What is the bfloat16 floating-point format, and what is its primary purpose?

The bfloat16 (brain floating point) format is a computer number format that uses 16 bits of memory. Its primary purpose is to accelerate machine learning and near-sensor computing by providing a wide dynamic range of numeric values, similar to 32-bit floating-point numbers, but with reduced precision and storage requirements.

Who developed the bfloat16 format, and why is it named 'brain floating point'?

The bfloat16 format was developed by Google Brain, an artificial intelligence research group at Google. The name 'brain floating point' originates from this group where the concept for the format was conceived.

How does the bfloat16 format compare in size to the standard IEEE 754 single-precision format?

The bfloat16 format occupies 16 bits in computer memory, making it a shortened version of the 32-bit IEEE 754 single-precision floating-point format (binary32).

What are the key structural differences between bfloat16 and IEEE 754 single-precision (binary32) in terms of bits allocated?

While both formats use 1 bit for the sign, bfloat16 allocates 8 bits for the exponent and 7 bits for the significand (with an implicit leading bit, totaling 8 bits of precision). In contrast, IEEE 754 single-precision (binary32) uses 8 bits for the exponent and 23 bits for the significand (totaling 24 bits of precision).

What is the main advantage of using bfloat16 for machine learning algorithms?

Bfloat16 is used to reduce the storage requirements and increase the calculation speed of machine learning algorithms. It achieves this by using fewer bits per number while retaining a similar dynamic range to 32-bit floats.

Which types of hardware processors and AI accelerators support the bfloat16 format?

Bfloat16 is supported by various hardware, including Intel Xeon processors (with AVX-512 BF16 extensions), Intel Data Center GPUs, Intel Nervana NNP-L1000, Intel FPGAs, AMD Zen microarchitecture (including AMD Instinct), NVIDIA GPUs, Google Cloud TPUs, AWS Inferentia, AWS Trainium, ARMv8.6-A architecture, and Apple's M2 and later A-series chips.

What popular software libraries and frameworks support the bfloat16 format?

Several widely used libraries and frameworks support bfloat16, including CUDA, Intel oneAPI Math Kernel Library, AMD ROCm, AMD Optimizing CPU Libraries, PyTorch, and TensorFlow.

How does the exponent encoding in bfloat16 work, and what is its bias?

The bfloat16 exponent is encoded using an offset-binary representation, similar to the IEEE 754 standard. The exponent bias is 127 (or 7F in hexadecimal). To find the true exponent, 127 must be subtracted from the value stored in the exponent field.

What are the minimum and maximum values for the exponent field in bfloat16, and how are they interpreted?

The minimum exponent field value is 00 (hexadecimal), which represents subnormal numbers or zero. The maximum exponent field value is FF (hexadecimal), which is used to represent infinity and NaN (Not a Number) values. Values from 01 to FE (hexadecimal) represent normalized numbers.

How are positive and negative infinity represented in the bfloat16 format?

Positive and negative infinity in bfloat16 are represented with their respective sign bits (0 for positive, 1 for negative), all 8 exponent bits set to 1 (FF hexadecimal), and all significand bits set to zero.

How are NaN (Not a Number) values represented in bfloat16?

NaN values in bfloat16 are represented with either sign bit, all 8 exponent bits set to 1 (FF hexadecimal), and the significand bits being non-zero. At least one bit in the significand must be set to 1 for it to be a NaN.

What is the primary trade-off bfloat16 makes regarding range and precision compared to IEEE 754 single-precision?

Bfloat16 is designed to maintain the approximate number range of the 32-bit IEEE 754 single-precision format while significantly reducing the precision. It sacrifices the detailed precision of binary32 for a comparable dynamic range in a smaller, 16-bit format.

What is the approximate maximum positive finite value that can be represented in bfloat16?

The maximum positive finite value representable in bfloat16 is approximately 3.38953139 x 10^38. This is slightly less than the maximum finite value of single-precision (binary32).

What is the minimum normalized positive value that can be represented in bfloat16, and how does it compare to single-precision?

The minimum normalized positive value in bfloat16 is approximately 1.175494351 x 10^-38 (which is 2^-126). This value is the same as the minimum normalized positive value for IEEE 754 single-precision floating-point numbers.

How does the conversion from IEEE 754 binary32 to bfloat16 typically occur, especially regarding rounding?

