# bfloat16 floating-point format

The **bfloat16 (Brain Floating Point) floating-point format** is a computer number format occupying 16 bits in computer memory; it represents a wide dynamic range of numeric values by using a floating radix point. This format is a truncated (16-bit) version of the 32-bit IEEE 754 single-precision floating-point format (binary32) with the intent of accelerating machine learning and near-sensor computing.[1] It preserves the approximate dynamic range of 32-bit floating-point numbers by retaining 8 exponent bits, but supports only an 8-bit precision rather than the 24-bit significand of the binary32 format. More so than single-precision 32-bit floating-point numbers, bfloat16 numbers are unsuitable for integer calculations, but this is not their intended use.

Floating-point formats |
---|

IEEE 754 |

Other |

The bfloat16 format is utilized in upcoming Intel AI processors, such as Nervana NNP-L1000, Xeon processors (AVX-512 BF16 extensions), and Intel FPGAs,[2][3][4] Google Cloud TPUs,[5][6][7] and TensorFlow.[7][8] ARMv8.6-A also supports the bfloat16 format.[9] As of October 2019, AMD has added support for the format to its ROCm libraries.[10]

## bfloat16 floating-point format

**bfloat16** has the following format:

- Sign bit: 1 bit
- Exponent width: 8 bits
- Significand precision: 8 bits (7 explicitly stored), as opposed to 24 bits in a classical single-precision floating-point format

The bfloat16 format, being a truncated IEEE 754 single-precision 32-bit float, allows for fast conversion to and from an IEEE 754 single-precision 32-bit float; in conversion to the bfloat16 format, the exponent bits are preserved while the significand field can be reduced by truncation (thus corresponding to round toward 0), ignoring the NaN special case. Preserving the exponent bits maintains the 32-bit float's range of ≈ 10^{−38} to ≈ 3 × 10^{38}.[11]

The bits are laid out as follows:

Compare to an IEEE 754 single-precision 32-bit float:

And contrast with an IEEE half-precision 16-bit float:

### Contrast with bfloat16 and single precision

S | E | E | E | E | E | E | E | E | F | F | F | F | F | F | F | f | f | f | f | f | f | f | f | f | f | f | f | f | f | f | f |

#### Legend

### Exponent encoding

The bfloat16 binary floating-point exponent is encoded using an offset-binary representation, with the zero offset being 127; also known as exponent bias in the IEEE 754 standard.

- E
_{min}= 01_{H}−7F_{H}= −126 - E
_{max}= FE_{H}−7F_{H}= 127 - Exponent bias = 7F
_{H}= 127

Thus, in order to get the true exponent as defined by the offset-binary representation, the offset of 127 has to be subtracted from the value of the exponent field.

The minimum and maximum values of the exponent field (00_{H} and FF_{H}) are interpreted specially, like in the IEEE 754 standard formats.

Exponent | Significand zero | Significand non-zero | Equation |
---|---|---|---|

00_{H} |
zero, −0 | subnormal numbers | (−1)^{signbit}×2^{−126}× 0.significandbits |

01_{H}, ..., FE_{H} |
normalized value | (−1)^{signbit}×2^{exponentbits−127}× 1.significandbits | |

FF_{H} |
±infinity | NaN (quiet, signaling) |

The minimum positive normal value is 2^{−126} ≈ 1.18 × 10^{−38} and the minimum positive (subnormal) value is 2^{−126−7} = 2^{−133} ≈ 9.2 × 10^{−41}.

## Encoding of special values

### Positive and negative infinity

Just as in IEEE 754, positive and negative infinity are represented with their corresponding sign bits, all 8 exponent bits set (FF_{hex}) and all significand bits zero. Explicitly,

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

### Not a Number

Just as in IEEE 754, NaN values are represented with either sign bit, all 8 exponent bits set (FF_{hex}) and not all significand bits zero. Explicitly,

```
val s_exponent_signcnd
+NaN = 0_11111111_klmnopq
-NaN = 1_11111111_klmonpq
```

where at least one of *k, l, m, n, o, p,* or *q* is 1. As with IEEE 754, NaN values can be quiet or signaling, although there are no known uses of signaling bfloat16 NaNs as of September 2018.

## Range and precision

Bfloat16 is designed to maintain the number range from the 32-bit IEEE 754 single-precision floating-point format (binary32), while reducing the precision from 24 bits to 8 bits.

## Examples

These examples are given in bit *representation*, in hexadecimal and binary, of the floating-point value. This includes the sign, (biased) exponent, and significand.

