https://hgpu.org/?p=27597
Mixing Low-Precision Formats in Multiply-Accumulate Units for DNN Training