Single precision is probably fine for a neural network. Neural networks are somewhat insensitive to noise and failure and single precision adds very little noise.