Consider the neural network model MLPO from Figure $7\mathrm{.}5.$ That model has $20M$ weights in five fully connected layers $($neural network researchers count the input layer as if it were a layer in the stack, but it has no weights associated with it$).$ For simplicity, let's assume that those layers are each the same size, so each layer holds $4M$ weights. Then assume that each layer has identical geometry, so each group of $4M$ weights represents a $2k**2k$ matrix. Because the TPU typically uses $8-$bit numerical values, $20M$ weights take up $20MB.$