Neural Network inference on commodity CPUs using sparsity

Up until now sparsity has looked like a promising way to reduce resource requirements but you paid a terrible price in accuracy. If they've really resolved this, it's big.