Expired60% confidenceFactTime unknown
In mixed-precision training, each FP16 parameter occupies 2 bytes of GPU memory, and each FP32 parameter occupies 4 bytes of GPU memory.
2
Sources
60%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
Valid until: 9/30/2026(expired)
Sources
Related Claims
Unverified在混合精度Adam训练中，单个参数的完整训练状态高达18字节（FP16权重2字节、FP32主权重4字节、一阶矩4字节、二阶矩4字节，加梯度）82% 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77% similarUnverifiedCompared to FP16, BF16 has the same 8-bit exponent width as FP32 and therefore a larger dynamic range, making it the mainstream precision choice for current large model training.74% 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71% similarUnverified叠加cv=5的5折交叉验证后，144种参数组合需训练144×5=720次模型68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/61142API
curl https://kongchang.com/api/v1/knowledge/claims/61142MCP
get_claim(id=61142)