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

7f7f = 0 11111110 1111111 = (2^{8}− 1) × 2^{−7}× 2^{127}≈ 3.38953139 × 10^{38}(max finite positive value in bfloat16 precision) 0080 = 0 00000001 0000000 = 2^{−126}≈ 1.175494351 × 10^{−38}(min normalized positive value in bfloat16 precision and single-precision floating point)

The maximum positive finite value of a normal bfloat16 number is 3.38953139 × 10^{38}, slightly below (2^{24} − 1) × 2^{−23} × 2^{127} = 3.402823466 × 10^{38}, the max finite positive value representable in single precision.

### Zeros and infinities

0000 = 0 00000000 0000000 = 0 8000 = 1 00000000 0000000 = −0

7f80 = 0 11111111 0000000 = infinity ff80 = 1 11111111 0000000 = −infinity

### Special values

4049 = 0 10000000 1001001 = 3.140625 ≈ π ( pi ) 3eab = 0 01111101 0101011 = 0.333984375 ≈ 1/3

### NaNs

ffc1 = x 11111111 1000001 => qNaN ff81 = x 11111111 0000001 => sNaN

## See also

- Half-precision floating-point format: 16-bit float w/ 1-bit sign, 5-bit exponent, and 11-bit significand, as defined by IEEE 754
- ISO/IEC 10967, Language Independent Arithmetic
- Primitive data type
- Minifloat

## References

- Tagliavini, Giuseppe; Mach, Stefan; Rossi, Davide; Marongiu, Andrea; Benin, Luca (2018). "A transprecision floating-point platform for ultra-low power computing".
*2018 Design, Automation & Test in Europe Conference & Exhibition (DATE)*. pp. 1051–1056. arXiv:1711.10374. doi:10.23919/DATE.2018.8342167. ISBN 978-3-9819263-0-9. - Khari Johnson (2018-05-23). "Intel unveils Nervana Neural Net L-1000 for accelerated AI training".
*VentureBeat*. Retrieved 2018-05-23....Intel will be extending bfloat16 support across our AI product lines, including Intel Xeon processors and Intel FPGAs.

- Michael Feldman (2018-05-23). "Intel Lays Out New Roadmap for AI Portfolio".
*TOP500 Supercomputer Sites*. Retrieved 2018-05-23.Intel plans to support this format across all their AI products, including the Xeon and FPGA lines

- Lucian Armasu (2018-05-23). "Intel To Launch Spring Crest, Its First Neural Network Processor, In 2019".
*Tom's Hardware*. Retrieved 2018-05-23.Intel said that the NNP-L1000 would also support bfloat16, a numerical format that’s being adopted by all the ML industry players for neural networks. The company will also support bfloat16 in its FPGAs, Xeons, and other ML products. The Nervana NNP-L1000 is scheduled for release in 2019.

- "Available TensorFlow Ops | Cloud TPU | Google Cloud".
*Google Cloud*. Retrieved 2018-05-23.This page lists the TensorFlow Python APIs and graph operators available on Cloud TPU.

- Elmar Haußmann (2018-04-26). "Comparing Google's TPUv2 against Nvidia's V100 on ResNet-50".
*RiseML Blog*. Archived from the original on 2018-04-26. Retrieved 2018-05-23.For the Cloud TPU, Google recommended we use the bfloat16 implementation from the official TPU repository with TensorFlow 1.7.0. Both the TPU and GPU implementations make use of mixed-precision computation on the respective architecture and store most tensors with half-precision.

- Tensorflow Authors (2018-07-23). "ResNet-50 using BFloat16 on TPU".
*Google*. Retrieved 2018-11-06. - Joshua V. Dillon, Ian Langmore, Dustin Tran, Eugene Brevdo, Srinivas Vasudevan, Dave Moore, Brian Patton, Alex Alemi, Matt Hoffman, Rif A. Saurous (2017-11-28). TensorFlow Distributions (Report). arXiv:1711.10604. Bibcode:2017arXiv171110604D. Accessed 2018-05-23.
All operations in TensorFlow Distributions are numerically stable across half, single, and double floating-point precisions (as TensorFlow dtypes: tf.bfloat16 (truncated floating point), tf.float16, tf.float32, tf.float64). Class constructors have a validate_args flag for numerical asserts

CS1 maint: multiple names: authors list (link) - "BFloat16 extensions for Armv8-A".
*community.arm.com*. Retrieved 2019-08-30. - "ROCm version history".
*github.com*. Retrieved 2019-10-23. - "Livestream Day 1: Stage 8 (Google I/O '18) - YouTube".
*Google*. 2018-05-08. Retrieved 2018-05-23.In many models this is a drop-in replacement for float-32