onnx--onnx/docs/docsgen/source/technical/float8.md

(onnx-detail-float8)=

Float stored in 8 bits

Papers

Two papers have been published in 2022 to introduce floats stored on a byte as opposed to float 32 stored on 4 bytes. The float precision is much lower but the training accuracy does not suffer too much.

FP8 Formats for Deep Learning from NVIDIA, Intel and ARM introduces two types following IEEE specification. First one is E4M3, 1 bit for the sign, 4 bits for the exponents and 3 bits for the mantissa. Second one is E5M2, 1 bit for the sign, 5 bits for the exponents and 2 for the mantissa. The first types is mostly used for the weights, the second one for the gradient.

Second paper 8-bit Numerical Formats For Deep Neural Networks introduces similar types. IEEE standard gives the same value to +0 (or integer 0) and -0 (or integer 128). They chose to give distinct float values to these two numbers. The paper experiments different split between exponent and mantissa and shows and E4M3 and E5M2 are the best ones.

As a result, four new types were introduced in onnx==1.15.0 to support a limited set of operators to enable computation with float 8.

• E4M3FN: 1 bit for the sign, 4 bits for the exponents, 3 bits for the mantissa, only nan values and no infinite values (FN),
• E4M3FNUZ: 1 bit for the sign, 4 bits for the exponents, 3 bits for the mantissa, only nan values and no infinite values (FN), no negative zero (UZ)
• E5M2: 1 bit for the sign, 5 bits for the exponents, 2 bits for the mantissa,
• E5M2FNUZ: 1 bit for the sign, 5 bits for the exponents, 2 bits for the mantissa, only nan values and no infinite values (FN), no negative zero (UZ)

The implementation is usually hardware dependent. NVIDIA, Intel and Arm implement E4M3FN and E5M2 is its latest graphical processor. GraphCore does the same only with E4M3FNUZ and E5M2FNUZ.

E4M3FN and E5M2

S stands for the sign. 10_2 describe a number base 2.

.. list-table:: Float8 types
   :widths: 10 10 10
   :header-rows: 1

   * -
     - E4M3FN
     - E5M2
   * - Exponent bias
     - 7
     - 15
   * - Infinities
     -
     - :math:`S.11111.00_2`
   * - NaN
     - :math:`S.1111.111_2`
     - :math:`S.11111.\{01, 10, 11\}_2`
   * - Zeros
     - :math:`S.0000.000_2`
     - :math:`S.00000.00_2`
   * - Max
     - :math:`S.1111.110_2`
     - :math:`1.75 \times 2^{15}= 57344`
   * - Min
     - :math:`S.0000.001_2 = 2^{-9}`
     - :math:`S.00000.01_2 = 2^{-16}`

Let's denote the bit representation as S.b_6 b_5 b_4 b_3 b_2 b_1 b_0. The float value is defined by the following expressions:

.. list-table:: Float8 types values
   :widths: 10 10 10
   :header-rows: 1

   * -
     - E4M3FN
     - E5M2
   * - exponent :math:`\neq` 0
     - :math:`(-1)^S 2^{\sum_{i=3}^6 b_i 2^{i-3} - 7} \left( 1 + \sum_{i=0}^2 b_i 2^{i-3} \right)`
     - :math:`(-1)^S 2^{\sum_{i=2}^6 b_i 2^{i-2} - 15} \left( 1 + \sum_{i=0}^1 b_i 2^{i-2} \right)`
   * - exponent :math:`=` 0
     - :math:`(-1)^S 2^{-6} \sum_{i=0}^2 b_i 2^{i-3}`
     - :math:`(-1)^S 2^{-14} \sum_{i=0}^1 b_i 2^{i-2}`

E4M3FNUZ and E5M2FNUZ

The previous types support positive and negative zero, positive and negative nan. Another type definition was introduced by GraphCore to make a better use of these four values. Every type including UZ in its name have only one zero and one nan (= negative zero). The other difference comes from the exponent bias. As a result, a float 8 FLOAT8E4M3FN, not null, not nan, cannot be simply converted into FLOAT8E4M3FNUZ due to this exponent bias difference. Even if the mantissa is the same, the exponent is not.

