BF16
BF16, brain float 16, is a 16-bit floating point format,
- 1 sign bit, 8 exponent bits, 7 mantissa bits
- Lower precision, same range as FP32
- Range: ~1e-38 to 3.4e38
- More stable for training, less likely to overflow
BF16, brain float 16, is a 16-bit floating point format,