Unverified50% confidenceFactExact time
FP32每个参数占4字节，FP16/BF16占2字节，INT8量化占1字节，INT4量化占0.5字节
1
Sources
50%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
苹果M7 Ultra芯片曝光：1.5TB统一内存能否重塑本地AI推理格局
Related Claims
UnverifiedThe FP4 format has only 4 bits (1 sign bit, 2 exponent bits, 1 mantissa bit) and can precisely represent only approximately 16 discrete values.83% similarUnverifiedFP32 uses 32 binary bits to represent a number (1 sign bit, 8 exponent bits, 23 mantissa bits) and can accurately represent approximately 7 significant decimal digits.77% similarUnverifiedBlackwell的FP4格式仅有4位（1位符号、2位指数、1位尾数），采用每128个FP4元素共享一个FP8精度缩放因子的块浮点表示77% similarUnverifiedIn mixed-precision training, each FP16 parameter occupies 2 bytes of GPU memory, and each FP32 parameter occupies 4 bytes of GPU memory.77% similarUnverifiedINT4量化将原本由65536个不同值（FP16）表示的权重压缩到仅16个离散级别76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/507765API
curl https://kongchang.com/api/v1/knowledge/claims/507765MCP
get_claim(id=507765)