Mixed precision

Compute interpretation

Low-precision training adaptation that increases accelerator throughput while preserving convergence with loss scaling and FP32 state where needed.

Supporting reading cards

• Mixed Precision Training (2017, multi_gpu_dense_training)
• BitNet b1.58 2B4T Technical Report (2025, efficient_edge_inference)

Obsolete or less central under later compute

Track this only through linked reading cards; do not treat this method page as standalone evidence.