BFloat16 offers a significantly higher range than FP16, which caps out at just 6.55 x 10^4, forcing certain AI researchers to "fallback" to the relatively inefficient FP32 math hardware. BFloat16 uses three fewer significand bits than FP16 (8 bits versus 11 bits), offering 8 exponent bits, while FP16 only offers 5 bits. BFloat16 is more resilient to overflow and underflow in conversions to FP32 than FP16 is, since BFloat16 is essentially a truncated FP32.
Future AMD GPUs to adopt BFloat16 floating point support?
Posted on Tuesday, October 22 2019 @ 10:33 CEST by Thomas De Maesschalck