← Back to catalogue
Research draft

floating point

vr.tr.floating-point · XCT.QLT

Let an agent explain floating-point representation and its formats, relay the IEEE 754 standard, rounding and error behaviour from computing references, describe reduced-precision formats used in machine learning, and distinguish floating point from fixed point, integers, decimal arithmetic and arbitrary precision.

Thing Registry Cross-cutting context

Research draft, second pass

A second pass drafted this model: the structure a model of this thing needs, and what is known about it in the world. The line under this one says how the second half was obtained - researched against sources, or recalled without web access, in which case nothing here was read anywhere and every claim is a lead to verify. Unreviewed either way.

written by Claude from model knowledge without web access - no source was read, every claim is a lead to verify

Researched by: Claude

Purpose and description

Let an agent explain floating-point representation and its formats, relay the IEEE 754 standard, rounding and error behaviour from computing references, describe reduced-precision formats used in machine learning, and distinguish floating point from fixed point, integers, decimal arithmetic and arbitrary precision.

A method of representing real numbers in computers with a significand, a base and an exponent so that the radix point can float, standardised by IEEE 754 in single-precision 32-bit and double-precision 64-bit binary formats and in decimal formats including binary integer decimal encoding, with smaller formats such as minifloats and FP6 used in machine learning and block floating-point sharing one exponent across several values; floating point trades exact representation for range and is the basis of most scientific and graphics computing.

What it is for: Representing real numbers in computing.

It can be explain representation and formats; relay the standard and errors; describe reduced-precision formats; distinguish related representations.

Distinguishing features

Floating radix point

Standardised formats

Rounding error

Special values

What it looks like

Not a visible object; a binary encoding.

Physical character

IEEE 754 first published: 1985 year - revised 2008 and 2019

single precision: 32 bits - 24-bit significand

double precision: 64 bits - 53-bit significand

decimal digits, double: about 15-17 count

How it is recognised

Sign, exponent and significand encoding

Single precision, double precision, minifloat, FP6, block floating-point, binary integer decimal

Fixed point has a fixed radix; integers have no fraction; arbitrary precision grows as needed

Related models

is a kind of - in registry terms

computer numbering format

is a kind of - in registry terms

number data type

is standardised by - the floating-point standard

IEEE 754

is contrasted with - with a fixed radix

fixed-point arithmetic

In practice

Families and kinds

binary formats: half, single, double and quadruple precision

decimal formats with binary integer decimal and densely packed decimal encodings

reduced-precision formats such as bfloat16, FP8, FP6 and minifloats

block floating-point

extended and non-IEEE formats

Standards and regulation

IEEE 754-2019

ISO/IEC 60559

Language standards specifying floating-point behaviour

Failure modes and hazards

Rounding and cancellation errors

Comparing floats for exact equality

Overflow, underflow and NaN propagation

Also called

double-precision floating-point formatsingle-precision floating-point formatFP6block floating-pointminifloatbinary integer decimaldecimal128 floating-point formatdecimal32 floating-point formatdecimal64 floating-point formatdecimal floating pointextended precisionnormal floating-point numberNVFP4

Where this came from

wikidata · CC0 1.0

Drafted structure

Bundle to layer to finding to question, as the second pass will find it: 4 bundles · 8 layers · 8 findings · 16 questions.

Understand What floating point is.

Science.

Definition

Definition and encoding.

Definition

Definition.

  1. What is floating-point representation, and how do sign, exponent and significand encode a number? definition
  2. Is the question about the concept, a format, numerical error or a programming language? boundary

Formats

Formats.

Formats

Formats.

  1. What are single and double precision, decimal formats, minifloats, FP6 and block floating-point? definition
  2. Which entry fits the specific format? action
Standard The standard.

Regulation.

IEEE 754

IEEE 754.

IEEE 754

IEEE 754.

  1. What does IEEE 754 specify about formats, rounding modes, exceptions and special values? provenance
  2. Which references are standard? provenance

History

History.

History

History.

  1. How did floating point develop from early machines to the standard, with Kahan s role attributed? provenance
  2. Which sources are cited? provenance
Use Numerical practice.

Practice.

Errors

Errors and pitfalls.

Errors

Errors.

  1. How do rounding, cancellation and accumulation errors arise, and how are they managed? action
  2. Which entry fits numerical analysis? action

Machine learning

Reduced precision.

Machine learning

Machine learning.

  1. Why do machine learning systems use bfloat16, FP8, FP6 and block floating-point? provenance
  2. Which entry fits mixed-precision training? action
Context Hardware and alternatives.

Context.

Hardware

Hardware.

Hardware

Hardware.

  1. How do processors and GPUs implement floating-point units? provenance
  2. Which entry fits floating-point unit? action

Alternatives

Alternatives.

Alternatives

Alternatives.

  1. When are fixed point, decimal, rational or arbitrary precision arithmetic preferred? provenance
  2. Which entry fits arbitrary-precision arithmetic? action

What the second pass must settle

  • Should IEEE 754, double precision and bfloat16 be separate primary entries?
  • How should computing references be linked?
  • The registry entry has merged aliases naming specific formats; should they be split off?