← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8688 · 生成模型:GPT-6

浮点范围与精度是不同维度

理解等级:中级

FP32由符号、8位指数及23位存储尾数组成,FP16指数5位尾数10位,BF16指数8位尾数7位。BF16保留更宽数量级范围,却比FP16尾数更短;不能说它恢复了FP32有效数字精度。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      The mantissa/significand (23 bits) For more details on how FP32 works, see “Demystifying Floating Point Precision”. The maximum and minimum value that can be represented using FP32 is 3.4028237 × 1038 and 1.175494 × 1038, respectively. This is referred to as the range of values that can be represented by this format. Similarly, a number represented in float16 (FP16), also referred to as half-precision floating

    参考来源

    - 位置:《Designing Large Language Model Applications》第09章《推理优化》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch09.html) - 结论依据:摘录中关于浮点范围与精度是不同维度的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。 - 原始内容:The mantissa/significand (23 bits) For more details on how FP32 works, see “Demystifying Floating Point Precision”. The maximum and minimum value that can be represented using FP32 is 3.4028237 × 1038 and 1.175494 × 1038, respectively. This is referred to as the range of values that can be represented by this format. Similarly, a number represented in float16 (FP16), also referred to as half-precision floating