Comment by ncdr

That's not true - they don't lock global weights when updating, so if you have lots of cores and threads, asynchronous updates will result in very poor accuracy, making training useless.