Initially, the conversion from IEEE 754 binary32 to bfloat16 often involved truncation, which is equivalent to rounding towards zero. However, depending on the hardware platform and its intended use (like matrix multiplication units), other rounding mechanisms such as round-to-nearest-even or non-IEEE Round-to-Odd modes are also employed.

How is the conversion from bfloat16 to IEEE 754 binary32 handled?

Since the 32-bit binary32 format can represent all exact values present in the bfloat16 format, the conversion from bfloat16 to binary32 is straightforward. It involves padding the significand bits with 16 zeros to match the binary32 format.

What are the bit allocations for the sign, exponent, and fraction in the standard IEEE 754 half-precision (binary16) format?

The IEEE 754 half-precision (binary16) format uses 1 bit for the sign, 5 bits for the exponent, and 10 bits for the significand (fraction).

How does Nvidia's TensorFloat-32 (TF32) format differ from bfloat16 in terms of bit allocation?

Nvidia's TensorFloat-32 (TF32) format uses 1 bit for the sign, 8 bits for the exponent, and 10 bits for the fraction (significand). This gives it the same exponent range as bfloat16 but a larger significand than bfloat16, though still less than standard single-precision.

Describe the structure of ATI's fp24 format in terms of bits for sign, exponent, and fraction.

ATI's fp24 format allocates 1 bit for the sign, 7 bits for the exponent, and 16 bits for the fraction (significand).

What is the bit structure of Pixar's PXR24 format?

Pixar's PXR24 format consists of 1 bit for the sign, 8 bits for the exponent, and 15 bits for the fraction (significand).

What are the hexadecimal and binary representations for the value 1 in bfloat16?

The value 1 in bfloat16 is represented as 3f80 in hexadecimal and as 0_01111111_0000000 in binary, where the parts are sign, exponent, and significand respectively.

What are the hexadecimal and binary representations for the value -2 in bfloat16?

The value -2 in bfloat16 is represented as c000 in hexadecimal and as 1_10000000_0000000 in binary, indicating a negative sign, a specific exponent, and a zero significand.

What are the hexadecimal and binary representations for positive and negative infinity in bfloat16?

Positive infinity is represented as 7f80 in hexadecimal (0_11111111_0000000 in binary), and negative infinity is represented as ff80 in hexadecimal (1_11111111_0000000 in binary). Both use all exponent bits set and zero significand bits.

What are the hexadecimal and binary representations for zero and negative zero in bfloat16?

Zero is represented as 0000 in hexadecimal (0_00000000_0000000 in binary), and negative zero is represented as 8000 in hexadecimal (1_00000000_0000000 in binary). Both have all exponent and significand bits as zero.

What are the hexadecimal and binary representations for an approximate value of Pi (π) in bfloat16?

An approximate value of Pi (π) in bfloat16 is represented as 4049 in hexadecimal, which translates to 0_10000000_1001001 in binary.

What are the hexadecimal and binary representations for an approximate value of 1/3 in bfloat16?

An approximate value of 1/3 in bfloat16 is represented as 3eab in hexadecimal, which translates to 0_01111101_0101011 in binary.

How are quiet NaN (qNaN) and signaling NaN (sNaN) represented in bfloat16?

Both qNaN and sNaN use all 8 exponent bits set to 1 (FF hexadecimal) and have a non-zero significand. A qNaN is represented as ffc1 (0_11111111_1000001 in binary), and an sNaN is represented as ff81 (0_11111111_0000001 in binary), with the specific significand bits determining the exact NaN value.

What is the purpose of the implicit leading bit in the bfloat16 significand?

The implicit leading bit, which is assumed to be 1 for normalized numbers, is a common convention in floating-point formats like bfloat16. It allows for an extra bit of precision to be represented without consuming an additional storage bit, thus enhancing the effective precision of the significand.

Are bfloat16 numbers suitable for integer calculations, according to the source text?

No, the source text indicates that bfloat16 numbers are unsuitable for integer calculations, but clarifies that this is not their intended use case.

What is the minimum positive subnormal value representable in bfloat16?

The minimum positive subnormal value in bfloat16 is 2^-133, which is approximately 9.2 x 10^-41. This value is derived from the minimum exponent (2^-126) multiplied by the smallest possible significand (2^-7).

What is the significance of preserving the exponent bits from binary32 when converting to bfloat16?

