[vvipgupta]

The same is true for data (aka gradients) consistency while training large ML models. Asynchronous SGD is as good (and maybe even faster) than synchronous SGD: https://papers.nips.cc/paper/2011/file/218a0aefd1d1a4be65601...