Scaling and distributed training

With mixed-precision AdamW (bf16 weights and gradients, fp32 master copy and two fp32 Adam moments), how many bytes per parameter does training state take, before activations?