Preserving the 8 exponent bits from binary32 in bfloat16 is crucial because it allows bfloat16 to maintain the same approximate dynamic range as the 32-bit format. This wide range is particularly beneficial for machine learning tasks where values can span many orders of magnitude.

What is the role of bfloat16 in mixed-precision arithmetic on supported platforms?

On platforms that support it, bfloat16 can be used in mixed-precision arithmetic. This means that bfloat16 numbers can be operated on directly and then potentially expanded to wider data types (like 32-bit floats) for further calculations, optimizing performance and memory usage.

How does the precision of bfloat16 compare to IEEE 754 single-precision?

Bfloat16 offers significantly less precision than IEEE 754 single-precision. Bfloat16 has 8 bits of significand precision (7 stored + 1 implicit), whereas single-precision (binary32) has 24 bits of significand precision (23 stored + 1 implicit).

What is the relationship between bfloat16 and the IEEE 754 standard?

Bfloat16 is described as a shortened version of the IEEE 754 single-precision (binary32) format. While it shares the concept of exponent bias and special value representations (like infinity and NaN) with IEEE 754, its significand precision and some rounding behaviors differ.

What does the 'brain floating point' name imply about the design philosophy of bfloat16?

The name 'brain floating point' suggests that the format was designed with biological neural networks or artificial neural networks in mind, prioritizing the wide dynamic range needed for complex computations in these fields over the high precision typically found in traditional floating-point formats.

What are the specific rounding modes mentioned for conversions involving bfloat16?

The text mentions several rounding modes: round toward 0 (truncation), round-to-nearest-even (used by Google TPU and NVIDIA), and a non-IEEE Round-to-Odd mode (used by ARM).

What is the purpose of the 'See also' section in relation to bfloat16?

The 'See also' section provides links to related topics and formats, such as the IEEE 754 half-precision format, Google Brain, and primitive data types, helping readers understand bfloat16 within a broader context of computer arithmetic and data representation.

How does the exponent bias of bfloat16 compare to that of IEEE 754 single-precision?

Both bfloat16 and IEEE 754 single-precision (binary32) use an exponent bias of 127 (7F hexadecimal). This means that the stored exponent value is the actual exponent plus 127.

What is the significance of the 'Not to be confused with binary16' note at the beginning of the article?

This note clarifies that bfloat16 is distinct from binary16, which is another name for the IEEE 754 half-precision floating-point format. This distinction is important because they have different bit allocations and characteristics.

What is the maximum finite positive value representable in IEEE 754 single-precision, and how does it compare to bfloat16's maximum?

The maximum finite positive value representable in IEEE 754 single-precision (binary32) is approximately 3.402823466 x 10^38. This is slightly higher than the maximum finite positive value of bfloat16, which is approximately 3.38953139 x 10^38.

What does the term 'near-sensor computing' imply in the context of bfloat16 usage?

Near-sensor computing suggests that bfloat16 is used in devices or processors located close to data sources (sensors). This implies a need for efficient, low-power computation directly on the data as it's generated, where bfloat16's speed and reduced memory footprint are advantageous.

What is the purpose of the table comparing different floating-point formats (IEEE half-precision, TensorFloat-32, ATI fp24, Pixar PXR24, IEEE single-precision)?

The tables visually compare the bit structure (sign, exponent, fraction) of bfloat16 against several other floating-point formats. This allows for a direct comparison of their respective sizes and allocations, highlighting how bfloat16 differs from formats like IEEE half-precision and standard single-precision.

What is the role of the 'implicit leading bit' in the significand of bfloat16?

The implicit leading bit, assumed to be '1' for normalized numbers, effectively increases the precision of the significand without using an extra bit. For bfloat16, this means the 7 stored fraction bits, combined with the implicit leading bit, provide a total of 8 bits of precision for the significand.

How does the conversion from binary32 to bfloat16 handle the significand bits?

When converting from binary32 to bfloat16, the significand field of the binary32 number is reduced. This reduction can be achieved through truncation (rounding towards zero) or other specified rounding mechanisms, effectively discarding the lower-order bits of the binary32 significand.

What is the significance of bfloat16's 8-bit exponent width?

The 8-bit exponent width in bfloat16 is significant because it matches the exponent width of the IEEE 754 single-precision (binary32) format. This is the key factor that allows bfloat16 to preserve the wide dynamic range of binary32 numbers.

