude finite posit rather than producing infinity.
• cfloat (fp8/fp16) may expose ±inf and NaN per its configuration.
• fixpnt has no non-finite values; out-of-range results saturate or wrap.
• lns carries its own encodings for zero and its exceptional value.

Casts to and from float32/float64 (and the integer types) round to the target format; converting a non-finite IEEE value into a posit yields NaR.

Quirks of low-precision arithmetic

As with any low-precision format, naive accumulation loses information. Summing many values in posit16 rounds after every partial sum:

import numpy as np
import universal_dtypes

x = np.ones(4096, dtype=universal_dtypes.posit16)
np.sum(x)  # rounds each partial sum in posit16 — not 4096 exactly

The standard fix is to accumulate in higher precision. Universal's posits also provide the quire — a fixed-point accumulator wide enough to hold an exact sum of products, rounding only once at the end. The quire is an algorithm concern (how you accumulate), so it lives in the linear-algebra layer, not in the dtype: it is exposed by mtl5-python through accumulator= on dot, norm, and the other reduction-bearing operations. universal_dtypes deliberately gives ordinary NumPy semantics for the dtype; when you need an exact reduction, reach for MTL5:

import mtl5

# exact posit16 sum-of-products via the quire (single final rounding)
mtl5.dot(x, np.ones_like(x), accumulator="quire")

The two answers will differ on ill-conditioned data — by design.

Relationship to MTL5

universal_dtypes is the element type and its scalar / element-wise arithmetic. mtl5-python is what you do with arrays of them: dense/sparse linear algebra, and the quire / accumulator mixed-precision policy. mtl5-python depends on universal_dtypes for the dtypes; universal_dtypes does not depend on MTL5. See docs/design.md for the full rationale and the package boundary.

License

universal_dtypes is licensed under the MIT License — see LICENSE. Copyright © 2026 Stillwater Supercomputing, Inc.

It builds on Universal (MIT), whose header-only number-system implementations provide the arithmetic behind each dtype.