There was work done by Doug Burger from microsoft, spliting an algorithm to between a digital Cpu and a neural compute unit(simulated in analog). In algorithms

There was work done by Doug Burger from microsoft, spliting an algorithm to between a digital Cpu and a neural compute unit(simulated in analog). In algorithms that we're not bottlenecked by the digital cpu, they've even seen gains of 50x in perf/w. But most algorithms they tested we're bottlenecked by the digital part, so the average benefit was ~3x. But maybe if you start the design using this "turing neural network", you could see consistently large gains.