Unverified50% confidenceFactExact time
FP8 (E4M3/E5M2) occupies 1 byte per parameter and is the mainstream low-precision inference format on current Hopper and Blackwell.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/15/2026
First Seen
Valid until: 1/13/2027
Sources
Related Entities
Related Claims
UnverifiedBlackwell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72% similarVerifiedFP8 has two variants: E4M3 (4-bit exponent + 3-bit mantissa), which has higher precision and is suited for forward inference, and E5M2 (5-bit exponent + 2-bit mantissa), which has a larger dynamic range and is suited for gradient computation.68% similarUnverifiedGemma 4采用混合专家架构（MoE），E2B后缀表示Effective 2B，每次推理只激活约20亿有效参数63% similarVerifiedFP32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62% similarUnverifiedLFM2.5-8B-A1B总参数量为8B，推理时仅激活1.5B参数61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1032579API
curl https://kongchang.com/api/v1/knowledge/claims/1032579MCP
get_claim(id=1032579)