NVIDIA presents Mixed Precision Training
Presented Mixed Precision Training — FP16 training with an FP32 master copy of the weights and loss scaling, cutting memory by about 50%
Why it matters
The prototype of mixed-precision training. The whole lineage of shrinking precision, FP16→BF16→FP8→FP4, is a variation on these two techniques
Source — Primary sources first; secondary reports are labelled “Press”
Paperarxiv.orghttps://arxiv.org/abs/1710.03740Added in Week 35, 2026 · See that week