What does the term 'transprecision' refer to in the context of floating-point formats like bfloat16?

Transprecision refers to the ability of a system or format to handle numbers with varying levels of precision. Bfloat16 is an example of a reduced-precision format used alongside higher-precision formats (like 32-bit floats) in mixed-precision computations, allowing developers to choose the appropriate precision for different parts of a calculation to optimize performance and accuracy.

What is the purpose of the 'Exponent encoding' section in the article?

The 'Exponent encoding' section details how the exponent values are represented within the bfloat16 format. It explains the use of an offset-binary representation with a bias of 127 and clarifies how different exponent field values (00, 01-FE, FF) correspond to special values like zero, subnormals, normalized numbers, infinity, and NaN.

How does the number of bits in the significand of bfloat16 compare to IEEE half-precision (binary16)?

Bfloat16 has 7 stored significand bits (plus an implicit leading bit for a total of 8 bits of precision), whereas IEEE half-precision (binary16) has 10 stored significand bits (plus an implicit leading bit for a total of 11 bits of precision).

What is the implication of bfloat16 having fewer significand bits compared to IEEE 754 single-precision?

Having fewer significand bits (7 stored vs. 23 stored) means bfloat16 has lower numerical precision than single-precision. This can lead to greater rounding errors in calculations where high precision is critical, but it is a deliberate trade-off for speed and memory efficiency in applications like machine learning.

What does the reference to 'Google Brain' in the context of bfloat16 development signify?

The mention of 'Google Brain' signifies that the development of bfloat16 originated from Google's dedicated artificial intelligence research division, highlighting its specific focus on advancing AI hardware and software capabilities.

What is the primary benefit of using bfloat16 for machine learning training and inference?

The primary benefit is the acceleration of machine learning tasks. By using a 16-bit format with a wide dynamic range, bfloat16 allows for faster computations and reduced memory bandwidth requirements compared to 32-bit floating-point formats, which is crucial for large-scale AI models.

Bfloat16 Floating-point Format Wiki: Bfloat16: Precision Meets Performance

Dive in with Flashcard Learning!

When you are ready...
🎮 Play the Wiki2Web Clarity Challenge Game🎮

Overview

A Specialized Format

The bfloat16 (brain floating point) format is a 16-bit computer number format designed to accelerate machine learning and near-sensor computing. It achieves this by using a floating radix point, similar to standard floating-point formats, but with a specific trade-off between range and precision.

Accelerating AI

Developed by Google Brain, bfloat16 preserves the approximate dynamic range of the 32-bit IEEE 754 single-precision format by retaining its 8 exponent bits. This allows it to handle the vast numerical scales common in AI computations more effectively than formats with less range.

Range vs. Precision

While bfloat16 offers the extensive dynamic range of 32-bit floats, it supports only 8 bits of precision (compared to 24 bits in binary32). This reduction makes it unsuitable for precise integer calculations but ideal for the iterative nature of neural network training where maintaining range is often more critical than extreme precision.

Bitwise Structure

Component Breakdown

The bfloat16 format is structured as follows:

Sign bit: 1 bit
Exponent width: 8 bits
Significand precision: 8 bits (7 explicitly stored, plus an implicit leading bit)

This structure is derived from the IEEE 754 single-precision format, retaining the exponent bits but reducing the significand bits.

Comparative Layouts

The following table illustrates the bit layout of bfloat16 compared to other relevant floating-point formats:

IEEE half-precision (binary16)

	0	exponent (5 bit)	fraction (10 bit)
	15	14	10			0

bfloat16

	0	exponent (8 bit)			fraction (7 bit)
	15	14			7	0

Nvidia's TensorFloat-32 (TF32)

	0	exponent (8 bit)			fraction (10 bit)
	18	17			10		0

ATI's fp24 format

	0	exponent (7 bit)		fraction (16 bit)
	23	22		15				0

IEEE 754 single-precision (binary32)

	0	exponent (8 bit)			fraction (23 bit)
	31	30			22				0

Exponent Encoding

Offset Binary Representation

The bfloat16 exponent is encoded using an offset-binary representation, with an exponent bias of 127 (0x7F). This is consistent with the IEEE 754 standard for single-precision floats, ensuring compatibility and maintaining the same range.

