Achieving even 100x improvement in the energy efficiency of training / inferencing neural networks sounds too good to be true. Still, as ASICs are getting to th

Achieving even 100x improvement in the energy efficiency of training / inferencing neural networks sounds too good to be true. Still, as ASICs are getting to their limits, it makes sense to try lower level changes in the architecture.