.. list-table:: Float8 types
   :widths: 10 10 10
   :header-rows: 1

   * -
     - E4M3FNUZ
     - E5M2FNUZ
   * - Exponent bias
     - 8
     - 16
   * - Infinities
     -
     -
   * - NaN
     - :math:`1.0000.000_2`
     - :math:`1.00000.00_2`
   * - Zeros
     - :math:`0.0000.000_2`
     - :math:`0.00000.00_2`
   * - Max
     - :math:`S.1111.111_2`
     - :math:`S.11111.11_2`
   * - Min
     - :math:`S.0000.001_2 = 2^{-10}`
     - :math:`S.00000.01_2 = 2^{-17}`

The float value is defined by the following expressions:

.. list-table:: Float8 types values
   :widths: 10 10 10
   :header-rows: 1

   * -
     - E4M3FNUZ
     - E5M2FNUZ
   * - exponent :math:`\neq` 0
     - :math:`(-1)^S 2^{\sum_{i=3}^6 b_i 2^{i-3} - 8} \left( 1 + \sum_{i=0}^2 b_i 2^{i-3} \right)`
     - :math:`(-1)^S 2^{\sum_{i=2}^6 b_i 2^{i-2} - 16} \left( 1 + \sum_{i=0}^1 b_i 2^{i-2} \right)`
   * - exponent :math:`=` 0
     - :math:`(-1)^S 2^{-7} \sum_{i=0}^2 b_i 2^{i-3}`
     - :math:`(-1)^S 2^{-15} \sum_{i=0}^1 b_i 2^{i-2}`

Cast

Cast from float 8 to float 16 (or E5M10), bfloat16 (or E8M7), float32 (or E8M23) is easier. The cast is exact. The conversion does not necessarily preserve the sign for specific values such as -0 or -NaN.

Cast to float 8 consists in finding the closest float 8 to the original float 32 value. It is usually done by shifting and truncating.

The conversion may with saturation, every value out of range becomes the highest available value. Next table summarizes all the case. [x] means the value rounded to the target mantissa width.

x	E4M3FN	E4M3FNUZ	E5M2	E5M2FNUZ
0	0	0	0	0
-0	-0	0	-0	0
NaN	NaN	NaN	NaN	NaN
Inf	FLT_MAX	NaN	FLT_MAX	NaN
-Inf	-FLT_MAX	NaN	-FLT_MAX	NaN
[x] > FLT_MAX	FLT_MAX	FLT_MAX	FLT_MAX	FLT_MAX
[x] < -FLT_MAX	-FLT_MAX	-FLT_MAX	-FLT_MAX	-FLT_MAX
else	RNE	RNE	RNE	RNE

The conversion may also be defined without any saturation.

x	E4M3FN	E4M3FNUZ	E5M2	E5M2FNUZ
0	0	0	0	0
-0	-0	0	-0	0
NaN	NaN	NaN	NaN	NaN
-NaN	-NaN	NaN	-NaN	NaN
Inf	NaN	NaN	Inf	NaN
-Inf	-NaN	NaN	-Inf	NaN
[x] > FLT_MAX	NaN	NaN	Inf	NaN
[x] < -FLT_MAX	NaN	NaN	-Inf	NaN
else	RNE	RNE	RNE	RNE

E8M0

The E8M0 data type serves as the common scale type for all OCP Microscaling (MX) Formats. It has eight bits for the exponent, and no sign or mantissa bits.

.. list-table:: E8M0
   :widths: 10 10
   :header-rows: 1

   * -
     - E8M0
   * - Exponent bias
     - 127
   * - Infinities
     -
   * - NaN
     - :math:`11111111_2`
   * - Zeros
     -
   * - Max
     - :math:`11111110_2 = 2^{127}`
   * - Min
     - :math:`00000000_2 = 2^{-127}`

When computing scale factors in MX formats, there are different casting choices one can make. For this reason, the ONNX spec for the Cast operator has introduced an additional "round_mode" attribute, which accepts the following:

• "up": round to nearest value away from zero
• "down": round to nearest value towards zero
• "nearest": round to nearest value and ties round up

It has been shown that rounding up with saturation achieves superior accuracy in LLM pretraining compared to other rounding modes.