Comment by swiftcoder

So the question becomes, how many other parts of the inference pipeline have left 1000x optimization opportunities lying on the table?