Bfloat16 - AshokBhat/notes GitHub Wiki

Bfloat16

• 16-bit floating-point format
• Format: 1 bit sign, 8 bit exponent, 7 bit mantissa
• Better suited for deep learning than FP16

Support

• By Intel, Arm, NVIDIA and Google TPU

See Also

• [FP16]] ](/AshokBhat/notes/wiki/[FP32) | [bfloat16]]