How does it compare in practice to modern forms of stochastic gradient descent with momentum (e.g., RAdam SGD) when applied to challenging, multi-million-variab

How does it compare in practice to modern forms of stochastic gradient descent with momentum (e.g., RAdam SGD) when applied to challenging, multi-million-variable non-convex optimization problems?