Efficient Code Search with Nvidia DGX

I wonder where the label ‘mini/micro’ batch came from (‘Training at bfloat16 numeric precision enabled them to use large micro-batch sizes of 256...’), given that batches were never that big to begin with.