E_min: 0x01 - 0x7F = -126
E_max: 0xFE - 0x7F = 127
Exponent Bias: 0x7F = 127

The true exponent is calculated by subtracting the bias from the stored exponent value. Special values (00_H and FF_H) are reserved for specific meanings.

Special Exponent Values

The minimum and maximum values of the exponent field (00_H and FF_H) are interpreted specially:

Exponent Field	Significand Zero	Significand Non-Zero	Equation
00_H	Zero (±0)	Subnormal numbers	(−1)^signbit × 2⁻¹²⁶ × 0.significandbits
01_H to FE_H	Normalized value		(−1)^signbit × 2^{exponentbits−127} × 1.significandbits
FF_H	±Infinity	NaN (Quiet, Signaling)

The minimum positive normalized value is approximately 1.18 × 10⁻³⁸, while the minimum positive subnormal value is approximately 9.2 × 10⁻⁴¹.

Conversion & Rounding

Binary32 to Bfloat16

When converting from IEEE 754 binary32 (32-bit float) to bfloat16, the exponent bits are preserved. The significand field is reduced by truncation (rounding toward zero) or other mechanisms. This preserves the range but sacrifices precision.

Different hardware platforms employ various rounding schemes:

Google TPU: Round-to-nearest-even.
ARM: Non-IEEE Round-to-Odd.
NVIDIA: Round-to-nearest-even.

Bfloat16 to Binary32

Converting from bfloat16 back to binary32 is straightforward because binary32 can represent all exact bfloat16 values. The conversion involves padding the significand bits with zeros, effectively extending the precision without loss.

This seamless conversion facilitates its use in mixed-precision arithmetic, where computations might occur in bfloat16 but intermediate or final results might be expanded to wider types for accuracy.

Special Values

Infinity

Positive and negative infinity are represented similarly to IEEE 754 standards:

val    s_exponent_signcnd
+inf = 0_11111111_0000000
-inf = 1_11111111_0000000

This involves setting all 8 exponent bits to 1 (FF_hex) and the significand bits to zero, with the sign bit determining positive or negative infinity.

Not a Number (NaN)

NaN values are encoded with all 8 exponent bits set (FF_hex) and at least one significand bit set to 1. The sign bit can be either 0 or 1.

val    s_exponent_signcnd
+NaN = 0_11111111_klmnopq
-NaN = 1_11111111_klmnopq

(where at least one of k, l, m, n, o, p, or q is 1). Signaling NaNs (sNaNs) and quiet NaNs (qNaNs) can be represented, though signaling NaNs have limited practical use.

Range and Precision

Maintaining Dynamic Range

The primary design goal of bfloat16 was to maintain the dynamic range of IEEE 754 single-precision (binary32), which extends up to approximately 3.4 × 10³⁸. This is achieved by using the same 8-bit exponent field.

The maximum positive finite value representable in bfloat16 is approximately 3.38953139 × 10³⁸.

Reduced Precision

To fit within 16 bits while retaining the 8-bit exponent, the significand precision is reduced to 7 explicit bits (plus the implicit leading bit), totaling 8 bits. This results in a precision of roughly two to three decimal digits.

This trade-off is acceptable for many machine learning tasks where the convergence of algorithms is more sensitive to the range of values than to minute precision differences.

Illustrative Examples

Key Values

Here are some examples of bfloat16 representations in hexadecimal and binary:

3f80 = 0 01111111 0000000 = 1
c000 = 1 10000000 0000000 = -2

7f7f = 0 11111110 1111111 ≈ 3.38953139 × 10^38 (Max finite positive value)
0080 = 0 00000001 0000000 ≈ 1.175494351 × 10^-38 (Min normalized positive value)

Special and Fractional Values

Examples of zeros, infinities, and common mathematical constants:

0000 = 0 00000000 0000000 = +0
8000 = 1 00000000 0000000 = -0
7f80 = 0 11111111 0000000 = +Infinity
ff80 = 1 11111111 0000000 = -Infinity

4049 = 0 10000000 1001001 ≈ π (pi)
3eab = 0 01111101 0101011 ≈ 1/3

ffc1 = x 11111111 1000001 => qNaN (Quiet NaN)
ff81 = x 11111111 0000001 => sNaN (Signaling NaN)

Related Concepts

Data Types

Half-precision (binary16): IEEE 754 16-bit float with 1-bit sign, 5-bit exponent, and 11-bit significand.
Single-precision (binary32): The standard 32-bit float with 1-bit sign, 8-bit exponent, and 23-bit significand.
TensorFloat-32 (TF32): Nvidia's format using 8 exponent bits and 10 significand bits, offering a compromise between bfloat16 and FP16.
Arbitrary-precision arithmetic: For calculations requiring precision beyond standard floating-point types.

