Practice
Learn
← Quit
0 / 0 correct
LLMs & AI
Hard
1 of 40
Why is bf16 popular for training instead of fp16?
A
bf16 is more precise than fp32
B
bf16 keeps fp32's exponent range, avoiding overflow/underflow without loss scaling, at the cost of mantissa precision
C
fp16 can't be used on GPUs
D
bf16 uses 8 bits