Calling ML training non-parallelizable is very strong, no? Matrix multiplication is highly parallelizable.