Unverified60% confidenceFactExact time
FP8 has two main sub-formats — E4M3 (4-bit exponent + 3-bit mantissa) and E5M2 (5-bit exponent + 2-bit mantissa) — optimized for forward propagation and backward propagation respectively.
2
Sources
60%
Confidence
Long-term
Relevance
7/10/2026
First Seen
Sources
Krea2 Tested: Does the FP8 vs BF16 Quality Gap Still Exist?
redditr/StableDiffusion7/9/2026
Related Claims
UnverifiedFP8 training schemes typically use E4M3 for the forward pass and E5M2 for the backward pass.74% similarUnverifiedMXFP8 has two variants: E4M3 (4-bit exponent, 3-bit mantissa) and E5M2 (5-bit exponent, 2-bit mantissa).71% similarUnverifiedFP4采用1位符号位、2位指数位和1位尾数位的布局69% similarUnverifiedNVFP4 采用 E2M1 编码（2位指数、1位尾数、1位符号位），以每16个元素为一组配置独立的 FP8 缩放因子69% similarUnverifiedThe M5 Ultra's FP8 and FP4 floating-point compute performance is identical to its BF16 performance, and the gap with the Spark narrows noticeably when dropping to FP4 precision.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/464365API
curl https://kongchang.com/api/v1/knowledge/claims/464365MCP
get_claim(id=464365)