Origins and Ecosystem

Google Brain: The research group that developed the bfloat16 format.
Tensor Processing Unit (TPU): Google's custom ASICs for machine learning, which heavily utilize bfloat16.
Mixed-precision arithmetic: Using different numerical formats (like bfloat16 and FP32) within the same computation to optimize performance and memory usage.
Computer number formats: General overview of how numbers are represented in computers.

Teacher's Corner

Edit and Print this course in the Wiki2Web Teacher Studio

Edit and Print Materials from this study in the wiki2web studio

Click here to open the "Bfloat16 Floating-point Format" Wiki2Web Studio curriculum kit

Use the free Wiki2web Studio to generate printable flashcards, worksheets, exams, and export your materials as a web page or an interactive game.

True or False?

Test Your Knowledge!

Gamer's Corner

Are you ready for the Wiki2Web Clarity Challenge?

Learn about bfloat16_floating-point_format while playing the wiki2web Clarity Challenge game.

Unlock the mystery image and prove your knowledge by earning trophies. This simple game is addictively fun and is a great way to learn!

Play now

Explore More Topics

Discover other topics to study!

bourgeoisie

capital loss

duchy of brze ść kujawski

ordre national du mérite

References

A full list of references for this article are available at the Bfloat16 floating-point format Wikipedia page

Feedback & Support

To report an issue with this page, or to find out ways to support the mission, please click here.

Disclaimer

Important Notice

This page was generated by an Artificial Intelligence and is intended for informational and educational purposes only. The content is based on a snapshot of publicly available data from Wikipedia and may not be entirely accurate, complete, or up-to-date.

This is not professional technical advice. The information provided on this website is not a substitute for consulting official hardware documentation, software libraries, or seeking advice from qualified computer architects, engineers, or AI researchers. Always refer to official specifications and consult with experts for specific implementation needs.

The creators of this page are not responsible for any errors or omissions, or for any actions taken based on the information provided herein.

Bfloat16: Precision Meets Performance

💡 Dive in with Flashcard Learning! 💡

Overview ℹ️

🔢 A Specialized Format

🚀 Accelerating AI

⚖️ Range vs. Precision

Bitwise Structure 💻

🧩 Component Breakdown

📊 Comparative Layouts

Exponent Encoding ⚡

⚙️ Offset Binary Representation

🧮 Special Exponent Values

Conversion & Rounding 🔄

↔️ Binary32 to Bfloat16

🧩 Bfloat16 to Binary32

Special Values ✨

♾️ Infinity

❓ Not a Number (NaN)

Range and Precision 📏

📈 Maintaining Dynamic Range

📉 Reduced Precision

Illustrative Examples 💡

🔢 Key Values

🧮 Special and Fractional Values

Related Concepts 🔗

📚 Data Types

🏢 Origins and Ecosystem

Teacher's Corner 🧑‍🏫

Edit and Print this course in the Wiki2Web Teacher Studio

True or False? 🤔

❓ Test Your Knowledge! ❓

Gamer's Corner 🎮

Are you ready for the Wiki2Web Clarity Challenge?

Explore More Topics

📜 Discover other topics to study!

References

📜 References

Feedback & Support 👍

To report an issue with this page, or to find out ways to support the mission, please click here.

Disclaimer ⚠️

📜 Important Notice

Dive in with Flashcard Learning!

Overview

A Specialized Format

Accelerating AI

Range vs. Precision

Bitwise Structure

Component Breakdown

Comparative Layouts

Exponent Encoding

Offset Binary Representation

Special Exponent Values

Conversion & Rounding

Binary32 to Bfloat16

Bfloat16 to Binary32

Special Values

Infinity

Not a Number (NaN)

Range and Precision

Maintaining Dynamic Range

Reduced Precision

Illustrative Examples

Key Values

Special and Fractional Values

Related Concepts

Data Types

Origins and Ecosystem

Teacher's Corner

True or False?

Test Your Knowledge!

Gamer's Corner

Discover other topics to study!

References

Feedback & Support

Disclaimer

Important Notice