Why it matters
FP16 was the first big win for ML precision reduction. Understanding it clarifies why BF16 arose and how mixed precision training works.
Advertisement
The architecture
Format: 1 sign bit, 5 exponent bits, 10 mantissa bits.
Range: roughly ±65504. Precision: much less than FP32.
Advertisement
How it works end to end
Dynamic range: narrower than FP32. Gradients can underflow.
Loss scaling: multiply loss by large factor before backprop to keep gradients in range. Standard mixed precision technique.
Hardware: tensor cores support FP16 with dramatic speedup vs FP32 (up to 8x).
Alternative: BF16 has wider range, less mantissa. Often preferable for training.