bfloat

• A bfloat16 has the same range as float16, but less precision.
• Precisely, the mantissa is given less bits, and the exponent is given more bits
• You use bfloat to shorten training time. Empirically, performance is preserved