UE8M0 data format

ue8m0 is an 8-bit unsigned floating-point data format with 8 exponent bits and 0 mantissa bits, designed for microscaling applications in ai workloads.

technical specification

format characteristics

8 exponent bits enabling power-of-two values from 2^-127 to 2^127
0 mantissa bits - functions as pure scaling factor
unsigned format with no sign bit
nan represented as 0xff, no infinity support
must be used in packed format as ue8m0x2

the format was documented in nvidia’s ptx isa version 9.0 (august 2025).¹

binary structure

standard fp8 (e5m2): [S][EEEEE][MM]
standard fp8 (e4m3): [S][EEEE][MMM]
ue8m0 format:        [EEEEEEEE]

microscaling applications

ue8m0 serves as a shared scaling factor for blocks of 32 elements in lower precision formats:²

mxfp8: 8-bit values with ue8m0 scale
mxfp6: 6-bit values with ue8m0 scale
mxfp4: 4-bit values with ue8m0 scale

this approach reduces memory bandwidth by up to 75% while maintaining acceptable accuracy for inference.³

for more on mxfp4 formats and gpt model requirements, see gpt-oss mxfp4 requirements.

hardware benefits

the zero-mantissa design simplifies hardware:

eliminates multiplication circuits
reduces silicon area
optimizes for scaling operations
aligns with 32-element tensor tiles

platform support

nvidia gpus

support introduced in ptx isa 9.0 (august 2025):¹

supported: gb200 and newer
not supported: h100, h20, a100, older architectures

deepseek models

deepseek v3.1 (august 21, 2025) was trained using ue8m0 fp8 scale data format.⁴

implementation

def ue8m0_encode(value):
    """encode float as ue8m0"""
    if value == 0 or math.isnan(value):
        return 0xff if math.isnan(value) else 0x00

    exponent = int(math.log2(abs(value)))
    biased_exp = exponent + 127

    return max(0, min(254, biased_exp))

def ue8m0_decode(byte_value):
    """decode ue8m0 to float"""
    if byte_value == 0xff:
        return float('nan')
    if byte_value == 0x00:
        return 0.0

    exponent = byte_value - 127
    return 2.0 ** exponent

precision considerations

error tolerance increases from 1e-5 (standard fp8) to approximately 7e-4 with ue8m0.⁵ this makes it suitable primarily for inference rather than training.