Scaling up Stochastic Gradient Descent for Non-convex Optimisation