IBM Aims To Reduce Power Needed For Neural Net Training By 100x