Infrastructure & Agents

What Is bfloat16?

bfloat16 uses 16 bits but keeps the same exponent range as 32-bit floats, trading some precision for a wide dynamic range. This makes it well suited to training deep learning models with fewer numeric stability issues than FP16.