TLDR is you can't. For a very simple example, storing the products of all 3x3 matrices * length 3 vectors in Float16 precision would take 2^193 bytes (which is

TLDR is you can't. For a very simple example, storing the products of all 3x3 matrices * length 3 vectors in Float16 precision would take 2^193 bytes (which is obviously impractical).