Open Format · Deep Learning · 8-Bit/4-Bit Precision
HiFloat: Advanced AI low bit-width floating-point format
HiFloat (HiF8/HiF4/...) is an efficient floating-point format for AI computing. It outperforms FP16/BF16 and FP8 by delivering higher computing efficiency under identical hardware constraints.
Experimental results for quantization of Gaussian-distributed random numbers, evaluated by MSE and SQNR
quantization_error_reduction
−24%
Assessment of chip PPA gains based on the assumption that the computing power of the HiF4 matrix is twice that of the 8-bit (FP8/INT8) format
incrementa_chip_area_savings
68%
power_reduction
10%
// HiF4: block floating-point format
- Bit width in average4.5bits/value
- Group Size64
- Special valuesNaN Zero Inf
- FormatE1M2
- Base scaleE6M2
- Dynamic range69binades
- Local dynamic range4.81binades
- Max positive value218× 1.3125
- Min positive value2-50
// HiF8: element-wise floating-point format
- Bit width in average8bits/value
- Group SizeUser-defined
- Special valuesNaN Zero Inf
- FormatDot field
- Base scaleUser-defined
- Dynamic range38binades
- Local dynamic rangeUser-defined
- Max positive value215
- Min positive value2-22
Hardware Support
Events and Meetings
Community News
Building the HiFloat Ecosystem Together
25+ partners
Academia & Research
Industry Institutions
Note: Partners are listed in alphabetical order.
























