Core Concepts
Takum arithmetic introduces a novel logarithmic number format that synthesizes the advantages of posits for low-bit applications with high encoding efficiency for numbers distant from unity, while addressing several issues previously identified in posits.
Abstract
The paper introduces a new number format called "takum arithmetic" that aims to address the limitations of IEEE 754 floating-point numbers and posits.
Key highlights:
Takum encoding: Takums use a variable-length characteristic (exponent) representation with a fixed 3-bit regime and a logarithmic significand. This allows for efficient encoding of numbers close to unity while maintaining a constrained dynamic range.
Logarithmic significand: Takums employ a logarithmic significand, which simplifies arithmetic operations like multiplication, division, and square root compared to linear significands. This makes takums well-suited for applications like neural networks.
Dynamic range: The dynamic range of takums is constrained to √e^-255 to √e^255, which the authors argue is suitable for general-purpose computational tasks without excessive magnitude.
Rounding: Takums use saturation arithmetic for rounding, clamping values outside the dynamic range to the smallest or largest representable number.
Advantages over IEEE 754 and posits: Takums exhibit higher coding efficiency for numbers distant from unity compared to posits, while avoiding the excessive dynamic range and redundant bit representations present in IEEE 754 formats.
The paper provides a formal definition of the takum encoding, proves key properties, and compares takums to existing number formats.
Stats
The dynamic range of takums is constrained to √e^-255 ≈ 4.2 × 10^-56 to √e^255 ≈ 2.4 × 10^55.
Quotes
"Takums exhibit an asymptotically constant dynamic range in terms of bit string length, which is delineated in the paper to be suitable for a general-purpose number format."
"It is demonstrated that takums either match or surpass existing alternatives."